Sie sind auf Seite 1von 243

The Pennsylvania State University

The Graduate School


Department of Agricultural Economics and Rural Sociology

COMMODITY FUTURES IN ASSET ALLOCATION

A Thesis in
Agricultural, Environmental and Regional
Economics and Operation Research
by

Shengwu Du

2005 Shengwu Du

Submitted in Partial Fulfillment


of the Requirements
for the Degree of

Doctor of Philosophy
December 2005

The thesis of Shengwu Du has been reviewed and approved* by the following:

Spiro E. Stefanou
Professor of Agricultural Economics
Thesis Adviser
Chair of Committee

Jeffrey R. Stokes
Associate Professor of Agricultural Economics

James W. Dunn
Professor of Agricultural Economics

Jingzhi Huang
Associate Professor of Finance

Stephen M. Smith
Professor of Agricultural and Regional Economics
Head of the Department of Agricultural Economics
and Rural Sociology

*Signatures are on file in the Graduate School.

iii

ABSTRACT
Commodity futures can supply fair investment return and unique diversification
benefits to a portfolio manager. The relative performance between commodity futures
and traditional financial assets is varied with business and market conditions. In general,
commodity futures are negatively correlated with stocks and bonds. This negative
correlation comes from the nature of the commodity market. Commodity futures can
supply efficient diversification benefit to stocks and bonds investors when it is needed,
i.e. when the stock or bond market displays poor performance or has a significant
downside movement, commodity futures market usually shows strong performance and
provides a good return.
Both commodity futures and traditional financial assets display the structure
break behavior and their risk/return characters change under different market and
economic conditions. Markov regime-switching model can be applied to address this
asset return dynamic and corresponding portfolio selection issue. The estimation results
of a simple two-regime model support that for three risky asset classes there is a
normal regime characterized by relatively high return and low risk and a bad regime
characterized by relatively low return and high risk. An investment strategy ignoring the
regime switching effect leads to inefficient asset allocation and poorer portfolio
performance. When the market is in the bad regime, commodity futures perform
relatively stronger than stocks and can provide more significant diversification benefit.
Meanwhile, the regime-switching model can predict the break change of financial market
and send signal to investors to increase commodity futures investment when financial
market switches to the bad regime.
A general regime-switching model extends the simple regime-switching model
by using economic variables as instruments to predict asset returns. This model helps
investigate asset return dynamics and optimal asset allocation under alternative economic
environments. The first general regime switching model uses the short rate as the

iv

instrumental variable to predict expected asset returns and regime switching probability.
This model shows that the short rate plays an important role in asset return regime
switching and optimal asset allocation. The optimal asset weights on risky assets
negatively relate to the T-bill rate. This is not surprising as the T-bill rate determines the
risk-free asset. When the return of risk-free asset is increasing, investors expect a greater
return from the risky assets and have a greater chance to switch investment to the
risk-free asset if the expectation is not met. However, the fraction of commodity futures
in the risky asset portfolio increases with the short rate level.
The second general regime switching model is the IPI model which uses the
monthly change of industrial production index (IPI) as the instrumental variable to model
risky asset returns. The estimation results of this model support that the IPI growth rate
plays an important role for efficient asset allocation since it determines both the regime
transition probability and the expected returns of risky assets. Commodity futures return
is always positively correlated with the IPI growth rate and the optimal weights on
commodity futures increase with the IPI growth rate.
To further investigate the diversification benefit of commodity futures, an
alternative approach is provided. Commodity futures are included into the initial
portfolio with various stocks and bonds components and a new risk measure, VaR, is
adopted to measure the risk of these portfolios. Three different VaR estimations are
proposed: parametric VaR with regime-switching model, historical simulation VaR, and
the EVT VaR. These estimation results again strongly support that commodity futures
can reduce portfolio risk and improve portfolio performance. However, for different
initial traditional portfolios, the optimal weight on commodity futures is not identical.
Allocating 5-10% capital on commodity futures investment, as most intuitional
investors have done, is not the optimal investment strategy, especially for equity
investors. Meanwhile, extreme negative market movement for the traditional asset does
not occur synchronously with that for commodity futures; hence, commodity futures do
provide good downside protection.

TABLE OF CONTENTS

LIST OF TABLES...................................................................................................... VII


LIST OF FIGURES .....................................................................................................IX
ACKNOWLEDGMENTS ...........................................................................................XI
CHAPTER 1 INTRODUCTION AND LITERATURE REVIEW............................ 1
1.1 COMMODITY FUTURES: AN ALTERNATIVE ASSET ................................................ 1
1.2 RESEARCH ON COMMODITY FUTURES INVESTMENT ............................................ 5
1.3 STATEMENT OF PURPOSE AND PROSPECTS .......................................................... 10
CHAPTER 2 QUANTITATIVE PORTFOLIO SELECTION MODEL ................ 16
2.1 OVERVIEW ............................................................................................................ 16
2.2 STATIC PORTFOLIO SELECTION THEORY ............................................................ 17
2.3 DYNAMIC PORTFOLIO SELECTION THEORY ....................................................... 28
2.4 MARKOV REGIME SWITCHING PORTFOLIO SELECTION MODEL ....................... 36
2.5 PORTFOLIO ANALYSIS WITH VALUE AT RISK ....................................................... 41
CHAPTER 3 COMMODITY FUTURES INDEX INVESTMENT ........................ 50
3.1 INVESTING IN COMMODITY FUTURES INDEX ...................................................... 50
3.2 STAND-ALONE INVESTMENT PERFORMANCE ...................................................... 53
3.3 INVESTMENT PERFORMANCE COMPARED WITH STOCKS AND BONDS ............... 55
3.4 COMMODITY FUTURES INVESTMENT AND INFLATION ........................................ 59
3.5 COMMODITY FUTURES INVESTMENT AND BUSINESS CYCLE .............................. 61
3.6 COMMODITY FUTURES INVESTMENT AND EVENT RISK...................................... 65
3.7 CONCLUSION ........................................................................................................ 66
CHAPTER 4 STATIC ASSET ALLOCATION WITH REGIME-SWITCHING . 80
4.1 INTRODUCTION OF REGIME SWITCHING MODEL ............................................... 80
4.2 REGIME SWITCHING IN COMMODITY FUTURES .................................................. 87
4.3 SINGLE REGIME SWITCHING MODEL FOR THREE ASSET CLASSES ................... 90
4.4 OPTIMAL PORTFOLIO ALLOCATION WITHOUT RISK-FREE ASSET ..................... 93
4.5 OPTIMAL PORTFOLIO ALLOCATION WITH RISK-FREE ASSET .......................... 101
4.6 CONCLUSION ...................................................................................................... 105
CHAPTER 5 DYNAMIC PORTFOLIO SELECTION WITH COMMODITY
FUTURES ................................................................................................................... 139
5.1. GENERAL REGIME-SWITCHING MODEL .......................................................... 139
5.2 PORTFOLIO SELECTION WITH T-BILL MODEL .................................................. 141
5.3 PORTFOLIO SELECTION WITH IPI MODEL ........................................................ 147

vi

CHAPTER 6 PORTFOLIO SELECTION WITH THE VALUE AT RISK ......... 177


6.1 WHY USE THE VALUE AT RISK ........................................................................... 177
6.2 TRADITIONAL VAR ANALYSIS ............................................................................ 178
6.3 PORTFOLIO VAR WITH THE EXTREME VALUE THEORY ................................... 184
6.4 CONCLUSION ...................................................................................................... 193
CHAPTER 7 CONCLUSIONS AND IMPLICATIONS ........................................ 209
REFERENCES........................................................................................................... 216

vii

LIST OF TABLES

Table 3-1: The component breakdown of GSCI main index ......................................... 68


Table 3-2: Spot returns, collateral yield, and roll yield of GSCI ................................... 69
Table 3-3: Summary statistics of historical total returns on GSCI indexes ................... 70
Table 3-4: Correlation structure of GSCI sub-indexes................................................... 70
Table 3-5: Historical total returns on GSCI, Stocks, and Bonds.................................... 71
Table 3-6: Correlation of GSCI indexes with stocks and T-bond .................................. 71
Table 3-7: Portfolio performance of downside risk protection ...................................... 71
Table 3-8: GSCI correlation with different inflation components ................................. 72
Table 3-9: The NBER business cycle chronology ......................................................... 72
Table 3-10: Average asset returns in four phases ........................................................... 72
Table 3-11: Year returns with events .............................................................................. 73
Table 4-1: Hansens LR tests for regime-switching GSCI indexes.............................. 107
Table 4-2: Estimates of regime switching model for commodity futures index .......... 107
Table 4-3: Hamiltons LM test for model misspecification ......................................... 108
Table 4-4: Regime switching model estimates for GSCI, Stocks and Bonds .............. 108
Table 4-5: Correlation of transition probability among three asset classes ................. 108
Table 4-6: Parameters estimation of simple MRS model for three assets.................... 109
Table 4-7: Sharpe ratio of three asset classes............................................................... 109
Table 4-8: Correlation among three asset classes ........................................................ 110
Table 4-9: Sharpe ratio comparing under different regimes ........................................ 110
Table 4-10: Sharpe ratio comparing for optimal portfolio without risk-free asset ...... 110
Table 4-11: Sharpe ratio comparing for simulated one period returns over 10 years
(without risk-free asset) ........................................................................................ 111
Table 4-12: Accumulated end wealth over 10 years simulation without risk-free asset
............................................................................................................................... 111
Table 4-13: Average weight on GSCI over 10 years investment simulation without
risk-free asset ........................................................................................................ 111
Table 4-14: Optimal risky asset allocation when a risk-free asset existing ................. 112
Table 4-15: Sharpe ratio comparing for optimal portfolio with risk-free asset............ 112
Table 4-16: Sharpe ratio comparing for simulated one period returns over 10 years
(with risk-free asset).............................................................................................. 112
Table 4-17: Cumulated wealth at the end of 10 years investment simulation (with
risk-free asset) ....................................................................................................... 112
Table 4-18: Average optimal weight on GSCI over 10 years investment simulation
(with risk-free asset).............................................................................................. 112
Table 5-1: T-bill model parameters estimation............................................................. 156
Table 5-2: Variance-covariance matrix estimation from T-bill model ......................... 157
Table 5-3: Correlation matrix estimation from T-bill model........................................ 157
Table 5-4: Parameters estimations of Industrial production index model.................... 158

viii

Table 5-5: Variance-covariance matrix estimation for IPI model ................................ 159
Table 6-1: Percentage of Reward-to-VaR increase in different regime ....................... 194
Table 6-2: Historical simulation result of portfolio VaR decrease ............................... 194
Table 6-3: Historical simulation result of portfolio Reward-to-VaR increase ............. 195
Table 6-4: Descriptive statistics of three assets daily returns ...................................... 195
Table 6-5: Maximum estimation results of left tail on three risky asset returns .......... 196
Table 6-6: Percentage of portfolio EVT VaR decrease ................................................ 196
Table 6-7: Percentage of portfolio EVT Reward-to-VaR increase............................... 197

ix

LIST OF FIGURES

Figure 2-1: Efficient frontier .......................................................................................... 20


Figure 3-1: Risks and returns tradeoff for GSCI indexes .............................................. 73
Figure 3-2: The risks and returns characteristics of GSCI, Stocks, and Bonds ............. 74
Figure 3-3: Time series plot graph of return on GSCI, stocks, and bonds..................... 75
Figure 3-4: Correlation under different investment horizon .......................................... 76
Figure 3-5: Efficient frontier with GSCI, Stocks, and Bonds ........................................ 77
Figure 3-6: Downside risk protection ............................................................................ 77
Figure 3-7: Standardized year change of GSCI total return index and CPI................... 78
Figure 3-8: Correlation with inflation ............................................................................ 78
Figure 3-9: Correlation with different inflation components ......................................... 78
Figure 3-10: Average returns of commodity futures, stocks and bonds with business
cycle ........................................................................................................................ 79
Figure 3-11: Average returns of commodity futures, stocks and bonds under different
phase........................................................................................................................ 79
Figure 4-1: Smooth probability of state 1 for commodity future indexes.................... 113
Figure 4-2: Smooth probability of state 1 for three asset classes................................. 113
Figure 4-3: Smooth probability of state 1 for three asset classes MRS ....................... 113
Figure 4-4: Mean-variance efficient frontier with three asset classes.......................... 114
Figure 4-5: Optimal portfolio weight on three assets................................................... 115
Figure 4-6: risk and return comparing (no risk-free asset) .......................................... 118
Figure 4-7: Mean and standard deviation of simulated one-period return (no risk-free
asset)...................................................................................................................... 120
Figure 4-8: Accumulated wealth over 10 years investment simulation (no risk-free
asset)...................................................................................................................... 122
Figure 4-9: Optimal weight on GSCI over 10 years investment simulation (no risk-free
asset)...................................................................................................................... 126
Figure 4-10: Optimal capital allocation with three risky assets and a risk-free asset .. 130
Figure 4-11: Optimal weight on 4 asset classes ........................................................... 131
Figure 4-12: Risk and return comparing (with risk-free asset) .................................... 132
Figure 4-13: Mean and standard deviation of simulated one-period return (with
risk-free asset) ....................................................................................................... 134
Figure 4-14: Accumulate wealth over 10 years investment simulation (with risk-free
asset)...................................................................................................................... 135
Figure 4-15: Optimal weight on GSCI over the 10 years investment simulation (with
risk-free asset) ....................................................................................................... 137
Figure 5-1: Smoother probability of regime 1 at period t for T-bill model.................. 160
Figure 5-2: Transition probability for T-bill model...................................................... 161
Figure 5-3: One-period ahead optimal asset allocation (T-bill model) ........................ 162
Figure 5-4: Multi-period ahead optimal asset allocation (T-bill model)...................... 166

Figure 5-5: Transition probability for IPI model ......................................................... 168


Figure 5-6: Smoother probability of regime 1 at period t for IPI model...................... 169
Figure 5-7: One-period ahead optimal asset allocation (IPI model) ............................ 170
Figure 5-8: Multi-period ahead optimal asset allocation (IPI model).......................... 173
Figure 6-1: VaR in regime 1 for different portfolios.................................................... 198
Figure 6-2: Reward-to-VaR in regime 1 for different portfolios.................................. 199
Figure 6-3: VaR in regime 2 for different portfolios.................................................... 200
Figure 6-4: Reward-to-VaR in regime 2 for different portfolios.................................. 201
Figure 6-5: Historical simulation VaR for different portfolios .................................... 202
Figure 6-6: Historical simulation Reward-to-VaR for different portfolios .................. 203
Figure 6-7: VaR estimates and the number of exceedances ......................................... 204
Figure 6-8: the EVT VaR for different portfolios ........................................................ 206
Figure 6-9: the EVT ES for different portfolios........................................................... 207
Figure 6-10: the EVT Reward-to-VaR for different portfolios .................................... 208

xi

ACKNOWLEDGMENTS

This dissertation was influenced and improved by advice, knowledge, and


assistance of several people. First and foremost, my thanks go to my dissertation advisor
Dr. Darren Frechette. His help is invaluable for completing and improving this project. I
am grateful to my committee chairman Dr. Spiro Stefanou for his guidance, support, and
encouragement. I am proud that his guidance will be foundation of my future scholarly
work. I would like to express my deep gratitude to my other committee member: Dr.
Stokes, Dr. Huang, and Dr. Dunn for their critical, yet constructive guidance and
encouragement. The financial support of the Department of Agricultural Economics and
Rural Sociology is acknowledged.

Chapter 1 INTRODUCTION AND LITERATURE REVIEW

Commodities are becoming mainstream in the investment world. There is a


growing number of institutional investors increasing their asset allocation to commodity
futures to achieve better investment performance. My primary interest in this dissertation
lies in studying the return and risk of commodity futures in relation to strategic asset
allocation of institutional investors, i.e. the allocation to broad asset classes such as
stocks, bonds, and commodity futures. To professional financial investment institutions
such as mutual funds, pension funds, and school endowments, the strategic asset
allocation is the single most important determinant of the risk and return characteristics
of their portfolio.

1.1 Commodity Futures: An Alternative Asset


During the past decade, the investment management industry has undergone
numerous changes. The great volume of international capital flow makes the equity
markets a single global asset class and distinctions between international and domestic
stocks are beginning to fade. International equity investments can not provide suitable
diversification for a U.S. stock portfolio. Institutional investors are increasingly
integrating alternative investment assets into their strategic asset allocation in order to
achieve smooth and consistent performance by reducing their overall portfolio risk
without sacrificing expected return. Those alternative assets include commodities, hedge
funds, managed futures, private equity, and so on.
For a long time, commodities were considered as an inappropriate investment
category because of their substantial price variation. Recently, a number of studies have
confirmed that commodities show a unique risk/return character and can provide

efficient diversification benefit to traditional portfolio assets compared to other real


assets (Irwin and Landa [1987], Arkrim and Hensel [1993], Froot [1995]). In addition,
returns for commodities show positive correlation with unexpected inflation and exhibit
better performance under high inflation economic regimes (Georgiev [2001] and Gorton
and Rouwenhorst [2004]). Therefore, commodities can provide a good inflation hedge
for portfolio managers. For many investors, the question is no longer whether
commodities are an asset class, but what is the best exposure to this asset class. There are
four ways to obtain economic exposure to commodity assets: purchasing the underlying
commodity, owning the securities of commodity-based companies, trading commodity
futures and options contracts, and buying commodity-linked notes.
A commodity futures contract is an agreement to deliver or accept a specified
quantity of a commodity at a predetermined price at a designated time within a specified
time period. It is the first derivative traded in the world. Although commodity futures
account for a relatively small fraction of global futures and options trading at this stage,
they have been attracting the attention of some major hedge funds and other institutional
investors who sought to diversity their portfolios by investing in nontraditional assets
during the past decade. Commodity futures and option are increasing in trading activity,
especially metals and agricultural commodities.
Georgiev (2001) and Gorton and Rouwenhorst (2004) have shown that direct
commodity futures investment is the principal means by which institutional investors can
easily obtain exposure to commodity price movements. Commodity futures contracts are
traded on exchanges that share the same advantages as stock exchanges: transparent
pricing, daily liquidity, clearinghouse security, uniform contract size and terms, and a
central marketplace. Moreover, futures contracts do not need physical commodity
storage and delivery and can be very flexible to take either a long or short position with
just small margin deposits. The advantages of commodity futures make them attractive
for the institutional investors who are seeking an easy way to gain exposure to
commodities. However, commodities futures prices are subjective to high volatility,

which leads to one of the big disadvantages of investment in commodities futures, the
high investment risk. Price movements of commodity futures contracts are influenced by
supply and demand relationships, weather, government, agricultural trade, fiscal,
monetary and exchange control programs and policies, national and international
political and economic events, and changes in interest rates. Most of these factors are
unpredictable and subject to high variation. In addition, governments intervene directly
and by regulation in certain commodity markets. Such intervention is often intended to
influence prices directly. Meanwhile, trading commodity futures on margin involves a
high degree of risk. The liability of the account holder is not limited to the initial
investment or the equity in the program account.
Two methods can be used to hedge against the substantial risk of investing in an
individual futures contract. One way is to invest in a diversified basket of commodity
futures contracts. There are now a number of investable passive commodity futures
indexes which earn return from holding only long positions in unleveraged physical
commodity futures. The other way to hedge risk is through managed futures accounts.
Instead of passively holding long positions, managed futures accounts, which are
discretionarily controlled by commodity futures fund managers or commodity trade
advisors, actively use a technical trading system and professional skills to gain a
competitive return.
Commodity futures indexes are significantly different from managed futures
accounts in several ways. First, commodity futures indexes are designed to provide
passive exposure to the commodity futures markets while managed futures accounts are
actively traded with a skill-based strategy. Second, commodity futures indexes include
only physical commodity futures contracts, but managed futures accounts tend to invest
across the spectrum of futures markets, and financial futures contracts actually are
typically a large part of the portfolio (Irwin and Yoshimaru 1999). Third, commodity
futures indexes only take long positions; in contrast, managed futures accounts may
invest both long and short. Finally, commodity futures indexes are unleveraged while

managed futures accounts tend to apply leverage in the purchase and sale of commodity
futures contracts.
Empirical studies have concluded that passive commodity futures indexes show
not only good stand-alone performance but also substantial portfolio diversification
benefits (Abanomey and Mathur [2001], Ankrim and Hensel [1993], Anson [1999],
Becker and Finnerty [1994], Georgiev [2001], Gorton and Rouwenhorst [2004], Kaplan
and Lummer [1998], Johnson and Jensen [2001], and Jensen, Johnson and Mercer
[2002]). Adding commodity futures to a portfolio of stocks and bonds has the ability to
reduce the risk of the portfolio for a given level of return. However, prior empirical
research regarding managed futures is unsettled (Edwards and Ma [1988], Edwards and
Park [1996], Elton, Gruber, and Rentzler [1987], Irwin, Krukemyer, and Zulauf [1993],
Irwin and Brorsen [1985], Irwin, Zulauf, and Ward [1994], Schneeweis, Savanayanna,
and McCarthy [1991], Schneeweis, Spurgin, and McCarthy [1996]). There is no evidence
that public commodity pools provide any benefits either as a stand-alone investment or
as part of a diversified portfolio, although some studies have indicated that private
commodity pools and commodity trade advisors managed accounts can be valuable
additions to a diversified portfolio. In addition, the performance of managed futures
industry is inconsistent and unpredictable, which makes selecting properly managed
futures accounts impossible based on their historical performance.
This dissertation discusses portfolio performance of passive commodity futures
indexes instead of managed futures based on the following considerations. First,
commodity futures index consist only of physical commodities futures contracts, while
managed futures invest largely in finance futures (Irwin and Yoshimaru 1999). Finance
futures price movement is decided by the underlying assets such as stocks, bonds, and
interests rate. Adding financial futures into portfolio is a hedging issue instead of a
diversification issue. Second, managed futures, like hedge-funds, are actively trading
funds using skill-based strategy. Their performance is mostly decided by their
managers experience, trading skill, and inherent luck (Hartzmark 1991). This

dissertation plans to examine the diversification benefits achieved from the passive
addition of a new asset class such as commodity futures. Third, the data for commodity
futures indexes are much longer and more easily obtained. Most managed futures
accounts are not public, and their performance data are not available.
1.2 Research on Commodity Futures Investment
Greer (1978) studied the conservative use of commodity futures and found that
diversified, buy-and-hold, and unleveraged commodities were not as risky as stocks. An
unleveraged commodities index can serve as an inflation hedge for a stock portfolio. He
demonstrated that a combination of a commodity futures index and a stocks index
provided better performance with a high mean return for a given risk level. The optimal
relative position between these two asset classes was decided by the predicted inflation
rate.
Bodie (1983) examined how a diverse basket of commodity futures contracts
can supplement a portfolio of common stocks, bonds, and bills to improve the risk and
return trade-off based on the mean-variance optimization framework. Using historical
returns data from 1953 to 1981, he found that a broad-based position in commodity
futures tended to perform well when there was unanticipated inflation and provided a
substantial diversification benefit.
Irwin and Landa (1987) made use of the mean-variance optimization method to
compare the portfolio diversification effects of three alternative investment assets: real
estate, futures, and gold. They derived the optimal mean-variance portfolio allocation for
different risk levels and found that real estate should be held in substantial proportions at
lower portfolio risk levels. Both buy-and-hold commodity futures and futures funds were
held at lower risk levels. But buy-and-hold commodity futures holdings diminished
rapidly as portfolio risk increased, while futures funds holdings increased steadily. Gold
was not held in any efficient portfolios.

Arkrim and Hensel (1993) studied the inflation hedging effect and portfolio
improvement role of fully collateralized commodities futures using Goldman Sachs
Commodity Index (GSCI) and Investable Commodity Index (ICI) as proxy. Based on a
long term monthly return series from 1972 to 1990, commodity spot returns were
correlated positively with unexpected inflation, which negatively impact returns of
traditional financial assets. A mean-variance optimizer was used to calculate the efficient
portfolios for two risk-aversion levels and the optimization results showed that adding
commodity futures can generally reduce portfolio volatility without weakening returns.
They also examined the case for commodity futures using the reverse optimization
method which derives the required return for commodity futures to be included at the
efficient portfolio. The implied expected return necessary to hold 2.5% in commodity
futures is only 12 basis points over the risk-free asset, which is much lower than average
returns of commodity futures.
Becker and Finnerty (1994) examined the risk and return properties of
equity/bond portfolios before and after the inclusion of a diversified portfolio of long
commodity futures. Inclusion of commodities (using Commodity Research Bureau Index
(CRB) and GSCI as proxy) improved portfolio performance by enhancing the risk and
return characteristics. The improvement of the risk/return characteristic is more
substantial during the high inflation periods such as the decade of the 1970s than during
the low inflation period such as the decade of the 1980s. This result confirms that
commodity futures can serve as an inflation hedge.
Froot (1995) compared three classes of real assets: real estate, commodity
futures, and the stocks of companies that are commodity producers. He found that
commodity futures were a more effective hedging tool against unexpected inflation than
real estate and the stock of commodity producing companies. Commodity futures
rendered the other real assets ineffective when they were the initial hedge in a portfolio.
However, when they were added to the portfolio as the secondary hedge after other real
assets had already been added, commodity futures still received a significant portfolio

weight as a diversifier. The same conclusion did not hold for real estate and
commodity-based equity.
Satyanarayan and Varangis (1996) examined the diversification benefits of
commodity futures in a global portfolio context. They found that commodity futures
returns were correlated negatively with the returns to all developed markets and with
three of six emerging markets. The reason is that emerging markets tend to be net
suppliers of commodity inputs.
Kaplan and Lummer (1998) found that, in the long run, fully collateralized long
positions in GSCI futures contract not only provided good diversification for stock and
bond portfolio but also hedged against inflation risk; however, their portfolio benefits
could not be achieved over short periods of time. Over the long run, commodity
investment was correlated negatively with stocks and bonds investments and had better
performance during a rising inflation period than during a falling inflation period. But
over a short period such as March 1992 to February 1997, commodity futures showed
positive correlation with traditional assets. This result is in sharp contrast with preceding
empirical studies and leads to more careful thinking on commodity futures investment.
Anson (1999) examined the portfolio diversification role of four commodity
futures indexes: GSCI, Chase Physical Commodity Index (CPCI), ICI and JPMorgan
Commodity Index (JPMCI). Using quarterly returns from 1947 to 1997, he maximized
mean-variance utility over four asset classes: large-capitalization stocks S&P500,
small-capitalization stocks (NASDAQ), long-term bonds, and commodity futures. The
results showed that how the addition of commodity futures to a diversified portfolio was
determined by the investors level of relative risk aversion. Despite the popular
perception that commodity futures are too risky for the typical investment, he
demonstrated that the more risk averse the investor is, the higher his utility will be by
investing in commodity futures.
Gibson (1999) examined the rewards of multiple-asset-class investing in a

broader equity context. He found that a commodity futures index (GSCI) investment has
a pattern of returns that is the most dissimilar to the other asset classes (stocks, bonds,
and real estate) and accordingly produces the strongest diversification effect when
combined with other asset classes.
Georgiev (2001) investigated the relative risk and return advantages of direct
commodity futures investment using GSCI as a proxy. Investment in GSCI was shown to
result in a significant diversification benefit based on monthly returns data for a sample
period from January 1990 through December 2001. This benefit stemmed from the
unique exposure of commodities to market forces such as unexpected inflation as well as
the potential of positive roll return in periods of high spot price volatility. This
diversification benefit was beyond that achievable from commodity-based stock and
bond investment (indirect investment on commodities).
Using monthly returns of GSCI from 1973 to 1997, Jensen, Johnson and Mercer
(2000, 2002) examined the investment performance of commodity futures as a
stand-alone investment and as a portfolio component. In the overall sample period, the
stand-alone performance was poor for commodity futures since they had lower returns
and higher risk. In the portfolio context, the diversification benefit was significant for
commodity futures and mean-variance optimization yielded substantial capital allocation
to them. By dichotomizing the sample period into expansive-versus-restrictive monetary
environments using the most recent changes in the Federal Reserve discount rate as the
classifying criterion, the authors studied the investment performance of commodity
futures under different economic conditions. The results were dramatically different in
the two economic conditions. During restrictive monetary period, commodity futures
tended to display strong return/risk performance as a stand-alone asset and significantly
enhanced the portfolio performance. In contrast, during the expansive period, the
return/risk performance was very poor and no portfolio improvement was founded. The
authors also hypothesized that the sensitivity to Federal Reserve monetary policy is
likely to vary across commodities and expanded their investigation by examining the

performance of five GSCI sub-indexes. They found that the diversification benefit varied
across commodities with metals, energy, and agricultural futures providing quite high
returns during the restrictive period and livestock contracts performing much better
during the expansive period.
Nijman and Swinkels (2003) showed that commodity futures investments helped
pension funds in a mean-variance framework by shifting up the efficient frontier. They
tested the statistical significance of the shift for multiple investment horizons using the
regression analysis developed by Huberman & Kandel (1987). They concluded that
commodity investment could significantly reduce the risk of real pension funds, and the
timing strategies between commodities and stocks can improve the strategic
mean-variance frontier even further. The timing strategies were based on four macro
economic variables: the yield on 10-year government bonds, the term spread, the default
spread, and the inflation rate. Since there is some evidence that expected returns and
covariances change with current economic condition, the authors examined both
unconditional and conditional spanning of commodities by the traditional asset classes.
The results show that commodities can improve the mean-variance level both
conditionally and unconditionally for real pensions.
Gorton and Rouwenhorst (2004) studied return/risk properties of commodity
futures as an asset class by constructing an equally-weighted and fully-collateralized
index. Over a long-term period from July 1959 to March 2004, commodity futures have
offered the same average return and Sharpe ratio as equities and showed negative
correlation with equities and bonds but positive correlation with inflation, unexpected
inflation and changes in expected inflation. Those results strongly support that
commodity futures should enter the investment mainstream not only as a stand-alone
asset class but also as a portfolio component. In addition, the authors demonstrated that
the negative correlation between commodity futures and the other asset classes was
largely due to different behavior over the business cycle.

10

In conclusion, prior researchers present ample evidence about the benefits of


commodity

futures

investment.

Since

commodity

futures

are

negatively

or

non-significantly correlated with traditional assets, such as stocks and bonds, in


according with Markowitzs mean-variance theory, adding commodity futures into the
portfolio should improve the portfolio performance by reducing the total risk without
sacrificing expected return. In addition, commodity futures returns exhibit negative
correlation with inflation rates, which tend to reduce stock and bond returns. Commodity
futures can serve as a good inflation hedging tool for portfolio construction. However,
the portfolio performance improvement role of commodity futures is not consistent at all
times. The risk/return property of commodity futures will vary with the underlying
economic condition. To study this variation is the primary object of this dissertation.
1.3 Statement of Purpose and Prospects
The main purpose of this dissertation is to examine the portfolio performance of
commodity futures indexes under a dynamic regime switching economic environment. It
is hypothesized that the risk/return characteristics of commodity futures and financial
assets will be different under distinct economic regimes and moreover the economic
regimes are related to each other rather than totally independent. When portfolio
managers make a decision about asset allocations based on time-varying expected returns
and variations, they should consider current economic conditions and also take the
regime switching probability effects into account.
This dissertation will make at least four unique contributions to the existing
literature. First, instead of assuming that the distribution of asset return is identical
independent distributed (i.i.d.) normal; this paper applies a regime-switching model to
examine the structure break behavior existing in asset return dynamics and investigate
the relevance with asset allocation. Second, in addition to studying static portfolio
selection for a single period, this paper examines the multi-period portfolio allocation
problem with commodity futures under different economic conditions. Third, the Value

11

at Risk (VaR), a new risk measure is used to examine the risk reduction effect of
commodity futures. Traditional approaches use the variance of return distributions as the
risk measure. Finally, while traditional analyses examine the diversification benefits of
commodity futures under normal market conditions, this work extends the literature by
using the Extreme Value Theory (EVT) to examine the extreme market movements and
event risk for a portfolio including commodity futures and traditional assets.
Prior literature on the strategic benefits of commodity futures adopts the
mean-variance optimization framework and measures asset returns performance using
historical sample averages. The mean-variance model was developed by Markowitz
(1952) and Tobin (1958). The optimal solution of portfolio weights in their model is a
function on the first and second moments of the asset return distributions. According to
this model, making a portfolio will significantly reduce the total risk without sacrificing
expected return when the correlation among individual assets is negative or closes to
zero, which is the case for a portfolio of commodity futures and traditional assets.
Although the mean-variance model provides an elegant mathematical optimization
method to decide the efficient portfolio allocation, there are three important
shortcomings of this theory when applied to portfolio analysis.
First, the mean-variance model can only handle the static portfolio allocation
problem and the optimization procedure is based on historical average returns. The key
assumptions of this model are that asset returns are normally distributed and investment
risk is measured by the variance. Time-varying aspects of the assets performance are not
considered in this model. Those aspects include skewness, kurtosis, serial correlation,
and time-varying means and variances, all of which can only be examined using a
dynamic model.
Samuelson (1969) and Merton (1969, 1971) initiated multi-period portfolio
optimization study using stochastic dynamic control theory. Their continuous time
models show that if the investor has constant relative risk aversion and returns are

12

identical independent distributed, the optimal long-term portfolio is the myopic portfolio,
i.e. the portfolio optimal for the investor with a one period investment horizon. If returns
are not identical independent distributed, this is in general no longer true and the optimal
portfolio for the long-term investor that is allowed to rebalance will differ from the
myopic portfolio. This difference is called intertemporal hedging.
This dissertation introduces Ang and Bekaerts regime-switching portfolio
selection model to examine portfolio improvement role of commodity futures in both
one-period and multi-period context. In the one-period context, a simple (without
instrument) regime switching model is applied to model asset return dynamics. The asset
risk and return are distinguished in different regimes and the regime variable follows a
first-order Markov chain. For the one-period ahead optimal portfolio allocation, the
optimal asset weights derived from a mean-variance optimization are decided by the
regime expectation in the next period. In the multi-period context, asset return dynamics
are decided by a regime variable as well as an economic instrument. A dynamic
programming approach is adopted to estimate the optimal portfolio weights at each
investment horizon and under different economic conditions. Markov regime-switching
model is adopted to capture the investment performance of risky assets and the solution
of optimal weights is based on this special data generation process. The expected returns
and variation of risky assets vary through time and exhibit regime switching
characteristics, so the investment opportunity set is not constant over the investment
horizon and the optimal capital allocation to commodity futures should be dependent on
the underlying regime.
A second critical limitation to be addressed is that the mean-variance model
measures investment risk based on the variation of the return distribution. A number of
studies (see Balzer [1994] for an overview) have shown that the variance is only a
suitable measure of risk when the expected returns are normally distributed. Higher
moments should be considered when the distribution of asset returns is non-normal. In
addition, investors are mainly concerned about downside risk caused by adverse market

13

movement, instead of general volatility. Value at Risk (VaR), measuring the left (right)
tail risk for holding a long (short) market position at a given confidence level, can be
obtained for general distribution and is consistent with investors intuition. While VaR
has become a popular risk measurement tool in the investment world, no literature
appears to examine the risk reduction performance of commodity futures based on the
VaR risk measure. This study will fill in this gap.
Empirical evidence shows that the returns of commodity futures and financial
assets are not normally distributed and exhibit unusual levels of skewness and kurtosis.
(For example: Deaton and Laroque [1992], Yang and Brorsen [1992], Myers [994],
Vercammen [1995], Hilliard and Reis [1999], Wei and Leuthold [2000], Roberts [2001],
Kraus and Litzenberger [1976], Duffie and Pan [1997], Timmermann [2000]) As a
consequence, portfolio analysis based solely on mean and variance may be leading to
wrong conclusions and decisions. Amin and Kat (2003), Bacmann and Scholz (2003)
and Bacmann and Pache (2003) examined hedge funds diversification effect and showed
that while hedge funds combine well with stocks and bonds in the mean-variance
framework, this is no longer the case when skewness is considered. Alexander and
Baptista (2002) developed a mean-VaR model for portfolio selection instead of the
traditional mean-variance model and concluded that the optimal portfolio solution with a
mean-VaR model converges to the solution of a mean-variance model only when the
asset returns distribution is normal and the confidence level for the VaR measure is very
high. In 2003, they proposed a new investment performance measure called
reward-to-VaR ratio as a complement to the traditional Sharpe ratio. The reward-to-VaR
measure ranks portfolio performance differently from the Sharpe ratio under
non-normality. All of above discussed studies encourage us to examine the risk reduction
effects of commodity futures using VaR as risk measurement.
The third important weakness of the mean-variance model is that optimal
portfolio allocations are based on historical average performance of asset returns. Event
risk and market extreme movement are not considered. It is well known that financial

14

markets are subject to extreme variations, mostly because of financial turmoil, large
credit defaults, war, nature disaster, and political crisis. These market extreme
movements will result in substantial negative returns. Fund managers monitoring their
portfolio risk only based on historical average information might suffer a serous loss
when the market moves extremely in an adverse direction. A question is naturally raised:
while commodity futures show wonderful diversification effects under normal market
conditions, can they serve to hedge against event risk for institutional investors to
achieve smooth performance?
Previous studies have showed that commodities, in contrast with traditional
financial assets, tend to have positive exposure to event risk (Deaton and Laroque [1992],
Georgiev [2001], Anson [2002]). No statistical models have been designed to formally
test this issue. This dissertation fills in this gap by taking advantage of the Extreme
Value Theory (EVT) to describe the investment risk associated with extreme events for a
portfolio among stocks, bonds and commodity futures. The Extreme Value Theory
provides a very powerful tool to analyze the extreme tail distribution of a random
variable. Recently, more and more research has applied this tool to analyze the extreme
variations in financial markets. For example, Diebold, Schuermann, and Stroughair
(1998) provided an interesting discussion about how to use the EVT to measure the VaR
of equities investment; Bacmanna and Gawron (2004) examined the fat tail risk in
portfolios of hedge funds and traditional investments. This empirical study can provide
some statistical evidence about the diversification effect of commodity futures under
extreme market conditions.
To sum up, the conclusion of past study about the benefit of commodity futures
investment is questioned and disputable since it is based on a static model and an
improper risk measure. This dissertation provides some complementary studies by
extending static analysis to a dynamic context and by assessing portfolio risk with the
VaR measure. Markov regime-switching model is used to capture the time-varying
investment opportunity for stocks, bonds and commodity futures. The issue of the

15

diversification benefits and risk reduction of commodity futures is examined under


distinct economic conditions. Moreover, the EVT is adopted to examine the
diversification benefit under extreme market movement. This dissertation intends to
provide a comprehensive examination on commodity futures investment and supplies
more persuasive evidence on its diversification benefit to traditional portfolio.

16

Chapter 2 QUANTITATIVE PORTFOLIO SELECTION MODEL

2.1 Overview
Asset managers construct a portfolio of assets with many different single assets.
In general, there are two ways to construct a portfolio in the investment world: the
discretionary approach and the quantitative approach. The discretionary approach uses
fundamental analysis or technical (charts) analysis to predict the forward market
movement and asset price variation. Decisions concerning the selection of instruments
and asset allocation are made based on personal experience and analytical skills.
Personal insights play a large part in this portfolio construction process. The quantitative
approach uses statistical models, simulation techniques, and optimization processes to
discover the best asset selection. This method holds that the portfolio selection problem
can be expressed by a mathematic structure based on assumptions, parameters, input
variables, and output variables. Statistical models and the prediction of asset price
variations play the key roles for this portfolio approach. In the past two decades, the
rapid growing of financial markets, especially, the derivative market has complicated
investment analysis and more quantitative-oriented strategies. Meanwhile, the
development of new technology makes it possible to apply well-defined portfolio theory
and advanced statistic model to portfolio construction practice.
The inauguration of modern portfolio theory came with a static portfolio model
of Markowitz (1952) and Tobin (1958) that discussed how to construct an optimal
portfolio based on a mean-variance (M-V) framework. This normative theory was
extended by Sharpe (1964) and Lintner (1965) to a general equilibrium model of asset
prices, which is the famous Capital Asset Pricing Models (CAPM). Roll and Ross (1976)

17

criticized the usefulness of the CAPM and published the arbitrage pricing theory (APT)
as an alternative of CAPM.
The M-V, CAPM, and APT were all static portfolio theories, which basically
discussed the myopic portfolio selection under one period context. In multi-period case
when rebalancing is allowed, Samuelson (1969) and Merton (1969, 1971) apply
stochastic dynamic control theory to show that if the investor has constant relative risk
aversion and returns are i.i.d., the optimal long-term portfolio is the myopic portfolio, i.e.
the portfolio optimal for the investor with a one period investment horizon. If returns are
not i.i.d., in general, this is no longer true and the optimal portfolio for the long-term
investor that is allowed to rebalance will differ from the myopic portfolio. Recent
literature that investigates intertemporal hedging includes Brandt (1999) and Ang and
Bekaert (1999) who find the hedging demand to be relatively small whereas Brennan,
Schwartz and Lagnado (1997), Barberis (2000) and Campbell and Viceira (1999, 2000)
find the intertemporal hedging demand to be quite large.
Static portfolio theory is not new to agricultural economists. Johnson (1960)
discussed the optimal portfolios involving commodity inventory and short positions in
futures contracts using mean-variance theory. Agricultural economists have modified the
financial portfolio theory to study agricultural market issues such as optimal hedging
with futures and options, optimal crop selection, risk premium and so on. Tomek and
Peterson (2001) provided a comprehensive review of this field. This section provides an
introduction about the recently developed dynamic portfolio methods.

2.2 Static Portfolio Selection Theory


2.2.1 The Mean-Variance Model
In the 1950s, there was no specific measure for investment risk. Harry
Markowitz (1952) derived the expected rate of return for a portfolio of assets and an

18

expected risk measure. He derived the formula for computing the variance of a portfolio
and showed that variance was a meaningful measure of portfolio risk under a reasonable
assumption. The formula for the variance of a portfolio not only indicated the importance
of diversifying your investments to reduce the total risk of a portfolio, but also showed
how to effectively diversify.
Mathematic programming
We assume there are N investment assets available and T time periods. Let rit be
the one period return of asset i at time period t (i=1, , N: t=1,, T; the return could be
the continuously compounded return or the simple return, they are not very different
when one period is not long). Let the means

and variance of rit be

E (rit ) = it and Var (rit ) = it2 respectively, and let the covariance of rit and rjt be

Cov(rit , rjt ) = ijt . In general, it , it2 , and ijt are dependent on time t. The Markowitz
theory is a one-period theory with T=1 in which an investor at time 0 is supposed to
make an optimal portfolio among the N assets for time T=1 with respect to his objective
function based on predicted returns and variances. Hence in the static Markowitz theory,
the variability of it , it2 , and ijt over time is irrelevant since T=1. Thus we can leave
out the t at subscript. Let w=(w1, , wN) is the portfolio weight vector with denoting the
portfolio value ratio of the i-th asset. Then the portfolio return is:
R p ( w) = w r = w1 r1 + L + wN rN

(2.1)

The mean and variance of portfolio return are respectively given by:
N

p ( w) = wi i
i =1
N

( w) = wi w j ij
2
p

(2.2)

i =1 j =1

Hence if it , it2 , and ijt are known, then an investor at time 0 can optimize his
objective function basing on expected returns and variances:

19

max u ( p ( w), p2 ( w))


N

s.t.

w =1
i =1

(2.3)

a i w i bi
where ai and bi respectively denote the lower bound and upper bound for the portfolio
value ration of the i-th asset. Usually, the objective function is a utility function defined
as u = p ( w) p ( w) . denotes the investors risk aversion coefficient. This
mathematic programming problem is the basic part of the mean-variance theory and the
solution to it requires quadratic programming techniques because there are
second-order terms in the variance formula.
It is obvious that the mean-variance model is a two-step process. The first step is
to obtain the inputs to the model, i.e. the expected returns and covariance
matrix. it , it2 , and ijt . The second step is to solve the mathematic problems based on
the input variables. This model will always give the same solution when presented with
the same inputs and user parameter selections. Unless the inputs are changed or the
model formulation is altered, the model itself is not subject to intuition or emotional
reaction to current market conditions. Another advantage of this model is that a formal
structure is provided in the model to individually evaluate the different input variables
and risk factors. Sensitivity studies can be conducted by changing one input at a time and
inspecting the changing of the utility function value. This kind of analysis can give us
insight into the portfolio effect of an individual risk factor.
Portfolio effect

In general, making a portfolio will reduce the total risk of investment, which is
called the portfolio effect. The portfolio effect consists of two effects: the Markowitz
effect and the diversification effect. The Markowitz effect is the negative correlation
while the diversification effect is the uncorrelated effect. It is the principle of the
Markowitz effect that the variance of the portfolio with X and Y will be significantly
reduced and close to zero if the random return of two assets X and Y are strongly

20

negatively correlated. In fact, if XY = 1 , then the variance of a portfolio w1X+w2Y will


be zero if w1/w2= Y / X , where X , Y , XY are, respectively, the variances of X and
Y, and the correlation of X and Y. On the other hand, the diversification effect makes use
of the uncorrelated structure of the returns of assets. In fact, if individual assets return
r1,,rN are uncorrelated and their variances are bounded with a number M
( i2 M for all i=1,...N ,) then the variance of portfolio with equal weight wi =1/N

var( w1 x1 + K + wN xN ) = w12 12 + K + wN 2 N2 ( w12 + K + wN 2 ) M = M / N 2

(2.4)

which is very small with large N.


Figure 2-1: Efficient frontier

U3
U3
U2
U2

Risk

U1

Y
U3
X
U2
U1

Return

Efficient frontier

Another important feature of the mean-variance model is the efficient frontier.

21

Assuming numerous assets are available, we can derive the optimal portfolio with a
different combination of assets by solving the mathematic programming with different
value of the risk aversion coefficient . The efficient frontier is the envelop curve that
encompasses all of those optimal combinations. It defines as the set of portfolios those
having the highest return for each given level risk, or the minimum risk for each given
level of return. As Figure 2-1 shows, investors select the point of tangency between the
efficient frontier and the highest utility curve. Since the utility functions differ among
investors, each investors optimal portfolio point may differ from that of other investors.
Model assumption and limitation

Every quantitative model is conditioned on a set of assumptions, which most of


time is also a limitation for the model. The mean-variance model is based on several
assumptions regarding investor behavior:
1. Investors utility functions take quadratic form and are based on expected return
and variance (standard deviation). However, the quadratic utility function has
some undesirable characters, especially; the absolute risk aversion is an increasing
function on the investors wealth, which suggests the quadratic utility can only
represent investors desires over a restricted range of wealth.
2. Investors estimate the investment risk based on the variability of expected return. It
can be shown that variance is only a suitable measure of risk if the individual
expected returns are normally distributed. If the returns are not normally
distributed and the investor has a non-quadratic risk preference, then the
mean-variance model is not appropriate. Many authors have proved that the
distribution of financial assets return is skewed and has a fat tail. Duffie and Pan
(1997) provide an extensive overview about the fat tail issues. Deaton & Laroque
(1992), Myers (1994), and Yang & Brosen (1992) examined the distribution of
commodity prices and concluded that it is skewed to the right and displays
substantial kurtosis.

22

3. There exists a valid method to predict the expected return, variances and
co-variances. The typical method the model uses the simple historical average,
which is unreliable for prediction. In practice, no matter what nonparametric or
parametric methods are employed, measurement errors exist for the predictions
because of the random property of assets returns. Those measurement errors will
make the answer for the optimal portfolio wrong. Likewise, there is a serious curse
of dimensionality. For a portfolio with n investments, the number of independent
elements required for the mean-variance model including means, variances, and
covariance is (n)*(n+3)/3. Huge number of required input variables will make the
mean-variance model unworkable.
4. The portfolio optimization is static and based on a single time period. The very
nature of the time aspect of the asset performance history is not considered in this
model. Those natures include skewness, kurtosis, serial correlation of gain or loss,
and the time varying means and variances.
5. There are no transaction costs or other trading barriers. Then the optimal portfolio
based on the solution of the mathematic programming might be impossible because
of the significant transaction cost and trading rules.

2.2.2 The Capital Asset Pricing Model


Sharpe-Lintner-Mossin model

Following the development of the Markowitz model, Sharpe, Lintner and Mossin
independently derive a generalized theory of capital asset pricing under the market
equilibrium condition by considering the implication of assuming the existence of a
risk-free asset. It is often referred to as the Sharpe-Lintner-Mossin form of capital asset
pricing model. The basic logic for this model is simple. It proposes that there is a market
portfolio which includes all the market assets and a linear relationship exists between the
return on an asset and the return on the market portfolio. Then an assets return can be
described completely by a single factor, i.e. the market portfolio return, and the asset risk

23

can be also separated into a systematic part (related with market risk) and an
unsystematic part (independent with market risk). Investors are compensated for taking
on systematic risk but not for taking on the unsystematic one which can be diversified
away with a portfolio investment strategy.
Moreover, the CAPM considers the significant implications for the potential
return and risk and alternative risk-return combinations when a risk-free asset exists in
the market. The correlation and covariance of any risky asset with a risk-free asset are
zero, so that any combination of a risky asset or portfolio with the risk-free asset
generates a linear return and risk function. Therefore, combining the risk-free asset with
any point on the Markowitz efficient frontier will derive a set of straight-line portfolio
possibilities. The line tangent to the efficient frontier will dominant all other lines. This
dominant line is called capital market line (CML) and all investors should target along
this line by borrowing or lending money at the risk-free rate, which will depend on their
risk preferences. Tobin (1958) called this division of the investment decision from the
financing decision the separation theorem. Since all investors want to invest in the risky
portfolio at the point of tangency, this portfolio, i.e. the market portfolio, must contain all
risk assets in proportion to their relative market value.
Given the CML and the dominance of the market portfolio, the relevant risk
measure for an individual risky asset is the system risk which is defined by its covariance
with the market portfolio, that is, the beta. When this covariance is standardized and by
the market portfolio variance, we derive the well-known normalize beta measure of
system risk and a security market line (SML). Since all individual assets should plot on
this SML, we can determine their expected return based on their betas.
Single index model

Following the CAPM theory, Sharpe (1963) and Cohen and Pogue (1967)
develop a single-index model to estimate the asset return and risk. The single index
model assumes that the future performance of each investment in the portfolio can be

24

represented by a linear function on a common market index, which serves as the proxy of
the market portfolio.
Ri = i + i Rm + ei

2.5

Where
Ri = return of investment i
Rm = return of the market index--- a random variable

i = constant coefficient of investment is return which is independent of the


markets performance

i = a constant that measures the expected change in Ri given a change in Rm


ei = residual error of the linear equation fit of investment i --- a random variable
with zero mean
Note that both ei and Rm are random variables and they each have a probability
distribution and a mean and variance. Let ei and m respectively denote their standard
deviation.
Estimators of i , i , and ei are often obtained from time series-regression
analysis based on two general assumptions for the single index model. First, it is
assumed that ei is uncorrelated with Rm. This implies that ei represents the unsystematic
risk. Secondly, ei is assumed to be independent of ej for all the investments. This implies
that the only reason individual investments vary together, systematically, is because of a
common co-movement with the market. The unsystematic risk is independent with each
other.
The expected return of investment i expressed in terms of the performance of the
market index is:
E ( Ri ) = i + t E ( Rm ) = i + t Rm

(2.6)

25

where Rm is the expected market index return;


The variance of the investment is return is

2 = i2 m2 + ei2 , (2.7)
i

and the covariance of returns between investment i and j is

ij = i j m2

. (2.8)

Note that the expected return for investment i has two components: a constant part i
and a market-related part i Rm . Likewise, investment is variance also has unsystematic
part ei2 and systematic part i2 m2 . In contrast, the covariance depends only on the
market index risk. This is just what we have said earlier that the only reason individual
investment move together is a common response to market movements.
Given the single index model holds in practice, we can turn to the calculation of
the expected return and variance of any portfolio and derive the optimal portfolio
efficient frontier following the mean-variance framework. The expected return on any
portfolio is given by:
N

i =1

i =1

i =1

R p = E ( R p ) = Wi Ri = Wi i + Wi t Rm

(2.9)

where Wi is the portfolio ratio for investment i.


And the variance for a portfolio is given by:
N

i =1

i =1 j =1;
j i

i =1

i =1 j =1;
j i

p2 = Wi 2 i2 + WiW j ij = Wi 2 i2 m2 + WiW j i j m2 + Wi 2 ei2 (2.10)


i =1

There are many alternative ways to estimate the parameters of the single-index model.
It is clear that expected return and risk of a portfolio can be estimated if we have an
estimate of i , i , R m , m2 and ei . This is a total of 3N+2 estimates, which is much
less than that of the mean-variance model (N*(N+3)/2).
The single index model reduces the input variables for the mean-variance

26

problem and results in a faster solution. The main advantage of the single index model is
that it can easily handle as many as 1,000 investments in a portfolio. There are no other
special advantages to this approach. The single index model still follows the basic the
mean-variance approach and cannot solve the problems such as skewness and fat tail

2.2.3 The Arbitrage Pricing Theory

The arbitrage pricing theory extends the CAPM model by adding more
explanatory variablesthat is, by going from univariate to multivariate analysis. It is a
significantly more practical theory than the CAPM because it is very flexible and
essentially statistical in nature. The APT does not need the strong assumptions in
deriving the CAPM such as quadratic utility function and normally distributed security
returns. It simply attempts to characterize and predict future asset returns from a number
of factors. The APT proposes that the expected excess return on an asset is approximately
a linear function of the risk premiums on K systematic factors in the economy:
Ri = i1 * ( f1 R f ) + i 2 * ( f 2 R f ) +& K + ik * ( f k R f )

(2.11)

where
Ri is the expected excess return on asset i
Rf is the risk-free rate of return

ik is the sensitivity of asset i to risk factors k


fk is the return on risk factor k
(fk-Rf) is the risk premium (excess return) associated with risk factor k.
This equation simply states that the excess return on any asset can be explained
by the excess returns on a number of factors, multiplied by their respective sensitivity
coefficients with regard to the asset in question. Similarly to the CAPM, the APT is a
normative equilibrium theory. If all asset returns behaved according this equation, then
we would observe the empirical relation:

27

Ri = i + i1 * Z1 + i 2 * Z 2 +& K + iK * Z K + ei

(2.12)

where
Ri is the excess return on asset i

i is the non-factor-related or asset-specific return on asset i


ik is the sensitivity of asset i to risk factors k
Zk is the excess return on risk factor k
ei is the residual return on asset i .
By definition:
1. Residual variance of asset i equals ei2
2. Variance of index k equals k2
By construction and assumption:
1. Mean of ei equals zero (E(ei)=0) for all N assets
2. Covariance between factors l and k equals zero for all K indexes.
3. Covariance between the residual of asset i and factor k equals zero for all assets
and factors.
4. Covariance between the residual of asset i and the residual of asset j equals zero
(E(eiej)=0).
Assumption 4 implies that the only reason assets vary together is because of
common co-movement with the K factors. The expected return, variance and covariance
between securities when the multi-factor model describes the return structure are equal
to:
1. Expected return for asset i is
K

Ri = i + ik Z k

(2.13)

k =1

where Z k is the expected mean risk premium for the kth risk factor.
2. Variance of asset is return is

28

i2 = ik2 k2 + ei2

(2.14)

k =1

3. Covariance between asset i and j is


K

ij = ik lk k2

(2.15)

k =1

From above three equations, it is clear that the expected return and risk can be
estimated for any portfolio if we have the estimates of i , ik , and ei2 for each asset
and Z k and k2 for each risk factor. This is total of 2N+2K+KN estimates. This is larger
than the number of inputs required for single index model but considerably less than the
inputs needed when no simplifying structure was assumed. Notice that now analysts
must be able to estimate the responsiveness of each asset to several economic and market
influences.
Unlike the CAPM, the APT does not tell us exactly what the risk factors should
be. The choice of factors to include in the model is left completely to the analyst to
decide. There simply are no ex ante guidelines for selecting explanatory factors, and
since we can identify statistical relations between almost any set of variables, the choice
naturally becomes very subjective and difficult. In practice, the multi-factor model is
basically treated in these two ways:
1.

Factors Z1,Zk are explicit and fundamentals factors or macroeconomic factors


are used for regression analysis.

2. Factors Z1,Zk are implicit and only asset price or return data is analyzed through
the factor analysis approach.

2.3 Dynamic Portfolio Selection Theory

Empirical studies suggest that most of investors have a long-term investment


horizon and that asset returns show general historical autocorrelation. Then the question

29

arises whether the optimal multi-period portfolio has any relationship to optimal holding
in a single-period setting. Should the optimal portfolio allocations change as time passes?
Since Samuelson (1969) and Merton (1969, 1971, 1973), a number of authors have
developed different dynamic portfolio theories to deal with this question. The basic point
of departure for those dynamic portfolio theories is to propose that asset returns at each
period of the investment horizon are not constant but vary conditionally with a set of
exogenous state variables in according with some specified model. The future values of
the state variables are assumed to be given, then the moments of future return
distributions are assumed to be accurately known, and the optimal set of changing asset
allocations can be determined.
Dynamic portfolio optimization uses the power utility function of terminal wealth
U(WT) as the objective function:
WT 1
if >1

1
U (WT ) =
ln WT if =1

(2.16)

where T denotes the investment horizon, WT is the terminal wealth, and denotes the
coefficient of relative risk aversion. An investor who has a power utility function is
called a constant relative risk aversion (CRRA) investor because the power utility
function shows the property of constant relative risk aversion. The CRRA investor does
not care about the path taken to arrive at the terminal wealth, which implies that the
investor is not concerned about the volatility of the returns to arrive the terminal value.
Another common concept in dynamic portfolio models is the budget constraint
between one time period and next. The change in wealth between the two time periods is
a function of the allocations to the investments and the expected asset returns.
Considering portfolio growth and capital consumption, the value of the investors
portfolio with N assets at the end of time t+1 is

30

Wt +1 = (Wt Ct ) * X it Yit (2.17)


i =1

where W denotes the wealth, C is the capital consumption, Xi is the portfolio allocation
given to asset i, and Yi is the expected return for asset i. It is well-know result from
Samuelson (1969) and Merton (1969, 1971, 1973) that if the return distribution is
constant over the investment horizon, then optimal asset allocations will be constant
over the horizon and identical with the optimal portfolio in the single period context,
but if the asset returns are not independent and identically distributed, then the optimal
policy will be time varying and the dynamic portfolio policy is different from the
myopic portfolio policy. This difference is referred to as hedging demand.
The methodologies for solving dynamic portfolio optimization problem cover a
wide range of techniques. Both continuous-time and discrete-time approaches have been
proposed. Solution tools include dynamic programming, stochastic control process,
Monte Carlo simulation, and Markov regime switching process. Except for the
simulation model, none of these methods can handle the large numbers of assets and
state variables that arise in realistic portfolio choice problem. We will survey several
important approaches in very broad terms.
Mertons continuous-time stochastic dynamic programming

In a continuous-time setting, Merton develops a stochastic dynamic programming


formulation for capital growth and capital consumption. A complex stochastic partial
differential equation is derived to describe the optimal control process for changing the
asset allocations and the capital consumption rate over time. The investor maximizes his
utility function of consumption and terminal wealth by controlling the decision variables
of asset allocation wi (i=1,2,,N) and consumption C at each time period t.
T

V (W , X , t ) = max E0 U [C (t ), t ] dt + B [W (T ), T ]
t =0

(2.18)

where E0 is the conditional expectation operator, V is the objective value function, W is

31

the wealth, C(t) is the consumption at time t, U is the utility function for consumption,
T is the investment horizon, and B denotes the utility of terminal wealth or bequest
function. The elements of K state vector X are defined as a vector Ito process:
dX = A( X ) + B( X )dZ

(2.19)

where A(X) is the vector of Ito process time function [A1, A2,,AK], B(X) is the
diagonal matrix of Wiener process function [B1, B2,,BK], and dZ is the vector of
Wiener process elements [dZ1, dZ2,,dZK].
The optimality condition for the objective utility function V is given by the
partial different equation:

+
+
U
(
C
,
t
)
V
V

t
w wi ri + rf * W C +

i =1

N N
K N
1

2
0 = Max Vk Ak + Vww wi w j ijW + VkW wiWgk i ki (2.20)
C ,w
i =1 j =1
k =1 i =1
k =1 K K 2

+ 1

Vkl Bk Bh kh
2

k =1 l =1

where ri is the excess return for asset i, rf is the risk-free return, ij is the time varying
covariance between asset i and asset j, i is the standard deviation for asset i return,

ki is the covariance between state variable k and asset i, kh is the covariance


between state variable k and state variable h, and Vt is the notation for V / t , the rest
may be deduced by analogy.
This equation could be solved for wi and C by two first-order conditions and a
boundary conditions:
0 = U (C , t ) V (W , t , X )
N

0 = VW ri + VWW wiW ij + VkW Bk i ki


i =1

V (W , T , X ) = B(W , T )

. (2.21)

k =1 i =1

But actually solving the partial different equation in closed form is impossible under

32

most circumstances and this model is difficult to apply to realistic investment analysis.
Despite the mathematical difficulties, Merton manages to show that if the expected
return and the covariance matrix are constant over time, the asset allocations are also
constant and equal to the one-period solutions for a the mean-variance investor.
Brennan, Schwartz, and Lagnados stochastic control model

Employing a set of continuous differential equations, Brennan, Schwartz, and


Lagnado (1997) developed a stochastic optimal control model to study the multi-period
portfolio problem with three asset classes: stocks, long-term bonds, and cash. They
assume three state variables to control the process: the interest rate level, the stock
dividend yield, and the bond yield. The means and variances of the asset returns depend
on three conditional factors. A joint stochastic process is defined in which the
continuously rebalanced time change in the asset returns and the dividend yield are
functions of there exogenous factors as well as the returns of the other assets. The
coefficients for the joint stochastic process are estimated using nonlinear seemingly
unrelated regressions. Then the optimal control formulation is solved using a discrete
approximation grid for the continuous process.
Numerical results are obtained for a CRRA investor with different investment
horizons. They find that the investors time horizon has a significant effect on the
composition of the optimal portfolio. Investors with longer horizon case generally
commit more capital to stocks and long-term bonds. The authors point out that the reason
for this is the mean reversion in both bond and stock returns that makes these assets less
risky from the viewpoint of a long-term investor. Their results provide an empirical
evidence that hedging demand does exist and optimal portfolio for multi-period horizon
is different with that for single period investment. The out of sample simulation results
also support that the predictability of future asset returns is sufficient to yield significant
improvement in portfolio performance.
The critical limitation for this approach is that the solution times increase linearly

33

with the number of assets and geometrically with the number of state variables.
Transaction costs can not considered because it will enter as a new state variable,
generating a problem of unworkable size.
Brandts conditional Euler equation approach

Brandt (1999) develops a nonparametric Euler equation approach to examine


how portfolio and consumption choice depends on variables that forecast time-varying
investment opportunities. He studied a CRRA investor who allocates wealth to the NYSE
index and a 30-day Treasury bill. In order to avoid errors due to model misspecification,
a nonparametric method is proposed to formulate how asset returns relate to forecasting
variables or how investment opportunities change through time. Then portfolio asset
allocation is a nonparametric conditional function of the values of the selected state
variables. For each period of history, the state variables are observed and a sample
observation is generated. These sample observations are used to trace out the shape of
the function relating the state variables to the resulting asset allocations. To validate the
methodology, a solution is first obtained for the unconditional case without state
variables. The results for asset allocations are not statistically different across all
investment horizons for a given level of risk aversion. A conditional case is then
examined using the following state variables: the dividend yield, default premium, term
premium, and lagged excess return. The future values of the state variables are assume to
be known with certainty to determine the asset allocations over the forward time horizon.
Only one of the four state variables is used for each evaluation of the objective function.
The results show that the optimal decisions depend on the investors horizon and
rebalancing frequency. The optimal multi-period portfolio choice differs from the myopic
portfolio choice. Just as in Brennan, Schwartz, and Lagnado (1997), the number of state
variables causes geometric growth in the size of problem formulation.
Barberis dynamic programming with Bayesian approach

Barberis (2000) considered the problem that the relationship between future

34

values of the state variables and future values of asset returns is subjected to uncertainty.
There are actually two problems: the uncertainty of the future value of state variables and
the uncertainty of the relationship between the state variables and the future rates of
return. The first problem is state variable risk and the second one is a combination of
model risk and estimation risk. Barberis examined part of the problems by calculating
the impact of estimation risk on asset allocation over a range of horizons for a CRRA
investor.
Only two assets are considered: the NYSE index and a 30-day Treasury bill. Two
state variables are employed to predict the asset returns: the stock dividend yield and the
Treasury bill rate. Using a dynamic programming formulation for the asset allocations as
a function of the state variables, he considers two model formulations. The first one has
no estimation risk or the model coefficients affecting future returns, and the second one
is a Bayesian approach admitting uncertainty of the coefficients. Simulation techniques
are then employed to evaluate the distribution of return uncertainty.
Using the data from 1986 to 1995, Barberis finds that investor should increase
stock allocations at longer horizons due to the hedging demand effect as described in
Merton. But the portfolio allocation can be seriously misleading if the allocation
framework ignore the estimation risk. When this kind of risk is considered, the
long-horizon investors in general still allocate more to stock than the short-horizon
investors, but the difference is not as large. When the possibility of leaning better
estimates of the model coefficients is accounted for, the allocation to stocks reverses
course and decreases with the length of the horizon. For example, an investor who
rebalances annually, and who takes no account of estimation risk, allocates about 80
percent to stocks irrespective of the length of the horizon. If both coefficient uncertainty
and the possibility of learning are considered, the allocation to stocks is reduced to about
40 percent over a 10-year horizon.
Brandt, Goyal, and Santa-Claras simulation approach

35

Former discussed dynamic portfolio models can only handle limited state
variables. A simulation approach developed by Brandt, Goyal, and Santa-Clara (2001)
can circumvent the problem of treating many state variables. In this procedure, the
relationship between asset returns and the state variables can take any form. The
investors utility function need not to be CRRA. Any function such as maximizing the
Sharpe ratio, minimizing maximum drawdown, or minimizing the portfolio value at risk
can be accommodated. The only requirement is that the objective function must be four
times differentiable. Virtually, any feature of realistic portfolio evaluation can be
accommodated. This includes transaction cost, taxes, and intermediate capital
consumption. Additional constraints can easily be added to the system, such as
constraints on borrowing and short selling. The joint dynamic of the asset returns and the
state variables can also be evaluated in a Bayesian framework, with the result that state
variables uncertainty and return uncertainty can be specifically incorporated.
The simulation process follows several steps. First, the expected utility function
is defined as a Taylor series expansion and a closed-form solution can be obtained.
Fourth-order expansions are found useful to accommodate the effects of skewness and
kurtosis of asset returns. Secondly, a large number of random simulated investment paths
are generated following the predetermined rules involving the state variables and the
asset returns. For each simulated path, the optimal portfolio weights at each period are
recursively determined using dynamic programming to maximize the Taylor series utility
function. A regression analysis is conducted at teach time period on the moments of the
utility function to determine their dependence on the values of the state variables. This
approach is computation-intensive but flexible to handle a variety of objective functions
and investment return functions. It also obtains results similar to those of other dynamic
portfolio models when the model structure assumptions are the same.
Campbell, Chan, and Viceiras vector autoregressive process

Campbell, Chan, and Viceira (2003) proposed another procedure that can handle

36

a large number of investment and state variables. This approach assumes that asset
allocations are linear functions of the state variables. The assets include a value-weighted
index on the combination of the NYSE, NASDAQ, and AMEX as a proxy for stock
returns, a 5-year Treasury bond return, and a 90-day Treasury bill rate. The exogenous
state variables employed include the Treasury bill yield, stock dividend yield, and the
term spread between a 5-year zero coupon bond and the Treasury bill rate. No borrowing
and short selling restrictions are included in the formulation. Transaction costs are not
considered.
The model assumes that future values of the state variables follow a first-order
vector autoregressive (VAR) process. The VAR estimates obtained are based on quarterly
data from 1952 to 1999. Each exogenous state variable only relates significantly to its
own previous value. Other variables do not help much. Using numerical procedures to
solve a set of nonlinear equations resulting from assumptions about asset allocation, the
optimal portfolio set is determined at each period. The asset allocation to each asset can
be split at any time into components due to the myopic allocation and due to hedge
demand allocation. This is useful for evaluating how much hedging demand has to do
with terminal wealth.

2.4 Markov Regime Switching Portfolio Selection Model

Ang and Bekaert (1999) develop a regime switching process to study dynamic
portfolio choice problem for international investment diversification. They model the US,
UK and German equity returns using a regime-switching model and examine the time
varying investment opportunity set which may be characterized by correlations and
volatilities that increase in bad times. In most of dynamic portfolio models,
time-variation in expected returns characterizes the changes in the investment
opportunity set and the time-variation is captured by a linear function of the state
variables. In contrast, the expected returns, volatilities and correlations vary with the

37

regime, rather than with state variables, in Ang and Bekaerts benchmark model.
Moreover, they also combine predictability by state variables with regime switches in the
data-generating process (DGP).

Their regime switching DGP produces a normal

regime with low correlations, low volatilities and a bear regime with higher
correlations, higher volatilities and lower conditional means. Both regimes are persistent
and such persistence cannot be captured by transitory jumps independent of equity
returns. International diversification is still valuable with regime changes and currency
hedging imparts further benefit.
To characterize uncertainty in the portfolio allocations resulting from uncertainty
in the parameters of the DGP, they use the delta-method from a classical econometric
perspective. This approach can formally test for the presence of inter-temporal hedging
demands (the difference between the investors one-period ahead and long-horizon
portfolio choice), for the presence of regime-dependent asset allocation for investors
with different horizons, and for the statistical significance of international diversification.
They also attempt to quantify and contrast the utility cost using the certainty equivalent
notion of ignoring the regime switching effect and of disregarding the inter-temporal
hedging. The empirical results show that the costs of ignoring regime switching may be
small or large depending on the presence of a conditionally risk-free asset. Another
important finding is that the intertemporal hedging demands under regime switches are
economically negligible and statistically insignificant. This result holds even with a
conditionally risk-free asset and when the short rate predicts equity returns. Investors
have little to lose by acting myopically instead of solving a more complex dynamic
programming problem for horizons greater than one period.
Dynamic programming model for portfolio selection

The dynamic model concerns the portfolio choice at time t for a constant relative
risk aversion (CRRA) investor who maximizes the expected utility of end-of-investment
horizon wealth by rebalancing the portfolio with N assets at the beginning of each month.

38

Formally, the investors problem is:


Vt (Wt , Z t ) = max Et [U (WT )]
t ,L, aT 1

(2.22)

subject to the sequence of budget constraints:


N

Ws +1 = Ws Rs +1 ( s ) = Ws si exp(rsi+1 )

s t

(2.23)

i =1

where t = ( t1 ,L , tN ) denotes the portfolio Weights vector at time period t, Wt


denotes the accumulated wealth at time t, Zt is the information set available at time t,
and rti is the continuously compounded return on asset i between t-1 and t. The
notation Et [] indicates that the expectation is taken with respect to the information
available at time t. The objective function is a power utility function of terminal wealth
U(WT) defined as:
WT 1
if >1

U (WT ) =
ln WT if =1

(2.24)

where T denotes the investment horizon, WT is the terminal wealth, and denotes the
coefficient of relative risk aversion. An investor who has a power utility function is
called a constant relative risk aversion (CRRA) investor because the power utility
function shows the property of constant relative risk aversion. The CRRA investor does
not care about the path taken to arrive at the terminal Wealth, which implies that the
investor is not concerned about the volatility of the returns to arrive at the terminal
value (see Campbell and Viceira [2002] and Elton et al. [2003] for detailed discussion
about the utility function).
The multi-period portfolio choice is a discrete time and continuous state dynamic
programming problem. At time t, the investor selects the optimal portfolio weights
conditional on the accumulated wealth Wt and the information set Zt at time t. When the

39

investor makes this decision, he will take into account the fact that at every future period
s ( s t ) the portfolio weights will be optimally revised conditional on the
then-available wealth Ws and information set Zs. This dynamic nature of the multi-period
portfolio choice is best understood by expressing the optimal problem in a single-period
setting as:
Vt (Wt , Z t ) = max Et [u (WT )]
t ,LT 1

= max Et { max Et +1 [u (WT )]}


t

t +1 ,LT 1

(2.25)

= max Et {Vt +1 (Wt i t exp(r i t +1 ), Z t +1 }


t

i =1

The second equation follows from the law of iterated expectations. The third
equality, which is the so-called Bellman equation, is the basis for any recursive solution
of the dynamic portfolio choice problem. The first-order-conditions (FOCs) for the
optimal solution at each time are:
N

Et [ jVt +1 (Wt i t exp(rti+1 ), Z t +1 ) exp(r j ) = 0

(2.26)

i =1

where jV is the first derivative on the portfolio weight on asset j for j=1, 2, , N. In
order to derive the explicit optimal solution, we need to specify the value function
Vt(Wt, Zt). Let Qt +1,T = Et +1 [( RT ( T* 1 )L Rt + 2 ( t*+1 )]1 and QT,T=1. where t* means the
optimal weights vector at time period t. Using the restriction that the weights must sum
to unity, Rt+1 can be rewritten as:
N 1

N 1

i =1

i =1

Rt +1 ( t ) = ti exp(r i t +1 ) + (1 ti ) exp(rt +N1 )

(2.27)

Then:
Vt (Wt , Z t ) = max Et [u (WT )]
t +1 ,LT 1

= max Et [Qt +1,T Wt1+1 ]


t

. (2.28)

= max Et [Qt +1Wt1 Rt1+1 ( t )]


t

N 1

N 1

i =1

i =1

= max Et [Qt +1Wt1 ( ti exp(r i t +1 ) + (1 ti ) exp(rt +N1 ))1 ]


t

40

Taking the first derivative of the value function with respect to t1 ,L , tN 1 , the N-1
first order conditions are:
exp(r1t +1 ) exp(r N t +1 )
= 0 . (2.29)
Et Qt +1 R ( t )
M
exp(r N 1 ) exp(r N )
t +1
t +1

t +1

These FOCs make up a system of nonlinear equation involving integrals that can,
in general, be solved for t* only numerically. Note that t* has effectively N-1
degrees of freedom, as the weight in the Nth asset will make the portfolio weights sum to
one.
Introducing regime switching

Suppose there are K regimes and the regimes are known by the investors at time t.
Let st denotes the regime variable at time t. The optimal weights now become functions
of the regime variables st, t* = t* ( st ) . Moreover, the investor will want to hedge
against the regime switches. The random variable Qt,T is also dependent on the value of st
now. The FOCs for a particular regime j at time t is then the weighted average of the
regime dependent first order conditions. The Weights are the probabilities that the
underlying switching process will move from regime j at time t to regime k = 1, ...,K at
time t+1. If these probabilities are denoted by Pjkt, then the first order conditions under
regime switching can be written:

exp(rk1,t +1 ) exp(rkN,t +1 )

Q
= 0 . (2.30)

p
*
E
*
R
(
)
*
|
s
j

M
=

jkt
t
k ,t +1
j ,t
k ,t +1,T

exp(r N 1 ) exp(r N ) t
k =1
k ,t +1
k ,t +1

The backward induction algorithm can be used to derive the value of Qk, t+1,T. Define
Qi,T-1,T given sT-1=i as:

41

Qi,T-1,T = Pij ,T 1 Ri1,T ( i*,T 1 )


j =1
K

= Pij ,T 1 (exp(rj ,T ) ' i*,T 1 )1

(2.31)

j =1

and Qi,T-2,T given sT-2=i as


K

Qi,T-2,T = Pij ,T 2 Q j ,T 1,T Ri1,T ( i*,T 2 )


j =1
K

= Pij ,T 2 Q j ,T 1,T (exp(rj ,T 1 )


'
j =1

*
1
i ,T 2

. (2.32)

Continue this process for Qi,T-3,T and so on. The portfolio optimization problem is
solved conditional on the current regime. Thus, the total number of weights calculated
is (N-1)*K.
There is no closed form solution for the optimal weights and the FOCs must be
solved numerically. The expectations involved can be calculated by numerical quadrature
method. Formally, a quadrature rule to calculate the expectation of the function f (x) is
defined by:
I

h( x) f ( x)dx h(xi )wi

(2.33)

i =1

where f(x) is some density function for x, the (x1xI) and (w1wI) are the abscissa
and weights, respectively. Based on this numerical method, the FOCs under regime
switching can be approximated by:
K

P
k =1

jk ,t

Qk ,t +1,T

exp(rki1 ,t +1 ) exp(rkiN,t +1 )
R ( )

M
wi ,t +1 = 0
ki ,t +1
jt

N
1
N

i =1

exp(rki ,t +1 ) exp(rki ,t +1 )

(2.34)

which can be solved by a non-linear equation solver.

2.5 Portfolio Analysis with Value at Risk

Value at Risk, a new method to measure investment risk, has started a risk
management revolution in the last few years. The concept of VaR and the methodologies

42

for computing it were developed in response to the financial disasters by the large
commercial and investment banks during the late 1980s. VaR has become the standard
risk measurement and risk control tool in financial markets and has been widely used by
derivates dealers and most investment institutions. VaR can integrate market risk across
all assets classes, stocks, bonds, derivatives, and commodities. Furthermore, VaR can be
adapted to account for credit risk, liquidity risk, and operational risk.
Overview

The phrase value-at-risk first came into wide usage following its appearance in
the Group of Thirty report released in July 1993 and the release of the first version of
RiskMetrics in October 1994. Since then this approach has become popular among
practitioners, the academic community and most significantly among central banks.
A more detailed description of the historical development of VaR can be found in Duffie
and Pan [1997] and Linsmeier and Pearson [1996]. The RiskMetrics initiative of
J.P.Morgan played a very positive role, and made VaR a widely used risk management
tool. Since VaR focuses on downside risk and is usually reported in dollars or returns, it
is often considered easier to understand by managers and outside investors that may not
be well versed in statistical methods. VaR is commonly used for internal risk
management purposes and is further being touted for use in risk management decision
making by non-financial firms (Jorion, 2000; JP Morgan Risk Metrics).
Regulators also have shown great interests in VaR. The prudent regulation of
financial institutions requires the maintenance of minimum levels of capital as reserves
against financial risks. In the wake of several financial disasters involving the trading of
derivatives products, such as the Barrings Bank collapse, regulatory agencies such as the
Basel Committee on Banking Supervision, the U.S. Federal Reserve Bank, and the
Securities and Exchange Commission embrace VaR as a transparent measure of
downside market risk that could be useful in reporting risks associated with portfolios of
highly market-sensitive assets such as derivatives (Linsmeier and Pearson, 1997).

43

Using VaR to manage and assess market risk requires a robust method to
precisely compute and forecast time-varying VaR under substantially oscillatory markets.
The well established techniques on volatility forecasting have been adopted and modified
to estimate VaR. For instance, historical volatility estimates (Hendricks [1996], Jorion
[2000]), implied volatility from options prices (Jorion 2000, Hopper [1996], Mahoney
[1996]), various ARCH/GARCH models (Jorion [2000], JP Morgan Risk Metrics [1995],
Hendricks [1996], Hopper [1996], Duffine and Pan [1997], Brooks et al [2002]), regime
switching models (Duffie and Pan [1997], Venkataraman [1997]), and stochastic model
(Duffie and Pan [1997]) have all been suggested for use in the computation of VaR.
Many new techniques have also developed to provide well-calibrated estimates of VaR.
These methods include full Monte-Carlo simulation (Jorion, [2000]; Linsmeier and
Perarson, [1996, 1997]), bootstrapping methods (Jorion [2000], Duffie and Pan [1997],
Sortino and Forsey [1996]), extreme value theory (Diebold et al [1998], Embrechts et al
[1997]), regression quantile method called CAViaR (Engle and Manganelli [2000]), and
semiparametric estimation (Fan and Gu 2002). The accuracy of various VaR estimates
was compared and studied by Beder (1995) and Dave and Stahl (1997).
Most of studies in the VaR literature focus on the computation of the VaR for
financial assets such as stocks and bonds. Introducing VaR into agricultural market risk
management is a relatively new trend. Boehlje and Lins (1998) and Gloy and Baker
(2001) allude to its potential for applications in agribusiness. Manfredo and Leuthold
(1999) firstly provided a systematic review on VaR. They pointed out that agricultural
markets provide a unique laboratory for exploring the VaR technique. For example, VaR
could apply to agricultural risk management in a realistic portfolio context such as the
multi-product hedge ratio, the soybean crushing margin, and the cattle feeding margin.
VaR could be beneficial in making hedging decisions, managing cash flows, setting
position limits, and constructing portfolio allocation for an agricultural business company.
As well, Manfredo and Leuthold (2001) calculate VaR measures using alternative
estimation techniques to quantify the market risk of cattle feeders. Odening and Hinrichs

44

(2002) examined problems that may occur when conventional VaR estimators are used to
quantify market risks of feeder pigs, finished hogs, and hog finishing margins at a
German hog market. They found that VaR estimators based on Extreme Value Theory is
a useful complement to traditional VaR methods, especially for long-term VaR forecasts.
There have been some VaR applications in agricultural economics. The Ag-Risk
program developed by the Ohio State University and the University of Illinois at
Urbana-Champaign uses VaR analysis to determine the potential downside revenue for
planting corn, wheat, and soybeans. Publicly traded agribusiness companies must comply
with SEC regulations concerning the reporting of positions in highly market sensitive
assets including spot commodity, futures, and options positions. Commodity Futures
Trading Commission (CFTC) proposed that VaR should be used to as a risk reporting
measure for firms to disclose their market risk exposure at agricultural trade options. The
Chicago Mercantile Exchange (CME) uses the Standard Portfolio Analysis of Risk
(SPAN) system to calculate collateral requirements on the basis of overall portfolio risk
using VaR technique. SPAN has become widely used by futures and options exchanges
as a mechanism to set margin requirements.
Large institutional investors, like commodity futures funds, can use the different
VaR measures to monitor the risk of their long or short position on commodity futures
contracts. Brorsen and Irwin (1987)s survey shows that commodity futures funds
managers allocated their capital basing on variances of futures returns. Since the
distribution of futures price change has been proved to be non-normal and asymmetrical
with fat tails, traditional risk measures such as variance will underestimate the market
risk for commodity futures investments. Using VaR, which deals with the tail behavior
and downside risk, to allocate their capital, commodity futures funds managers may
achieve better investment performance by increasing trading returns and maintaining the
investment risk level.
Theoretical constructs of VaR

45

Jorion (2000) defined VaR as the predicted worst-case loss at a specific


confidence level (e.g., 95%) over a certain period of time (e.g., 1 day). For example, a
VaR measure of 1 million dollars at the 5% risk level (95% confidence level) means that
overall losses would not exceed 1 million dollars more than 5% of the time over a given
holding period under normal market conditions. Based on a firm scientific foundation,
VaR provides users with a summary measure of market risk using a single number.
Assume the initial investment is W0 and R is the rate of return on the portfolio.
The portfolio value at the end of target horizon is WT=W0(1+R). The expected return and
volatility of R are and . At the given confidence level c, the worst possible return is
R* and W*=W0(1+R*) is the lowest portfolio value. For a general distribution of future
portfolio value f (W ) , Jorion (2000) defines VaR as the worst possible realization W*
such that the probability of exceeding this value is c:
c=

W*

f (W )dW

(2.35)

or such that the probability of a value lower than W*, p = P(W W * ) , is 1-c:
W*

1 c =

f (W )dW = P(W W * ) = p

(2.36)

The number of W* is called the quantile of the distribution, which is the cut-off value
with a fixed probability of being exceeded.
The major advantages of VaR are its easy-to-understand definition and the fact
that it aggregates the likely loss of a portfolio of assets into one number. Standard risk
criteria like stochastic dominance or certainty equivalents rely on the entire distribution.
Opposed to that, VaR considers just the left tail of the distribution. That means risk is
viewed as bad outcome. Moreover VaR is easy to calculate and it is not difficult to
include multiple uncertain market factors (e.g. commodity prices, futures prices, or

46

interest rates) into the analysis. However, the implication of this indicator for risk
management is not straightforward. The choice of the confidence level is somewhat
arbitrary, and in general, consistency with the expected utility theory is not guaranteed.
VaR quantifies the probability that a loss exceeds a certain level, but the magnitude of
such a loss is not specified (Harlow, 1991). Another major challenge for VaR analysis is
that it is hard to specify the probability distribution of extreme returns used in the
calculation of the VaR estimate. Any statistical method used for VaR estimation has to
have the prediction of tail events as its primary goal.
Standard methods of VaR calculation

There are the two basic types of methods for VaR estimation: parametric and
non-parametric. Parametric methods will include the variance-covariance approach and
some analytical methods. The non-parametric model includes historical simulation and
the Monte-Carlo approach. All methods have specific advantages and disadvantages. A
detailed treatment of these methods can be found in Jorion (2000) and Dowd (1998).
Linsmeier and Pearson (1996) and Manfredo and Leuthold (1999) provided a brief
discussion about the pros and cons of these estimation methods.
Variance- covariance Method
The variance-covariance method is a parametric method, based on the assumption
that the asset returns are normally distributed. Historical data is used to measure the
major parameters: means, standard deviations, correlations. If the market value of the
portfolio is a linear function of the underlying assets, the distribution of the portfolios
returns is normal as well. Therefore, the 5% quantile corresponding to VaR can be
calculated at 1.65 below the mean (2.33 will give the 1% level), and the VaR at 5%
risk level can be determined as:
VaR = Wt0 *1.65 * h

(2.37)

Where Wt0 is the initial Wealth level, is the standard deviation and h is time horizon.

47

For a portfolio consisting of N assets, the volatility of the portfolio return is calculated
by:

p =

w w
i =1 j =1

ij

(2.38)

where wi and wj are the weights of assets i and j and

ij

is the covariance between

them.
One significant advantage of this scheme is that for many market parameters all
of the relevant data are well known. J.P.Morgans RiskMetrics2 is probably the best
source for this type of data in many markets. Another strong side of this approach is that
it is flexible, simple and widely used. It also enables the addition of specific scenarios
(such as time varying volatility) and enables the analysis of the sensitivity of the results
with respect to the parameters.
However, this method relies heavily on the important assumption that all of the
asset returns are normally distributed. Historical distributions of market returns are far
from being normal. This problem is well known and is related to skewness and fat tails
(kurtosis). In that case the standard variance-covariance method will underestimate the
VaR for high confidence levels. One can find several techniques of how to deal with
these tails in the paper by Duffie and Pan [1997]. Even having solved this problem, one
must remember that the direct usage of the variance-covariance method for wide
portfolios is restricted to simple linear instruments. Therefore when a significant portion
of the portfolio is not linear (with options, for example) this method can not be used
directly.
Monte Carlo simulations
Monte-Carlo simulation is probably one of the most popular methods. It does not
require a specific form of the distributions for asset returns. Instead, it builds a joint
distribution of market factors based on one of the following: historical data; data

48

implicitly implied by observed prices; data based on specific economic scenarios. Then
the simulation of random paths for market factors is performed, typically with a large
number of scenarios. The profit and losses of the portfolio at the end of a desired forecast
period are measured for each scenario. These numbers should be ordered, and the 5%
quantile or 1% quantile of the worst results is the VaR estimate.
This method has several important advantages. First, it can handle different return
distributions. This method does not assume a specific model and can be easily adjusted
to economic forecasts. The results can be improved by taking a larger number of
simulated scenarios. Options and other nonlinear instruments can be easily included in a
portfolio. In addition, one can track path-dependence because the whole market process
is simulated rather then the final result alone. One can easily perform stress testing on the
Monte Carlo simulation or perform a more detailed analysis of a specific set of scenarios.
One important disadvantage of this method is very slow convergence. There is a
high cost of computation for a complex portfolio VaR analysis. Any Monte Carlo type
simulation converges to the true value as1/ n , where n is the total number of simulated
trajectories. One must perform 10000 times more simulations to increase the precision by
a factor of 100. Variance reduction methods have been proposed to solve this problem.
One of the typical techniques is to reduce the parameters using known properties of the
portfolio, such as correlations between some markets and securities, or known analytical
approximations to options and fixed income instruments. Another class of useful
techniques for speeding the standard Monte Carlo approach is portfolio compression, by
which one can represent a large portfolio of similar instruments as a single artificial
instrument with risk characteristics similar to the original portfolio. This bunching
requires that all of the instruments that we wish to unify have very similar risk
characteristics.
Historical simulations

49

Historical simulation resembles the Monte Carlo simulation regarding the


simulation procedure. The difference is that there is no explicit assumption of the asset
returns distribution. Instead, one observes the historical behavior of current portfolio
over the last few years and measures the daily percentage changes in the portfolio value,
and then applies these changes to current portfolio and measures the corresponding
profits and losses. Thus VaR estimates are derived from the empirical profit-and-loss
distribution and no explicit assumption about the return distribution is required.
This approach is relatively simple and it does not require distribution assumption
and the development of an analytical model. Moreover it can easily incorporate
non-linear instruments such as options. A typical problem with this approach is how to
choose the right historical sample period as the simulation base. It is true that the further
one goes into the past for data, the more data are available, but the less relevant this
information is to todays market. This is not a simple trade-off. On the one hand, one
would like to have more data to observe the rare events, which will lead to heavy losses.
On the other hand, one does not want to build current risk estimates on very old market
data. Assume the last five years of data are used for VaR estimate. If there was a big loss
on a particular day, and then exactly five years later the big market jump does not appear.
This will lead to a jump in the dairy VaR estimate from one day to the next. This example
demonstrates that the results are not stable when using the historical simulations
approach.

50

Chapter 3 COMMODITY FUTURES INDEX INVESTMENT

This chapter presents the basic facts about investing on fully-collateralized


commodity futures using the historical data over the period from January 1970 to
November 2004. It begins with a brief introduction of commodity futures index
investment, followed by a discussion on stand-along investment performance and
relative performance comparing with stocks and bonds. Then the portfolio benefit of
commodity futures investment is investigated under different inflation environment and
business cycle phases. Lastly, an examination of different event risk exposure between
commodity futures and stocks and bonds is presented.

3.1 Investing in Commodity Futures Index

The commodity futures referred to in this dissertation is an investment in a fully


collateralized commodity futures index. This represents passive, unleveraged long-only
positions in a broad-based commodity futures index. In recent years, a variety of passive
investable commodity indexes have been created to provide direct access to commodity
investment. 1 These include exchange-traded commodity indexes such as the Goldman
Sachs Commodity Index (GSCI) and the Commodity Research Bureau (CRB) and other
commodity-linked investments such as the Chase Physical Commodity Index, Dow
Jones Commodity indices, as well as more exotic, futures-based products such as the
Mount Lucas Management Index (MLMI).
Most of the commodity indexes share all the benefits of a stock index: they are
transparent, liquid, and investable. Even if an investor, such as a pension fund, cannot

Schneeweis and Spurgin (1997) have provided a detailed discussion on the different commodity futures indexes
and a comparison of those indexes.

51

directly invest into commodity futures indexes, she may still gain exposure through a
commodity-linked note. 2 Commodity futures indexes can be used by institutional
investors in two ways. First, an investor can use a commodity futures index to obtain a
different return nature from that of equity or bond investment. Second, an investor can
fulfill his strategic purpose to diversify the investment portfolios risk and return by
adding commodity futures index into the portfolio. This dissertation explores the
fundamental reasons as well as other factors that form the basis of commodity future
investment.
The GSCI has been selected as the proxy of commodity futures investment for
this study. The GSCI is an index of spot prices of nearest futures contracts for a basket of
commodities representing five commodity sectors such as energy, metals, livestock,
precious metals, and agricultural products. Currently, the index comprises 24
commodities which are chosen from the most liquid markets. Table 3-1 displays a list of
the 24 commodities with their dollar weights in the GSCI and the sub-index weights.
Each commodity in the index is weighted by their average share of world output in the
past five years. Thus, commodities with a greater share of the world output receive a
larger value in the index. The weights are updated annually to ensure that the index
reflects the most recent economic activities. The GSCI was introduced in 1991 by
Goldman Sachs Company.
The main reason why the GSCI is used here is that this index is very popular in
the investment world. Most institutional investors use the GSCI to gain exposure to
commodity assets. Moreover, a long historical data of GSCI return is available and this
index also provides five commodity groups data. Most of the early commodity indices,
e.g. the exchange-traded Commodity Research Bureau (CRB) or the Dow Jones
Commodity indices, are not investable. As pointed out by Schneeweis and Spurgin
(1997), the method for determining commodity weights for the CRB Index is too
2

Commodity-linked notes are financial engineering innovative products. The value of a commodity-linked note is
tied to the prices of the specified commodity futures contracts or commodity options.

52

cumbersome, while the Dow Jones Commodity Index requires daily rebalancing of each
position, making it expensive to replicate. In addition, neither of these indices uses
production weights to reflect the relative significance of each commodity to global
output. Not surprisingly, futures and options contracts based on these indices are not
actively traded. The Goldman Sachs Commodity Index (GSCI), as a broad-based and
production-weighted commodity index, has achieved a great success for commodity
investment. In 1992, futures contract on GSCI started trading on the Chicago Board of
Trade (CBOT).
The return of commodity futures index

The returns from unleveraged fully-collateralized commodity futures investment


can be broken down into three components:
i.

Collateral Yield: interest earned on the collateral backing the futures contracts,

ii.

Future contract price change,

iii.

Roll Yield.

A commodity futures index is unleveraged. The futures positions are fully collateralized
by deposited T-Bills. The interest earned on the T-Bill used as collateral is called the
collateral yield 3, which could be a significant part of the total return to a commodity
futures index. Meanwhile, the return of the T-Bill will reflect the expected changes in
inflation rate. This is the way for a commodity futures index to hedge against
anticipated inflation. The future contract price change is the return coming from the
market movement of the underlying future contract. It shows high variation and debate
surrounds whether the mean of this return is zero or positive. The risk premium theory
concludes that the mean return should be positive as hedging traders transfer risk to
speculators who can only take the opposite position of hedgers when there is a risk
premium. However, arbitrage theory says that the mean return should reduce to zero as
3

In fact, an unleveraged commodity futures index is a portfolio of the T-bill and the futures contracts. As the return
from the T-bill is the more consistent part of the portfolios total return, the diversification benefit from the
commodity futures contracts could be significantly different from that of an unleveraged commodity futures index,
which is examined in this dissertation.

53

no arbitrage free profits can survive in an efficient market. History data does supply
some evidence to supply the risk premium theory. The roll yield is return from rolling
futures positions forward through time. The process of investing in collateralized
commodity futures requires rolling over into the next futures contract on a regular basis.
The act of continually rebalancing between futures contracts leads to a natural
mechanism to buy low and sell high. When the futures markets are in backwardation, a
positive return will be earned.
According to different return components, a commodity future index can have
three different return indexes:
1. Spot return indexSpot return measures the change in value of the nearby contract.
For example, an increase in the value of a contract from $15 to $18 gives a spot
return of (18-15)/15 = 20%.
2. Excess return indexExcess return comprises both spot return and roll yield.
3.

Total returnTotal return reflects all components of return to a fully collateralized


investment in commodities. It consists of excess return and the return on collateral.
Table 3-2 presents the annual total return for the GSCI over the period of

1971-2004 including the three sub-components of total returns. This table indicates that
each component of returns can add significant value to the total returns associated with
the commodity futures index. For instance, continuous compounded spot returns range
from -39.36% to 36.21%. Similarly, the roll yield ranges from -19.88% to 17.54% and
the collateral yield range from 1.05% to 14.94%. The contributions of these three
components made to total return are substantially varied in different time periods.

3.2 Stand-alone Investment Performance

Monthly price time series for GSCI composite, subgroups indexes are obtained

54

from Datastream over the period from January 1970 to December 2004. 4 All of spot
return, excess return, and total return level data are included in the sample. In Table 3-3,
the descriptive statistics of total returns on GSCI and the six subgroup indexes during the
sample period are presented. GSCI and its sub-indexes except the agricultural index
provide a significant positive return. The largest annualized return is 17.53% provided by
GSCI energy sub-index. GSCI composite supplies the second high level return 13.25%.
The annualized return for non-energy is 9.43%, just over half of that of energy sub-index.
The average returns for agricultural, livestock, industrial metals, and precious metals are
4.90%, 10.40%, 7.63%, and 5.75%, respectively.
One of the interesting findings is that the livestock sub-index has the best average
return in the non-energy commodity sector. The average return for the agricultural
sub-index is below the average return of 30-day T-bills over the sample period, which is
5.01%. We also analyze the excess returns for all the indexes. The results show that the
agricultural sub-index has a negative average excess return for the sample period, but all
other indexes supply positive average excess return and the mean levels are significantly
different from zero. An investor can receive an average positive excess return by just
passive buy-and-holding strategy. These results supply some evidence to support the risk
premium theory. One notorious feature of the commodity futures market is its high price
variation, which causes high investment risk. The high variance value and large range of
monthly returns on these indexes identify the risks that an investor should be cautious
with when investing on commodity futures.
Table 3-3 has reported the Sharpe-ratio 5 value in column 6. The risks and returns
tradeoff is also plotted in Figure 3-1. GSCI composite obtains the best investment
performance by having the highest Sharpe ratio value 0.1278. Since GSCI composite is a
portfolio of five different sectors of commodities, according the modern portfolio theory,
a diversified portfolio can achieve an expected return by reducing the related risk. Figure
4

The six subgroups are: energy, non-energy, metals, agricultural, livestock, and precious metals. The time period for
these subgroup is different.
5
Sharpe ratio is obtained by dividing the excess return by standard deviation.

55

3-1 shows that while the energy sub-index provides the best average returns of all the
commodity indexes, it also has the highest investment risk measured by standard
deviation. Energy commodity might not be a good investment vehicle for an investor
without the ability to undertake high risk. The agricultural index has the poorest
investment performance during the sample period as its corresponding Sharpe ratio value
is negative. Livestock has the best performance among all GSCI sub-indexes, but the
problem associated with livestock futures investment is the relative small market trade
volume, which could lead to serious liquidity risk for an institutional investor.
Pindyck and Rotemberg (1990) and Deb et al. (1996) found that there is a
significant contemporaneous correlation in the price movements of unrelated
commodities. This correlation feature does not identify in our data sample. Table 3-4
reports the correlation coefficients of GSCI six sub-indexes. Over the entire sample
period, the five commodity sectors do not show significant correlation. Only precious
metals and industrial metals are slightly correlated. These results support that GSCI
composite is a diversified portfolio and should have better performance than any of the
individual components.
A test for normal distribution is conducted to examine the distribution of the
monthly returns. Test results reject the null hypothesis of normality for all the GSCI
indexes. In fact, we can observe the non-normality characteristics from the value of
skewness and kurtosis presented in Table 3-3 in the last two columns. The distributions
of monthly returns on these GSCI indexes show significant positive skewness and have a
positive kurtosis coefficient. Their return distributions show fat tail behavior and there
are much more returns observations are distributed on the tail.

3.3 Investment Performance Compared with Stocks and Bonds

Let us compare the commodity futures performance with two traditional asset
classes: stocks and bonds. The financial return series employed in my study are: stock

56

return on the value-weighted CRSP 6index for the NYSE/AMEX/NASDAQ combined


market, including dividends, and 10 years T-bond index return from CRSP. The total
return performance of all these three asset classes has been reported in Table 3-5. Figure
3-2 presents the risk/return characteristics.
Over the period from 1970 to 2004, the return on GSCI has been comparable to
the return on the stocks and out-performed corporate bonds. The average return of GSCI
composite is 13.25%, slightly higher than that of stocks and much higher than the return
of bonds. The average return on energy products, which is 17.53%, is about 50% percent
higher than stocks average return and almost double the return of bonds. At the same
time, commodity futures also experiences larger investment risk. The standard deviation
of commodity futures returns is higher than those of stocks and bonds. Bonds have the
best risk/return tradeoff over the sample period. The Sharpe ratio of commodity futures is
slightly lower than bonds but higher than stocks. Overall, commodity futures are a
competitive investment asset class which can provide good market performance.
It has been generally argued that financial returns are not normally distributed
and display substantial kurtosis with more returns observations located at the tail end.
This is also true of commodity futures. Yang and Brorsen (1992) demonstrated that
commodity futures returns are considerably positively skewed compared to stock returns.
We can observe this by comparing the value of skewness and kurtosis for commodity
futures, stocks and bonds. The return distribution of stocks has negative skewness, while
the distribution of commodity returns has positive skewness. This means that,
proportionally, commodity futures have more weight in the right tail while stocks have
more weight in the left tail of return distribution. Commodity futures could be a good
tool to protect downside risk for equity investment. All the return distributions of stocks,
bonds, and commodity futures have positive kurtosis; that is, they are peaked relative
to the normal distribution. The non-normality of financial assets and commodity futures
requires us to use a different method to study the portfolio selection problem rather than
6

Center for Research in Security Prices: A research center at the University of Chicago Graduate School of Business

57

the traditional mean-variance model.


According the mean-variance model, a portfolio effect can be achieved by adding
an uncorrelated or negatively correlated asset into the portfolio. Figure 3-3 plots the
correlation between year return of commodity futures with these of stocks and bonds. A
cursory look at Figure 3-3 reveals a visual pattern between commodity futures, stocks
and bonds. In general, peaks and troughs in the changes of commodity futures appear
going different way with peaks and bottoms in the changes of stocks and bonds. We can
expect a negative correlation of commodity futures with stocks and bonds.
The correlation values for annualized returns have been reported in Table 3-6.
The correlation structures of different investment horizons are examined using
overlapping returns over quarterly, annual, and 3-year intervals. Inspecting the
correlation over different holding periods may reveal patterns in the data that are dimmed
by short-term price variations since financial market and commodity futures market are
both highly volatile in the short term. Figure 3-4 demonstrates the correlation structures
of commodity futures with stocks and bonds for different investment horizons.
Commodity futures show either negative correlation or insignificant correlation
with stocks and bonds. This suggests that commodity futures can be effective in
diversifying traditional financial asset investment. Moreover, Figure 3-4 demonstrates
that the negative correlation of commodity futures with stocks and bonds tends to
increase with the investment horizon. This suggests that the longer the holding period is,
the larger diversification benefit can be obtained by including commodity. Another
obvious feature revealing in Figure 3-4 is that commodity futures display positive
correlation with inflation, which is measured by the change of CPI level, and the
correlation is larger at longer horizons.
The low or negative correlations of commodity futures with stocks and bonds
suggest the potential diversification benefits of commodity futures. Figure 3-5 presents the
change of efficient frontiers with commodity futures. The benchmark efficient frontier is

58

graphed in blue color line marked as port1, which provides a graphical description of
the best portfolio performance that may be achieved using stocks and bonds. Then
different commodity futures indexes are included into the portfolio. When GSCI is added,
the efficient frontier is raised up. The benchmark frontier lies below the efficient frontier
with commodity futures. The addition of commodity futures improves the risk and return
tradeoff for stocks and bonds portfolio. When commodity futures are added into the
portfolio, investors can expect to achieve better return at a given risk level, or reduce the
investment risk to obtain a specified return level.
One of the great concerns for any investor is the downside risk caused by a
hostile or turbulent market movement. Commodity futures make a good choice to protect
downside risk. Panel A in Figure 3-6 shows the mean returns of three commodity futures
indexes during the 5% of worst months of stocks performance when the stock market fell
on average by 9.92% monthly. GSCI composite gains an average 0.34% monthly return,
GSCI energy sub-index provides 1.83% average monthly return, which is higher than its
average 1.34%, and GSCI non-energy sub-index has a negative returns. Both GSCI
composite and GSCI energy sub-index provide the diversification benefits when they are
needed most. GSCI energy sub-index experienced an earning of above average returns
and supplies the best hedging against downside risk of equities investment. Panel B in
Figure 3-6 displays the mean return of commodity futures, stocks, and bonds during the
5% of worst months of bonds performance when bonds fell on average by 4.2%. GSCI
indexes experienced a significant positive average monthly return, 1.81% for GSCI,
1.49% for GSCI energy, and 1.10% for GSCI non-energy. They all have a higher return
than their average. It seems that commodity futures work very well to hedge against
bond market downside risk.
Let us examine four different portfolios as following: portfolio 1 with 60% stocks
and 40% bonds; portfolio 2 with 55% stocks, 35% bonds, and 10% GSCI composite;
portfolio 3 with 55% stocks, 35% bonds, and 10% GSCI energy; portfolio 4 with 55%
stocks, 35% bonds, and 10% non-energy. Table 3-7 reports the relative performance of

59

these portfolios. When commodity futures are added, the average loss is decreased and a
downside protection is observed. Energy index provides the best downside risk
protection among the entire commodity futures index. The average month loss is
decreased by almost 2% after GSCI energy sub-index is added. Adding commodity
futures into stock and bond portfolio improves the expected return and decreases the
portfolio risk. The Sharpe ratio measure is improved.

3.4 Commodity Futures Investment and Inflation

Inflation has an adverse impact on financial markets. The government bonds are
nominally denominated assets and can completely hedge against expected inflation but
cannot hedge against unexpected inflation shock. When inflation is unexpectedly higher,
the return of bonds will significantly lose purchasing power. The return on stocks is
expected to support a better hedge against inflation than bonds since stocks represent
claims against real assets whose value can be expected to keep pace with inflation
variation. But accumulated empirical results suggest that stock returns are negatively
related to both expected and unexpected inflation.
Commodity futures have been suggested to be an effective hedge against both
expected and unexpected inflation. First, physical commodity prices such as oil are an
underlying source of inflation. As the cost of raw materials increases, so does the product
price. In fact, commodity prices are a component of the consumer price index and higher
commodity price mean higher inflation. Second, a rational prediction of commodity
futures spot prices has included information about the foreseeable trends; an unexpected
inflation will shift up the futures spot price and result in an extra gain for long futures
contracts holders. Third, one of the main components of commodity futures return is
collateral yield from T-bills, which is perfectly correlated with expected inflation. Figure
3-7 plots the standardized year change of GSCI total return index and CPI. In general,
the change of commodity futures runs parallel with inflation movement; the peaks and

60

bottoms of commodity futures returns display shortly before, shortly after or


simultaneously with peaks and bottoms of the inflation.
Figure 3-8 presents the correlation of stocks, bonds and commodity futures with
inflation over various investment horizons. Commodity futures have an opposite
exposure to inflation compared with stocks and bonds. At all horizons, commodity
futures are positively correlated with inflation, while the correlation of stocks and bonds
with inflation is negative. As the investment horizon becomes longer, the negative
inflation correlation of stocks and bonds and the positive inflation correlation of
commodity futures are becoming larger in absolute magnitude. In other words, the longer
the investment horizon, the more inflation hedging benefit can be achieved by adding
commodity futures into stocks and bonds portfolio. Another clear picture in Figure 3-8 is
that energy commodity index exhibits the highest degree of positive correlation with
inflation at all investment horizon. This result suggests that direct investing in Energy
futures could provide a significant inflation hedge.
Different components of inflation may have distinctive impact on assets returns.
In order to examine this difference, we need to measure the expected and unexpected
parts of inflation. There are several methods to decompose the inflation. One of the
widely used approaches is the Fama and Schwert (1977) model which uses the
short-term T-bill as a proxy for markets expectation of inflation. The unexpected
inflation is measured by the difference between the actual inflation rate and the normal
short-term interest rate. Figure 3-9 demonstrates the correlation of stocks, bonds, and
commodity futures returns with the different components of inflation derived from
Famas model. Bonds return is negatively correlated with unexpected inflation but
positive relate to expected inflation. The stock market displays negative correlation with
both expected and unexpected inflation. The negative sensitivities of stocks and bonds to
inflation stem mainly from sensitivities to unexpected inflation. The correlation with
unexpected inflation exceeds that with the total inflation.

61

Figure 3-9 gives us an anomalous result about the correlation of commodity


futures with the expected inflation component. The GSCI composite and non-energy
sub-index exhibit negative correlation with expected inflation and energy sub-index
shows positive correlation. But all of these correlation coefficients are not significant. To
explain why commodity futures return is not sensitive to expected inflation, we should
examine the correlation of expected inflation with three sources of commodity futures
return: collateral yield, roll yield, and spot price change. The collateral yield earned on
the Treasury bills is used as collateral for the futures investment. Treasury bills are
obvious to perfectly correlate with expected inflation. Roll yield, coming about when
futures prices are below spot prices (backwardation), is decided by the volatility of
futures market. Spot return, the return earned by the underlying futures contract price
changing, has been proved to be negatively correlated with the short-time interest rate
(Bjornson and Carter 1997). Since the short-term interest rate has been used as the proxy
for expected inflation, spot return displays significant negative correlation with expected
inflation and cause commodity futures total return to relate slightly to expected inflation.
Table 3-8 reports the correlation of GSCI composite spot return and roll yield with
current expected inflation and different lag value of unexpected inflation. Table 3-8
reveals that roll yield does not significantly relate to either expected or unexpected
inflation and spot return does present negative correlation with expected inflation.
As spot return derives from the futures price change, an unexpected inflation
shock usually raises up commodities price, causing an unexpected gain to long futures
contracts. We expect the spot return should display a positive response to unexpected
inflation. Table 3-8 reveals that the commodity futures price variation is followed in
short order by responses in the unexpected inflation rate. While this lag may be an
artifact related to reporting delays, it is still clear that unexpected inflation, which
adversely moves the financial market, is positively related to commodity spot returns.

3.5 Commodity Futures Investment and Business Cycle

62

Another factor driving the negative correlation of commodity futures with stocks
and bonds is the counter-cyclical movement between these asset classes. In general, at
the bottom of an economic cycle, the economy is well below its potential capacity, stocks
and the bond market have been at their highest performance, but commodity futures
returns have been at their lowest. When economic growth accelerates, stock and bond
returns begin to decline but commodity futures return increases. When the economy
heats up and runs above capacity, commodity futures outperform stocks and bonds. The
NBER 7 business cycle data are used to demonstrate this counter-cyclical movement.
The NBER maintains a chronology of the U.S. business cycle. The chronology
identifies the dates of peaks and troughs framing economic recession or expansion.
According to the relative relationship between economic output level and potential
output capability, a business cycle can be divided into 4 business phases: two or
recession and two for expansion. An expansion period (from trough to peak) can be
equally divided into early expansion (phase 1) and later expansion (phase 2). A recession
period (from peak to trough) can be broken into early recession (phase 3) and later
recession (phase 4). The relevant NBER business cycle chronology for our sample period
is shown in Table 3-9.
Figure 3-10 reports the average returns of commodity futures, stocks and bonds
over expansion and recession period. The business cycle impacts asset classes differently
because of the dynamics of the economy during each phase. During the expansion period,
GSCI energy index outperforms other asset classes by showing the highest average
monthly return; average monthly returns on stocks and GSCI composite are remarkably
similar, 1.06% and 1.09%, respectively; GSCI non-energy and bonds have the lower
return during that period, 0.8% and 0.7%, respectively. During the recession period, the
return on stocks and GSCI composite are decreased by more than 50% and the average
monthly return on GSCI energy becomes negative. Instead, bond presents very good
performance and achieves the 1.4% average monthly returns. The return on GSCI
7

The National Bureau of Economic Research is a private, nonprofit, economic research consortium.

63

non-energy also shows slightly decreasing but does not significantly changes. It seems
that stocks and commodity futures show similar pattern: well perform at the expansion
and poor perform under recession.
The similarity between stocks and commodity futures obscure some important
differences. When we examine the performance of commodity futures and financial
assets under different phases of business cycle, the striking differences between stocks
and commodity futures are evident. The annualized monthly average returns in four
phases have been reported at Table 3-10 and Figure 3-11. It is easy to observe the
asynchronous movement of these asset classes. During the expansion period, commodity
futures have better average returns and outperform stock and bond at phase 2 and stock
and bond obtain better returns and outperform commodity futures at phase 1. GSCI
composite and non-energy index achieve their best historical performance at the later
expansion phase. In contrast, during the recession period, stock and bond have much
stronger performance at the later recession than at the early recession; both of them
achieve their best average return at this phase. However, commodity futures have much
poorer performance at the later recession than at the early recession. Energy index
obtains its best performance at the early recession phase.
One of the striking features in Figure 3-11 is that energy futures have a very
strong performance over the early recession phase followed by a huge loss over the later
recession phase. In fact, as many economists have argued, (for example, Hamilton [2003,
1985]), an economic recession often connects to an oil shock. The unexpected increases
in oil prices disrupt economic activity and cause the raise of raw material costs and
declines in output as well as employment. Major economic recessions are associated with
the oil crises of the 1970s and 1980s. The general picture during the early recession stage
is that oil and energy-related prices rise unexpectedly, resulting in a windfall gain to
energy futures holders. As the economy moves far below the potential GDP growth, the
demand and supply of energy products change and a significant excess capacity implies
falling in price which causes a negative return to long futures investors.

64

It is necessary to give a brief overview of the impact of the business cycle on


different types of assets. Phase 1 is often characterized by a resurgence of growth led by
inventory accumulation and business investment. Given that rising demand can be met
with existing capacity, commodity prices keep stable; corporate margins expand; and the
earnings grow at a rapid rate. Stock investments will return a good gain and the stock
market outperforms commodity futures. Meanwhile, monetary policy often remains
accommodative as significant excess capacity is present and inflation keeps falling or
remains stable. The bond market will perform relatively well. As the economy moves
above its sustainable non-inflationary growth path, the problems of capacity constraints
and inflation become evident and the Federal Reserve Bank (the Fed) starts to increase
interest rates as a way of restraining inflation. The economic system then enters phase 2.
In this phase, bond shows poor performance as the coupon income often just offsets
capital losses caused by high inflation rate. The stock markets begin to flag as the rising
interest rates increases the discount factor for future dividends. However, commodity
markets show strong performance as supply is constrained by capacity limits and
demand keeps growing with economic acceleration.
When economic growth reaches a cyclical peak, it turns around and moves back
to the sustainable output level. The economic system starts an early recession phase 3.
This is a dangerous phase for stock and bond investment as the economic policy keeps
tight to moderate excess growth and demand. The combination of tight policy and
inflation conspires to distress stock and bond markets. Meanwhile, capacity remains tight
as demand still keeps strong at this phase, which helps commodity to achieve the
strongest performance of any major asset classes. Eventually, economic growth falls
below the potential level and inflation pressure abates. As a result, economic policy
becomes accommodative again and a later recession starts. Stocks and bonds register
their best performance at this stage in response to the good anticipation of an upturn in
the economy and relax monetary policy environments. However, slack demand and
excess capacity conspire to move commodity returns to the negative zone. Commodity

65

becomes the worst performing asset class in phase 4.

3.6 Commodity Futures Investment and Event Risk

One of the critical issues for portfolio diversification is event risk management.
Financial markets are very sensitive to the unexpected events; a bust in financial market
is often associated with an event shock and leads to a substantial loss to investors. We are
interested in whether adding commodity futures could make a portfolio less likely to be
exposed to significant event risk.
As most research has demonstrated, commodity futures returns tend to have
positive exposure to event risk because the event shocks that occur in the commodities
markets tend to be those that unexpectedly cut the supply of commodity to the market
and relegate the commodity prices into a high level, causing a significant gain to long
futures investors. For example, events such as natural disaster, droughts, floods, and crop
freeze all result in a shortage of supply of agricultural products; similarly, events such as
OPECs decision to cut crude oil supply cause a driving up in energy price. Furthermore,
these commodity event shocks are expected to be uncorrelated. The global supply and
demand factors for each individual commodity that determine the market price of each
commodity are very different and rarely relate to each other. The primary factors that
determine the price of crude oil are very different from those that decide the price of
soybeans, wheat, gold, or metals, and OPEC agreement to decrease oil supply should not
correlate with floods in the agricultural regions around world. So it is reasonable to
expected that the price movement of individual commodity to be uncorrelated with each
other. This has important implications for a commodity index.
Moreover, the event shocks to commodities markets are expected to be more
likely to negatively relate to the financial markets. The events causing the supply
disruption and price rise in commodity markets usually have a negative impact on
financial assets, whose expected returns will be reduced by the higher cost of raw

66

material inputs.
Table 3-11 exhibits the different patterns of GSCI and stocks exposed to main
events during the sample period from 1970 to 2004. In the early and middle of 1970s,
there were several oil crises and this was boon for commodity futures investment. The
average year return is more than 50% for GSCI. But for the stock market, this was a
disaster. In contrast, in 1998, there was a substantial increase of oil supply and petroleum
products, plenty of cheap raw materials hurt commodity market but made for a strong
performance in stock market. In 2000 and 2001, stock markets suffered a substantial loss
caused by the collapse of the internet boom and the 9/11 attack, but commodity futures
display strong performance.
In fact, the pattern that supply disruptions provide positive returns to commodity
futures investment has been verified when we consider the distribution of returns to
commodity futures. The return series of GSCI indexes display different degrees positive
skewness; that is, more return observations to the right of median than to the left of
median. As we have demonstrated, positive skewness return is a beneficial to protect the
downside risk and provides upward return bias to the portfolio. In conclusion,
commodity futures tend to be an efficient hedger against event risks which are associated
with financial assets.

3.7 Conclusion

The data from 1970 to present indeed provide evidence to support the
diversification benefit of commodity futures investment:
Commodity futures index supplies significant positive return to investors.
The return of different commodities is not significantly related with each other.
On average, the total return from an unleveraged commodity futures index is

comparable in magnitude and volatility to stock returns, and outperforms bond

67

returns in magnitude. The relative performance between commodity futures index


and financial assets is varied with business cycle.
Commodity futures are negatively correlated with stocks and bonds in general, but

not always.
Commodity futures index, especially energy index, is more efficient to hedge

against inflation and event risk.

68

Table 3-1: The component breakdown of GSCI main index


Constituents
Energy Weight

GSCI $Weights
72.68%

Sub-Index Weights
100.00%

Crude Oil
Brent Crude Oil
Unleaded Gas
Heating Oil
Gas Oil
Natural Gas

28.90%
13.63%
8.32%
8.14%
4.51%
9.18%

39.76%
18.75%
11.45%
11.20%
6.21%
12.63%

Agricultural Weight

12.39%

100.00%

2.84%
1.19%
3.22%
1.84%
1.19%
1.24%
0.57%
0.30%

22.92%
9.60%
25.99%
14.85%
9.60%
10.01%
4.60%
2.42%

6.16%

100.00%

3.47%
0.81%
1.88%

56.33%
13.15%
30.52%

6.67%

100.00%

2.90%
2.21%
0.30%
0.81%
0.45%

43.48%
33.13%
4.50%
12.14%
6.75%

2.09%

100.00%

1.88%
0.21%

89.95%
10.05%

Wheat
Red Wheat
Corn
Soybeans
Cotton
Sugar
Coffee
Cocoa
Livestock Weight
Live Cattle
Feeder Cattle
Live Hogs
Industrial Metal Weight
Aluminum
Copper
Lead
Nickel
Zinc
Precious Metals Weight
Gold
Silver

Note: the weight is decided on November 24th, 2004; for most recent weights, please check with
Goldman Sacks Group.

69

Table 3-2: Spot returns, collateral yield, and roll yield of GSCI
year

spot annual returns

roll yield

excess returns

collateral yield

total returns

1971

3.39%

8.24%

11.63%

4.51%

16.14%

1972

24.07%

3.80%

27.88%

4.08%

31.95%

1973

36.21%

9.09%

45.31%

7.04%

52.34%

1974

34.72%

6.46%

41.19%

8.06%

49.25%

1975

-39.36%

11.54%

-27.82%

6.08%

-21.74%

1976

-22.25%

-2.97%

-25.23%

5.21%

-20.02%

1977

0.51%

3.75%

4.26%

5.22%

9.48%

1978

26.16%

0.94%

27.10%

7.12%

34.22%

1979

17.71%

-2.04%

15.67%

10.14%

25.80%

1980

18.92%

-13.55%

5.36%

11.53%

16.90%

1981

-28.21%

-12.08%

-40.28%

14.94%

-25.34%

1982

-5.22%

0.00%

-5.22%

11.27%

6.05%

1983

2.07%

-1.15%

0.92%

8.76%

9.68%

1984

-4.54%

1.63%

-2.90%

9.97%

7.07%

1985

1.05%

1.15%

2.20%

7.72%

9.92%

1986

-21.58%

17.31%

-4.27%

6.22%

1.94%

1987

5.26%

10.92%

16.19%

5.92%

22.10%

1988

2.27%

6.75%

9.03%

6.64%

15.67%

1989

9.83%

12.51%

22.34%

8.35%

30.69%

1990

14.50%

11.51%

26.00%

7.79%

33.79%

1991

-17.35%

11.28%

-6.07%

5.74%

-0.34%

1992

-5.38%

-1.45%

-6.82%

3.59%

-3.23%

1993

-7.32%

-5.41%

-12.73%

3.09%

-9.64%

1994

5.02%

-10.14%

-5.13%

4.14%

-0.99%

1995

8.54%

-0.60%

7.94%

5.67%

13.61%

1996

14.30%

17.54%

31.85%

5.17%

37.02%

1997

-14.47%

2.54%

-11.93%

5.14%

-6.79%

1998

-35.24%

-19.88%

-55.12%

4.98%

-50.15%

1999

36.08%

-11.38%

24.71%

4.67%

29.38%

2000

26.90%

10.53%

37.44%

5.92%

43.36%

2001

-38.20%

-1.40%

-39.60%

3.88%

-35.72%

2002

25.49%

-8.64%

16.85%

1.68%

18.52%

2003

11.51%

7.67%

19.18%

1.05%

20.23%

2004

27.95%

-1.94%

26.01%

1.29%

27.30%

average

3.33%

1.84%

5.17%

6.25%

11.43%

Source: GSCI index data are downloaded from DataStream, the author estimates the collateral yield
and roll yield. All returns are continuous compounding returns.

70

Table 3-3: Summary statistics of total returns on GSCI (from 1970 to 2004)
Asset class

Annualized

StDev

Minimum

Max

Sharpe

Skewness

Kurtosis

Monthly

monthly

ratio

return

return

5.31%

-18.08%

20.42%

0.1278

0.1709

1.5950

5.48%

-17.54%

25.56%

-0.0030

0.5120

5.22%

-19.03%

18.78%

0.0847

-0.1025

Mean (P-value)
GSCI

13.25%

composite

(0.0002)

GSCI

4.90%

agricultural

(0.1294)

GSCI

10.40%

livestock

(0.0007)

GSCI

7.63%
(0.0047)

6.72%

-30.13%

32.52%

0.0314

0.4286

3.9398

5.75%
(0.0063)

6.45%

-32.56%

39.40%

0.0084

0.7902

7.0571

17.53%

9.03%

-27.39%

34.13%

0.1145

0.2040

1.3000

4.29%

-18.09%

20.42%

0.0827

0.3398

4.0250

industrial

2.7450

1.2520

metals
GSCI
Precious
metals
GSCI energy

(0.0056)
GSCI

9.43%

non-energy

(0.0009)

Table 3-4: Correlation structure of GSCI sub-indexes


Agricultural Energy Industry Livestock Non-energy Precious
metal
metal

Agricultural

1.00000

Energy

-0.02147

1.00000

(0.7289)

Industry
metal
Livestock

Non-energy
Precious
metal

0.13884

0.04650

(0.0243)

(0.4527)

0.08745

0.02000

-0.00922

(0.1573)

(0.7469)

(0.8818)

0.71999

0.03719

0.28325

0.57944

(<.0001)

(0.5482)

(<.0001)

(<.0001)

0.05904

0.13445

0.21381

0.03192

0.14916

(0.3403)

(0.0293)

(0.0005)

(0.6063)

(0.0155)

Note: the p-value is reported at parenthesis

1.00000
1.00000

1.00000

1.00000

71

Table 3-5: Historical returns on GSCI, Stocks, and Bonds (from 1970 to 2004)
Asset class

Annualized

Annualized

Minimum

Max

Sharpe

Mean (P-value)

StDev

Monthly

monthly

ratio

return

return

4.63%

-22.53%

16.56%

2.39%

-6.68%

5.31%

stocks

11.72%

Skewness

Kurtosis

0.1191

-0.4907

1.9351

9.99%

0.1293

0.2794

1.1238

-18.08%

20.42%

0.1278

0.1709

1.5950

9.03%

-27.39%

34.13%

0.1145

0.2040

1.3000

4.29%

-18.09%

20.42%

0.0827

0.3398

4.0250

(<.0001)
T-bonds

8.81%
(<.0001)

GSCI

13.25%

composite

(0.0002)

GSCI energy

17.53%

(0.0056)
GSCI

9.43%

non-energy

(0.0009)

Table 3-6: Correlation of GSCI indexes with stocks and T-bond


Stocks
T-bond

GSCI

AG

EN

IN

LV

NE

PM

-0.01203

0.08947

-0.05835

0.12090

0.05418

0.14461

-0.04504

0.8461

0.1479

0.3459

0.0502

0.3815

0.0190

0.4670

-0.05299

-0.02781

-0.04820

-0.14772

0.00209

-0.05920

-0.09904

0.3920

0.6535

0.4363

0.0165

0.9731

0.3389

0.1091

Table 3-7: Portfolio performance of downside risk protection


Portfolio

Annualized Annualized
Average downside
Sharpe ratio
mean return std deviation
downside protection

60/10 Stocks/Bonds

10.55%

37.45%

0.1455

-6.05%

N/A

55/35/10 Stocks/Bonds/GSCI

10.86%

34.51%

0.1669

-5.06%

0.99%

55/35/10 Stocks/Bonds/NE

10.47%

34.68%

0.1548

-5.13%

0.92%

55/35/10 Stocks/Bonds/EN

12.04%

32.72%

0.2121

-4.14%

1.91%

72

Table 3-8: GSCI correlation with different inflation components


Correlation with unexpected inflation

spot

roll

three months following

0.11

-0.04

two months following

0.14

0.05

one month following

0.31

0.04

current month

0.01

0.11

one month previous

0.00

0.02

two months previous

0.08

-0.03

three months previous

-0.01

0.01

-0.07

Correlation with current expected inflation


*Significant at 95% level

-0.17

Table 3-9: The NBER business cycle chronology


BUSINESS CYCLE REFERENCE
DATES
peak

trough

DURATION IN MONTHS
Contraction

Expansion

Cycle

peak to

Previous

Trough from

Peak from

trough

trough to

Previous

Previous Peak

this peak

Trough

106

117

December 1969

November 1970

11

116

November 1973

March 1975

16

36

52

47

January 1980

July 1980

58

64

74

July 1981

November 1982

16

12

28

18

July 1990

March 1991

92

100

108

March 2001

November 2001

120

128

128

Source; the NBER

Table 3-10: Average asset returns in four phases (from 1970 to 2004)
Phase 1

phase 2

phase 3

phase 4

stocks

14.18%

11.55%

-22.81%

31.52%

bonds

9.73%

4.77%

9.28%

22.30%

GSCI

7.34%

19.08%

14.50%

-3.00%

energy

10.21%

19.38%

58.27%

-75.78%

non-energy

4.67%

13.31%

7.04%

7.00%

73

Table 3-11: Year returns with events


YEAR
1973
1974
1981
1990
1998
2000
2001
2004

STOCKS
-18.77%
-27.93%
-3.98%
-6.08%
22.30%
-11.04%
-20.85%
13.04%

GSCI
52.34%
49.25%
-25.34%
33.79%
-50.15%
43.36%
18.52%
42.35%

Figure 3-1: Risks and returns tradeoff for GSCI indexes


120.00%
100.00%

GSCI
AG
LV
IM
PM
EN
NE

risk

80.00%
60.00%
40.00%
20.00%
0.00%
0.00%

5.00%

10.00%
return

15.00%

20.00%

74

Figure 3-2: The risks and returns characteristics of GSCI, Stocks, and Bonds

risk
120%
100%
80%
60%
40%
20%
0%
gsci

energy

nonenergy

stocks

T-bonds

stocks

T-bonds

return

14%
12%
10%
8%
6%
4%
2%
0%
gsci

energy

nonenergy

Sharpe ratio
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0

0.1278

0.1191

0.1146

0.1293

0.084

GSCI

EN

NE

STOCKS

T-bonds

75

Figure 3-3: Time series plot graph of return on GSCI, stocks, and bonds
Panel A: Non-energy index, Stocks, Bonds
0.6
0.5
0.4
0.3
0.2
stocks
bonds

0.1

non-energy

20
04

20
02

20
00

19
98

19
96

19
94

19
92

19
90

19
88

19
86

19
84

19
82

19
80

19
78

19
76

19
74

-0.1

19
72

19
70

-0.2
-0.3
-0.4

Panel B: GSCI composite index, Stocks, Bonds


0.6

0.4

0.2

19
70

stocks
bonds
GSCI

-0.2

-0.4

-0.6

Panel C: Energy index, Stocks, Bonds


0.8

0.6

0.4

0.2

19
70

-0.2

-0.4

-0.6

-0.8

stocks
bonds
energy

76

Figure 3-4: Correlation under different investment horizon


Panel A: GSCI composite index, Stocks, Bonds

GSCI
0.4
0.2

stocks

bonds
inflation

-0.2
-0.4
month quarter annual 3-year

Panel B: Energy index, Stocks, Bonds


energy
1
0.5

stocks
bonds

inflation

-0.5
month quarter annual 3-year

Panel C: Non-energy index, Stocks, Bonds

non-energy
0.6
0.4
0.2

stocks

0
-0.2

bonds
inflation

-0.4
month quarter annual

3-year

77

Figure 3-5: Efficient frontier with GSCI, Stocks, and Bonds

Figure 3-6: Downside risk protection


Panel A: average return of commodity futures during the 5% worst return period of stocks
mean return

5.00%
0.00%
GSCI

ENERGY

-5.00%

NONENERGY

STOCKS

T-BOND

-10.00%
-15.00%
Panel B: average return of commodity futures during the 5% worst return period of stocks
mean return
3.00%
2.00%
1.00%
0.00%
-1.00%
-2.00%
-3.00%
-4.00%
-5.00%

GSCI

ENERGY

NONENERGY

STOCKS

T-BOND

78

Figure 3-7: Standardized year change of GSCI total return index and CPI
3

cpi
0
20
04

20
02

20
00

19
98

19
96

19
94

19
92

19
90

19
88

19
86

19
84

19
82

19
80

19
78

19
76

19
74

19
72

19
70

GSCI

-1

-2

-3

Figure 3-8: Correlation with inflation


0.6
0.5
0.4
0.3
STOCKS
BONDS
NON-ENERGY
GSCI
ENERGY

0.2
0.1
0
-0.1
-0.2
-0.3
-0.4

monthly

quarterly

annual

3-year

Figure 3-9: Correlation with different inflation components


0.5
0.4
0.3

STOCKS
BONDS
NON-ENERGY
GSCI
ENERGY

0.2
0.1
0
-0.1
-0.2
-0.3

inflation

expeted
inflation

unexpected
inflation

79

Figure 3-10: Average returns of commodity futures, stocks and bonds with the
business cycle

0.2
0.15
0.1

non-energy
energy
GSCI
bonds
stocks

0.05
0
-0.05
-0.1

overall

expansion recession

Figure 3-11: Average returns of commodity futures, stocks and bonds under
different phases

1
0.5
non-energy
energy
GSCI
bonds
stocks

0
-0.5
-1

Phase 1

phase 2

phase 3

phase 4

80

Chapter 4 STATIC ASSET ALLOCATION WITH REGIME-SWITCHING

Hamiltons (1989) regime switching model is introduced to examine time-varying


aspects of asset return. Under the regime switching model, the risk/return characteristics
of asset classes will be different under distinct regimes and the regime variable is
autocorrelated rather than independent. When portfolio managers make a decision about
asset allocations based on time-varying expected returns and variations, they should
consider the current regime and take the regime switching effect into account. This
chapter starts with an overview on the regime switching model, and then an examination
of assets return time series, followed by a study of static asset allocation.

4.1 Introduction of Regime Switching Model


Overview

Almost any macroeconomic or financial time series for a sufficiently long period
will undergo episodes in which the behavior of the series seems to change quite
dramatically (Hamilton 1994). For example, macroeconomic systems exhibit periodical
switching between boom and recession; and futures markets periodically switch from
bear to bull. The Markov regime-switching model (MRS) is a time series econometric
technique that can handle the dramatic structural breaks with the advantage of allowing
breaks to happen several times. In a regime switching model, the parameters are not
constant through the sample period, but rather change with structural shifts, which divide
the period into distinct regimes with different parameter values. Moreover, the change in
regime is considered as a random variable instead of a perfectly foreseeable deterministic
event, and switching between regimes occurs stochastically according to a Markov

81

process (Hamilton 1994).


Since Hamilton (1989) introduced this method into the time series study as an
alternative way to model business cycles, a number of researchers have adopted this
method to account for specific features of economic time series such as the asymmetry of
economic activity over the business cycle (Diebold and Rudebusch [1996]) or the fat tail,
volatility clustering, and mean reversion in financial asset price series (Hamilton and
Susmel [1994], Schaller and Norden [1997], and Maheu and McCurdy [2000).
Furthermore, MRS suggests an alternative to model conditional heteroskedasticity, an
important feature of asset returns. Cai (1993) and Hamilton and Susmel (1994) proposed
a regime switching autoregressive conditional heteroskedasticity (switching ARCH)
model as the extension of Engle (1982)s original ARCH model. They showed that such
a specification can significantly improve the forecasts of stock volatility. Gray (1996)
generalized these models to incorporate the GARCH effects and proved that the
regime-switching GARCH is useful to model the volatility of interest rates. Timmermann
(2000) derived the moments for a range of MRS and showed that this model can generate
a wide range of coefficients of skewness, kurtosis and serial correlation. Recently, Ang
and Bekaert (1999) and Graflund and Nilsson (2002) applied a regime switching model
to examine the dynamic portfolio selection problem and concluded that optimal portfolio
construction should depend on the underlying economic regime.
Simple Markov regime switching model

Let us look at a simple two-regime switching model. In finance economics


literatures, it is generally assumed that the price of a financial asset follows a stochastic
process described by a stochastic different equation:
( Pt ) 1 dPt = t ( st )dt + t ( st )dWt

(4.1)

where t and t are drift parameter and diffusion parameter, Wt denotes a standard
Brownian motion. In regime switching setting, there is a latent regime variable st which

82

discretely switches the value of drift and diffusion. The regime at time t is denoted by a
state variable by st = j if and only if the time t regime is j (j=1,2,k, for k regimes
model).st is unobserved to econometrician and changes the value following a Markov
chain. In this simple model, the regime variable st can be 1 or 2 and assumed to follow
a time-homogenous first-order Markov chain, i.e. we assume that the most present
history of the chain matters and that the switching probabilities are constant. The
transition probabilities matrix is denoted by P={pjk}, j,k=1,2.and:
prob( st = 0 / st 1 = 0) = p 00
prob( st = 1 / st 1 = 0) = p 01
prob( st = 0 / st 1 = 1) = p10

(4.2)

prob( st = 1 / st 1 = 1) = p11

and
p 00 + p 01 = 1
p10 + p11 = 1

(4.3)

We can discretize the stochastic different equation (4.1) by assuming dt=ti-ti-1=h, where
h=T/n. Then equation (4.1) is approximated as:

rt ( st ) ( Pt 1 ) 1 ( Pt Pt 1 ) = ( st )h + ( st ) h t

(4.4)

where ti ~ N (0,1) rt is the assets returns at time t, which leads to:


rt ( st ) ~ N ( ( st ), 2 ( st ))

(4.5)

In other words, the asset return follows a mixture normal distribution with its
mean and variance switching depending on the unobserved state st, which follows a
Markov chain. This model can be extended to the n-asset case by replacing rt by a 1n
asset return vector Rt, t by a 1 bn mean return vector, and t by a nn standard
deviation matrix t as follows:

83

Rt ( st ) ~ MVN ( (1n ) ( st ), ( st )( st ) ') (4.6)


where MVN denotes a multivariate normal distribution. Then the n-asset returns follow
a multivariate normal distribution with regime switching. This simple two-regime
switching model has been empirically proved to be flexible enough to capture the
characteristics of asset return distribution such as asymmetric, high-peek and fat tails.
Basic estimation procedures

Basically the parameters estimation procedure follows the maximum likelihood


method, but due to the regime switch and hidden state variables, special recursive filters
proposed by Hamilton (1989, 1990, and 1994) is required to introduce the likelihood
function. Before we discuss the estimation procedure, let us define a few variables.
First, we define the conditional multi-normal density vector
1
1
f ( Rt / st = 1) 2 ( st = 1) exp(( Rt ( st = 1)) ' ( st = 1) ( Rt ( st = 1)))

t =
=

1
f ( Rt / st = 2) 2 ( st = 2) exp(( Rt ( st = 2)) ' ( st = 2) 1 ( Rt ( st = 2)))

(4.7)

where ( st ) = ( s t ) * ( st )'
Secondly, we define a state inference ti / tj :
Pr ob( sti = 1/ I tj )
(4.8)
Pr
ob
(
s
2
/
I
)
=
ti
tj

ti / tj =

Itj contain the history information of returns up to time tj, therefore ti / tj denote the
state probability at time ti given the information up to tj. This is called filter if ti=tj,
forecaster if ti>tj, and smoother if ti<tj. Following Hamilton (1994), when an initial
value 0 / 0 is given, the filter and one-step-ahead forecast can be calculated by the
following iteration scheme:

ti / ti =
ti +1 / ti

ti / ti 1 ti
1'*( ti / ti 1 ti )
= p * ti / ti

(4.9)

84

where denotes the element-by-element multiplication of vectors. The denominator


of equation (19) actually calculates the likelihood for observation Rti. Therefore, with T
observations, the log-likelihood function to be maximized is written as:

L( ) = log(1' ( t / t 1 t )

(4.10)

t =1

where = ( ( st = 1), ( st = 2), ( st = 1), ( st = 2, ), p11 , p22 ) denotes model parameters


which we are interested and need to be estimated.
Two methods have been developed for the maximum likelihood estimation. One
is numerical optimization and another is Expectation-Maximization (EM) algorithm. It is
well known that Markov switching models has the problem of a multimodal likelihood
surface. For this reason, the researcher should have a strategy to find the global
maximum. The EM algorithm shows more robust with ill-shaped likelihood shape and is
fairly easily implemented using the smoother defined by Kim (1994) as:

ti / T = ti / ti {P'[ ti +1 / T () ti +1 / ti }

(4.11)

where () denotes the element-by-element division of vectors. The initial value T / T is


given by the last step of the recursive calculation of equation (9) and then this is solved
backwardly.
Setting up an initial parameters value 0 , we can run the EM algorithm to update
the parameters until a pre-determined convergence criterion is achieved. There are two
main components for the EM algorithm at each iteration step:
First, we need run the iteration (9) to obtain the filter, forecaster and then run the
iteration (10) to obtain the smoother for each time period ti, i.e.
Pr ob( sti = 1/ IT ; )
(4.12).
Pr ob( sti = 2 / IT ; )

ti / T =

85

Second, we update the parameters according the following rule which will
maximized the likelihood function based on the prior distributions assumption:
T

( st = j )(i +1) =

prob(s
t =1
T

prob(s

t =1

( st = j )

( i +1)

= j / IT ; ( i ) )Rt

prob(s

t =1

= j / I T ; ( i ) )

= j / IT ; ( i )( Rt ( st = j )( i +1) )( Rt ( st = j )(i +1) ) '


T

prob(s

t =1

T 1

Pmn (i +1) =

prob(s
t =1

t +1

= j / I T ; )

= n, st = m / IT ; ( i ) )

T 1

prob(s
t =1

(4.13)

(i )

;
= m / I T ; )
(i )

m, n = 1, 2

where the superscript (i) shows the corresponding parameters estimated in the ith
recursive calculation.
Diagnostic test of the Markov switching specification

When using regime-switching model, the applied econometrician should be


always careful about whether there is a regime switching effect and how many regimes
that characterize the data. Without a formal examination and hypotheses tests on those
issues, a data-mining problem can occur and the model will be misspecified.
Unfortunately, the traditional likelihood ratio test cannot be used to test the hypotheses
about the regime effects. As Hamilton (1990) has pointed out, one of the regularity
conditions for the likelihood ratio test to have an asymptotic 2 distribution is that the
information matrix I be nonsingular. This condition cannot be satisfied if the
econometrician wants to fit an N-state model when the true process has N-1 regimes
since under the null hypothesis the parameters that describe the Nth state are
unidentified.
Hansen (1992, 1993) has suggested a method to get around the problems. He
regarded the likelihood function as an empirical process with the unknown parameters

86

and used the empirical process theory to derive a bound for the asymptotic distribution of
a standardized likelihood ratio statistic. Curcia and Shah (1993) showed how to use
Hansens work to test for Markov regimes switching models.
In order to use the LR test to address a problem posed by nuisance parameters,
Hansen applied empirical process theory (Shorack and Wellner 1987) to derive an upper
bound for type error of a modified LR statistic under the null, assuming nuisance
parameters are known under the alternative. Let us briefly review his test.
Let us write the log-likelihood function as follows:
n

L( , ) = li ( , ) (4.14)
i =1

where n is the sample size, denotes parameters identified under both the null and
alternative and = ( , ) where is a vector of nuisance parameters not identified
under the null. We want to test the null hypothesis that =0. According to Hansen, the
parameter vector is concentrated out of the sample likelihood function by
following:
)
( ) = arg max L( , ) given is defined.
This leaves the concentrated likelihood function as:
)
)
L ( ) = L( , ( )) (4.15)
And then we define the following statistics that will be used to construct the
standardized LR statistic for testing the number of regimes:
)
)
LR( ) = L( , ( )) L(0, , (0, ))
n
)
V ( ) = qi ( , ( )) 2

(4.16)

i =1

)
)
)
1
qi ( , ( )) = li ( , ( )) li (0, , (0, )) LR ( )
n
.

The standardized likelihood ratio statistics is:


SLR = sup

LR( )
V ( )

(4.17)

87

Hansen proves that Pr( SLR x) is bounded by an asymptotic distribution:


Pr(sup Qn x) pr (sup Q x)

(4.18)

where the distribution Qn is defined by:


Qn =

LRn ( ) E[ LRn ( )]
(4.19)
Vn ( )

Under the null hypothesis, E[ LRn ( )] 0 . Assuming an empirical Central Limit


Theorem holds, Qn ( ) Q( ) which is a Gaussian process with a known covariance
function. Then the distribution Qn ( ) can be simulated and the supremum obtained by
taking over all possible values of . This test is computationally intensive as we have
seen that a huge simulation is need to derive the Qn ( ) distribution with all possible
undefined parameters.

4.2 Regime Switching in Commodity Futures

While the MRS has become popular for macroeconomic and financial studies,
commodity futures research with this model is still new. Chow (1998) applied the regime
switching model to examine the co-integration problem of spot and future prices for four
precious metals (gold, silver, palladium, and platinum). He concluded that when taking
the regime switching into account, the null hypothesis that spot and futures prices are
co-integrated cannot be rejected. Fong and See (2001, 2003), using Gray (1996)s
switching GARCH model, provided clear evidence of regime shifts in conditional mean
and volatility for the daily return series of commodity future index and oil futures
contracts. Consistent with the theory of storage, the regime transition probabilities are
functions of the basis, and returns are more likely to switch to or remain in the
high-volatility state when the basis is negative than when the basis is positive. Alizadeh
and Nomikos (2004) investigated the performance of the time-varying hedge ratios

88

generated from a regime switching model. They believed that the dynamic relationship
between spot and futures returns may be characterized by regime shifts, which in turn
suggests allowing the hedge ratio to be dependent upon the state of the market.
Tests for regime-switching in commodity futures returns

The dynamic behavior of commodity future returns is quite different during a


disequilibrium market conditions and a relative equilibrium market. Different versions of
storage theory have showed this asymmetric behavior on commodity futures. The section
examines commodities regime switching behavior using a statistic approach. To
determine whether there is switching in commodity futures index returns, we consider
four specifications. First, we assume that commodity futures index returns are drawn
from a single Gaussian distribution with mean 0 and variance 0:
rt = 0 + 0 t (4.20)
where is i.i.d standard normal variable. This is the specification for the null
hypothesis of no switching.
Then we examine three alternative hypotheses:
1. regime switching in mean: rt = 0 ( st ) + 0 t (4.21)
2. regime switching in variance: rt = 0 + 0 ( st ) t (4.22)
3. regime switching in both mean and variance: rt = 0 ( st ) + 0 ( st ) t (4.23)
where st is the state variable equal to 1 (when it is in regime 1 at time t ) or 2( when it is
in regime 2 at time t), and the regime switches following a first-order Markov chain:
Pr( st = 1/ st 1 = 1) = p11
Pr( st = 1/ st 1 = 2) = p21
Pr( st = 1/ st 1 = 2) = p21
Pr( st = 2 / st 1 = 2) = p22
p11 + p12 = 1
p21 + p22 = 1

(4.24)

89

Table 4-1 presents the log-likelihood for the null hypothesis and each alternative
as well as the simulation results of LR test statistic value. Garcia (1992) has shown that
the 1% critical value for Hansens standardized LR statistic is 13.81% and 5% critical
value is 10.34. Our simulation test results strongly reject the null hypothesis of
non-regime switching on all of the three commodity futures index returns. The likelihood
ratio statistic for testing first alternative (regime switching in mean) against the null
(non-regime switching) is 107.24, 94.85, and 104.20, respectively, for GSCI composite,
energy, and non-energy. All of these values are much higher than the 5% critical value or
1% critical value. The results in column 3 (regime switching in the variance against the
null) and in column 4 (regime switching in both mean and variance against the null)
imply even stronger rejection of the null hypothesis. It is reasonable to conclude that the
regimes are detected.
Diagnostic tests of Markov switching specification

The LR test has shown to favor the specification of Markov regime switching
model as equation 4.24. Hamilton (1991) EM algorithm has applied to estimate the
model for commodity futures. The estimation results have been reported in Table 4-2. We
can observe that there are two distinctive regimes for commodity futures index: Regime
1 represents the relatively low returns and less volatility market and regime 2 is
characterized as the high return and high volatility market. Meanwhile, regime shows
strong persistence as the estimators of the transition probabilities are close to 1. The
estimation results imply that the commodity futures index return is characterized by a
state in which market is relatively stable and as a result the expected returns is relatively
low and a state in which the expected returns is much higher meanwhile the market is
also more volatile. This implication is consistent with the risk premium theory.
Recall that this regime switching specification is based on the assumption that
is an i.i.d standard normal variable. Hamilton (1996) has developed a series individual
tests for the serial correlation (AR(1)), ARCH(1), and higher-order Markov effects for all

90

these forms of misspecification of Markov regime switching model. His method is used
to conduct diagnostic tests on the specification. Table 4-3 presents individual tests results.
The diagnostic tests show no evidence of omitted AR and ARCH effects, i.e., our regime
switching specification can produce the autocorrelation of the level and squares of
commodity futures return time series. This result is consistent with the results of
Hamilton and Susmel (1993) and Timmermann (2000). Hamilton and Susmel (1993)
found that Markov regime switching models do a better job of capturing financial assets
return than previous ARCH and GARCH specifications. 8
Figure 4-1 plots the smooth probability of state 1 at time period t. 9 Most of
regime switching for commodity futures is driven by the incidents in the real economic
sector. For example, we are able to capture the depression of the oil price shocks of
1973-1974 and 1979-1980, the serious droughts in 1986, and the Gulf war in 1991 as
well as the sharply rising oil price at early year of this century. During the 1970s which is
called stagflation chartered by continuing inflation and stagnant business activity,
commodity futures markets remain in regime 2, the increased demand pushing up
commodity prices and commodity futures markets expecting high returns. By 1983,
inflation had eased, the economy had rebounded, and the United States began a sustained
period of economic growth, then commodity futures markets displays high probability to
stay in regime 1.

4.3 Single Regime Switching Model for Three Asset Classes

This section discusses the relationship between commodity futures regimes


switching behavior with stock and bond regime switching. The model specification of
equation 4.24 is used to estimate the three assets monthly returns. The estimates results
8 As Hamilton and Susmel (1993) have shown, if there were regime switching in the data-generating process, this
would induce a non-linearity that could show up in ARCH tests, since ARCH tests can detect non-linear structure in
time series.
9
It is the estimated probability based on full sample information. (see Kim (1994) for detail)

91

are reported in Table 4-4. Commodity futures and bonds display similar regime
switching behavior: their returns are characterized by two states: a regime with relatively
lower positive returns and less volatility and a regime with relatively higher positive
return as well as higher risk. Stock return is characterized by a bad regime in which the
mean return is negative and market is relatively more volatile and a normal regime in
which expected return is positive and markets volatility is also relatively low.
The difference of regime behavior between commodity futures and stocks could
be important to strategic asset allocation. We expect that commodity futures are in high
return regime when stocks are in bad state, and then we can achieve an efficient
diversification benefit. Table 4-5 reports the correlation of the smooth probability of state
1 at time period t for these three assets. The GSCIs regime switching behavior exhibits a
positive correlation with that of stock and an insignificant correlation with bonds. The
value of correlation between GSCI and bond is 0.42. This positive correlation of regime
switching behavior implies that when the stock market transfer into bad regime,
commodity futures have a greater chance to switch to a higher return regime.
Commodity futures do display diversification benefits in the regime switching sense.
Figure 4-2 presents the state 1 smooth probability for three assets returns.
Model specification and estimation

A multivariate regime switching model is to estimate the joint distribution of


three asset returns. Given the inconsistence of regime switching behavior among three
asset classes, a complicated high-order regime switching model might be sufficient to
capture the data characters. But as only 419 month return observations are covered in the
sample, e a simple two regime model is proposed as follow: 10
Rt ( st ) ~ MVN ( (1n ) ( st ), ( st )( st ) ') (4.25)
i.e., three asset returns are assumed to follow a multi-variant normal distribution with
10

Regime switching model has a problem of over-parameter. A model with more than 2 regimes could be
impossible to estimate efficiently.

92

two regimes.
Table 4-6 presents the full sample estimation results of my simple two regimes
model. These results suggest that both returns and variance-covariance of three asset
classes are regime-dependent. There asset classes have better expected return in regime 1
compared to regime 2. The annualized monthly returns for commodity futures, stocks
and bonds are 14%, 17%, 9%, respectively at state 1 and 11%, -2%, 7%, respectively at
state 2. The mean returns for commodity futures and bonds do not change significantly.
But stock shows significant changing. Commodity futures have the comparable
performance with stock at state 1 and better expected return than stock at state 2.
Commodity futures and stocks display much higher volatility in state 2 than in state 1
while bonds return variation in state 2 is just above half of the variation in state 1. In
both regimes, commodity futures have the highest return variation.
The Sharpe ratios of these three asset classes in different regimes are presented in
Table 4-7. The risk/return tradeoff of these three asset classes is much better at state 1
than at state 2. Stocks achieve the best return/risk tradeoff at state 1 but have the worst at
state 2. Bonds have the least level of Sharpe ratios at state 1 and the highest level at state
2. We can interpret that the first state corresponds to the normal market condition and
the second state to the bad one. One amazing feature for commodity futures returns is
that they display good consistence with regime switching, the expected return for
commodity futures is good in both regimes. A relatively strong performance in regime 2
for commodity futures makes them more valuable at this regime for investors to achieve
diversification benefits.
The regime-dependent covariance matrixes are worth noting. The covariance and
correlation coefficient matrixes are presented in Table 4-8. Commodity futures positively
relate with stock in state 1 and negatively relate in state 2 with the level of correlation
coefficient is negligible in magnitude. In the normal market condition, bond market and
stock market is highly positively correlated, but in the bad market condition, these two

93

markets show substantial negative correlations. The correlation between stocks and
bonds is 0.4369 at state 1 and -0.2165 at state 2. The correlation between commodity
futures and bonds is consistently negative in both regimes. The negative correlations and
high assets return volatility at state 2 imply that at this state it might be more important
for investors to diversify investment among these three assets class to reduce risk.
Note that there is only one regime variable in this model. An interesting question
is what this regime variable denotes or whose regime it is. It is obvious that this regime
variable does not mean the commodity futures market regime, or the stocks and bonds
regimes. We have learned that the regime switching behavior of these three asset classes
is not concurrent and can not be denoted by a single variable. However, it is almost
impossible to estimate a model with three different regime variables. For convenience,
the regime variable in my model is interpreted as an unobservable market power which
drives the integrated financial market switching between normal and bad, and as a result,
the performance of different asset classes will also varied with this latent variable.
The transition probabilities in Table 4-6 show that the first state is highly
persistent compared to the second. We can compute the expected duration of the regime
by 1/ (1 - pii) 11which shows the regime persistence. Regime 1 has an expected duration
more than three years while regime 2 has an expected duration about 15 months. In other
words, we can expect that normal market condition to last longer than bad market
condition. Figure 4-3 presents the smooth probability of regime 1 for time period t over
the sample period. An interesting fact is that many of the bad market conditions are
triggered by the incidents in the real economic sector. There is the clear spike at the stock
market crash October 1987 as well as a spike that coincides with the end of the bubble
economy at the end of 20th century.

4.4 Optimal Portfolio Allocation without Risk-free Asset


11

see Kim and Nelson (1999), pages 71-72

94

The significant changing of risk/return and correlation structure implies the


importance of regime switching in strategic portfolio selection. A regime-switching
mean-variance model is proposed to study the one-period optimal portfolio selection.
The first case examined here is portfolio selection among commodity futures, stocks, and
bonds; no risk-free asset is included. Under the mean-variance model, the optimal asset
weight vector will solve the following objective function:

max p ( w) * p2 ( w)
N

s.t.

w =1
i =1

(4.26)

where:
N

p ( w) = wi i
i =1
N

( w) = wi w j ij
2
p

(4.27)

i =1 j =1

denotes investors risk aversion and N=3 in our case. We can also add a constraint on
the asset weight as 0 wi 1 to restrict short sale. Using different input variables

it , it2 , and ijt will lead to different optimal weight solutions.


In order to examine the mean-variance strategy under different specifications,
three different set of input variables of it , it2 , and ijt are used to solve the problem: (a)
the historical mean and variance-covariance over full sample period, (b) the mean and
variance-covariance at state 1, (c) the mean and variance-covariance at state 2. The first
setting does not admit the regime switching effect. Figure 4-4 and Figure 4-5,
respectively, present the efficient frontier and optimal asset weight for these three
settings.
In Figure 4-4, the blue line in the middle represents the efficient frontier using the
historical sample moments, ignoring the regime switching effect. The green line near the

95

top represents the efficient frontier in normal regime (state 1) while the red one in the
bottom is the one applicable for bad regime (state 2). The risk-return tradeoff is generally
better in normal regime than in bad regime. Since the historical sample moments average
the asset performance, the risk-return tradeoff using historical average information
should be in the middle. A mean-variance investor who allocates his capital according
the historical sample moments will expect to either have higher risk or lower return than
the same type investor but accounting for regime switching. The Sharpe ratio results are
reported in Table 4-9. The significant risk-return tradeoff difference between the
historical average and the regime switching is revealed. For a wide range of risk aversion
coefficients, the Sharpe ratio level based on historical average is about 50% percent less
than that in regime 1 and 100% higher than that in regime 2. This significant difference
implies the importance of regime switching in portfolio selection.
Another important difference worth noting is the optimal weight solutions under
different specifications. Figure 4-5 plots the optimal weight vector with the risk aversion
for these three cases. Both constrained and unconstrained optimal weights are presented.
Panel A and Panel B present optimal weight results obtained with historical average
information. For higher risk tolerance investors who have a small risk aversion
coefficient, the optimal weights on three asset classes are almost identically equal to
one-third; they should equally allocate their capital on these three assets. As the risk
aversion level increases, investors have less ability to stand risk and increase their capital
allocation on bonds while decreasing the allocation to commodity futures and stocks.
The optimal weight to commodity futures is always greater than that placed on stocks.
The picture of optimal weights of three assets is totally different when the regime
switching behavior is considered. At state 1, stocks have very good performance but
bonds are performing poorly, high risk tolerant investors shorting bond to obtain more
exposure on the stock market. As the risk aversion level starts increasing, the optimal
weight on stocks sharply decreases and the weight on bonds increases. When the risk
aversion level reaches the high level (around = 10 area), invertors evenly allocate

96

capital on three risky assets. The optimal weight on commodity futures is relatively
stable, ranging from 30% to 40%.
On the other hand, at state 2, stocks show poor performance and bonds are
relatively good. Investors with high risk tolerant will short in the stock market in order to
have more exposure in the bond market. As investors risk aversion starts increasing, the
optimal weight on bonds slowly decreases and the weight on stock slowly increases.
High risk aversion investors keep about 80% capital on bonds, 10% on stock, and the
other 10% on commodity futures. The optimal weight on commodity futures is relatively
stable, ranging from 10% to 20%.
The optimal weight to commodity future is always positive and significant under
all the circumstances, which strongly supports that commodity futures should be
included into financial asset portfolio. Moreover, the optimal weight on commodity
futures remains stable and does not show much variation with risk aversion, which is in
contrast with the results of prior researchers. For example, Anson (1999) found that the
more risk averse, the higher benefit obtained by commodity futures. Instead, our results
support that a high risk aversion investor should allocate less capital on commodity
futures investment. On average, a low risk aversion investor should invest about 30% his
capital in commodity futures while a high risk aversion investor will keep about 10%
capital allocation to commodity futures investment.
If a mean-variance investor ignores the regime switching effects and allocates the
investment using the historical average information, Table 4-10 shows that he will
reward a poor risk-return tradeoff no matter in what market conditions. The improvement
of Sharpe ratio is especially significant in regime 2. For example, at risk aversion level 2,
the Sharpe ratio is only improved by 1% at state 1; but at state 2, the Sharpe ratio is
increased by about 10 times. Accounting regime switching effects are more important
when financial markets move from the normal regime to the bad regime.
We can compare portfolio performance for different capital allocation strategies

97

by assuming the data generate process to be MRS process. Figure 4-6 presents the results
for a range of risk aversion coefficient values. Investors can reduce the investment risk
using the non-regime switching strategy when the market is in regime 1; however, lower
expected return is also resulted. The risk and return are not significantly different. It is
more difficult for investors following the non-regime strategy in regime 2 when the
market is in bad state. The investment risk is significantly increased and the expected
return is reduced if a non-regimes investment strategy is followed.
The importance of adding regime switching into the mean-variance portfolio
selection has been showed above. However, there are some concerned about the MRS
model. This model may imperfectly match the true state-dependent moments of asset
returns because of sampling error and the possibility of model misspecification. As a
result, the performance of the regime switching portfolio selection can not be guaranteed.
In order to justify the significant impact of regime switching on static portfolio selection,
an out-sample test is needed. We can make the test by using a set of asset allocation
simulations based on different investment strategies. Let us assume investors with the
different risk aversions have initial capital 100 dollars and they start trading on
commodity futures, stocks and bonds in November 1994. The total investment horizon is
10 years, i.e. 120 months. At the end of each month (time period t), the investors make
investment decisions for next month (time period t+1) based on the information up to
current time period t and follow three different portfolio allocation strategies: strategy 0,
strategy 1, and strategy 2.
Strategy 0 is the non-regime strategy, which uses the historical average mean and
variance-covariance up to period t to obtain the optimal weights solution. Strategy 1 is a
regime switching strategy. Following this strategy, an investor infers the regime
probability from the current information. In particular, he computes: Prob(st/It) (where It
denotes the information set up to time period t), which is a by-product of the
optimization of the MRS model. When this probability is greater than 0.5, the investor
classifies the regime as 1, otherwise he classifies it as 2; meanwhile he assumes that the

98

regime will not change because of the high persistence of regime. Strategy 2 is another
regime switching strategy, which realizes the regime at current period in the same way as
strategy 1 does. However, when making decision for next month asset allocation,
strategy 2 derives the expected return vector and variance-covariance matrix as
following:
e1 ( st = 1) = p11 * 1 ( s = 1) + (1 p11 ) * 2 ( s = 2)
e2 ( st = 2) = p22 * 1 ( s = 2) + (1 p22 ) * 2 ( s = 1)
1 ( st = 1) = p11 * 1 ( s = 1) + (1 p11 ) * 2 ( s = 2) + p11 *(1 p11 ) *(e1 e2 )(e1 e2 ) '

(4.28)

2 ( st = 2) = p22 * 2 ( s = 2) + (1 p22 ) * 1 ( s = 1) + p22 *(1 p22 ) *(e1 e2 )(e1 e2 ) '

where 1 , 2 , 1 , 2 respectively, denote the estimated mean return vector and


variance-covariance matrix in regime 1 and in regime 2, e1 (e2 ) represents the expected
next period mean returns when current period is at state 1 (state 2). 1 (2) represents
the expected next period covariance matrix when current period is at state 1 (state 2).
The third term in the expected covariance matrix equation captures the jump caused by
the conditional change of mean return from one regime to the other. Thus strategy 0
considers the assets return as a simple Gaussian process; strategy 1 realizes the regime
switching behavior of asset return but does not account the future regime switching
effect; strategy 2 not only accounts for the regime switching in the past but also expects
future regime switching.
Two simulation results over the 10 years period are recorded: realized one-period
return and accumulated wealth. As both regimes show fair strong persistence, we do not
observe much difference between the results for two regime switching strategies
(strategy 1 and strategy 2). Figure 4-7 reports the mean and standard deviation of the
annualized one-period return for different risk aversion levels. Following the regime
switching strategies (strategy 1 and strategy 2) always results in better return. The higher
risk tolerance, the greater the return improvement by sticking to regime switching
strategies. For example, at the risk aversion coefficient level 2, the average one-period
return is increased by almost 50%. As the risk aversion level increases, this return

99

improvement effect becomes less noticeable. For high risk aversion investors (risk
aversion coefficient larger than 15), there is only a slight increase in return. An investor
can achieve much higher level of return if there is no weight constraint applied. While
the regime switching strategies can lead to better investment returns, they also increase
the investment risk. The standard deviation of the one-period return corresponding to
regime switching strategies is much greater than the one for the non-regime strategy,
especially in the unconstrained case. For example, at risk aversion level 2, the standard
deviation is doubled. When risk aversion increases, the difference becomes unnoticeable.
Our simulation results show that the regime switching strategies rise up both
return and risk level. Let us examine the risk-return tradeoff results to see whether
regime switching strategies can achieve better investment performance. Table 4-11
reports the Sharpe ratio for different risk aversion levels. The Sharpe ratio results support
that regime switching strategies are superior to non-regime switching strategy since they
achieve a higher Sharpe ratio level in all circumstances. Especially, this benefit is most
obvious at the middle risk aversion level. For example, at risk aversion level 5, the
Sharpe ratio is increased by about 40% by taking regime switching strategies. As most
institutional investors have middle level risk aversion, regime switching strategies are
very meaningful in practice.
Another fact which can support the regime switching strategies is that they can
achieve a greater accumulated wealth. Figure 4-9 shows how wealth cumulates over
time for three different strategies and Table 4-12 reports the end accumulated wealth
after 10 years investment simulations. Regime switching strategies consistently
outperforms non-regime strategy, especially at high risk aversion case. The accumulated
wealth lines for the regime switching strategies consistently lie above the line for the
non-regime strategy. Again, the difference is decreased as risk aversion level increases.
At extreme risk aversion level, the accumulated wealth does not present significant
differences among all the three strategies. An investor with high risk tolerance ability can
improve his accumulated wealth by nearly 30% if he follows the regime-dependent

100

investment strategies. An investor with mid level risk aversion can improve the
accumulated wealth by about 15-30%. In conclusion, the accumulated wealth results
strongly support that the regime switching strategies outperform the non-regime strategy.
Table 4-13 presents the average weight on commodity futures over the 10 years
investment simulation. On average, the optimal weight on commodity futures is located
between 20% and 35% and the specific level is varied with the investment strategy and
risk aversion coefficient. The higher risk aversion level, the less capital is allocated to
commodity futures. One interesting result is the average weight on commodity futures is
the almost same for different investment strategies. However, this similarity in the
average weight does obscure the striking difference of the allocation to commodity
futures between the non-regime strategy and the regime switching strategies. Figure 4-10
shows that the optimal weight on GSCI in regime switching strategies is highly fluctuant,
especially during the period from 40th to 80th month, which corresponds to the bubble
economy period started in 1998 and ended in 2002. During this time period, the financial
market experienced a transition from boom to bust, and asset returns displayed
significant variation. As a result, the optimal solution on asset weights significantly
varied under the regime switching strategies. The non-regime switching strategy uses the
historical average information and has much more stable estimates of asset returns and
variance, consequently, leading to a more stable optimal weight solution. Regime
switching investment strategies increase the asset allocation to commodity futures when
the financial market enters the poor-performed regime (such as the period of early 21st
century). As a result, the accumulated wealth achieved by regime switching strategies is
significantly increased. This result implies how important it is to recognize the regime
switching. 12

12

The large increase in commodity futures weight at the end of the simulation periods is mostly triggered by rally of
energy price rising. Since energy products account for more than 70% weight in GSCI, the role of energy products
on GSCI is important. Energy products are also the most active traded commodities.

101

4.5 Optimal Portfolio Allocation with Risk-free Asset

This section examines the static portfolio selection problem with three risky
assets and a risk-free asset: a commodity futures index (GSCI), an equity index (all for
the US), a 10-year T-bond, and a 30-day T-bill. Adding a risk-free asset makes the
mean-variance asset allocation totally different. The correlation of any risky asset with a
risk-free asset is zero, so that any combination of a risky asset or portfolio with the
risk-free asset generates a linear return and risk function. Therefore, combining the
risk-free asset with any point on the efficient frontier will derive a set of straight-line
portfolio possibilities. The line tangent to the efficient frontier will dominant all other
line. This dominant line is called capital market line (CML) and all investors should
target along this line by borrowing or lending money at the risk-free rate, which will
depend on their risk preference. Figure 4-11 gives two examples in our case, the optimal
risky asset allocation is marked by a black sign * and the optimal portfolio allocation
for a given risk aversion level is marked by a red sign +. Panel A presents the optimal
asset allocation for high risk aversion investors and Panel B represents low risk aversion
investors. The higher risk aversion level, the more the capital is allocated to the risk-free
asset. The lower risk aversion level, the more a risk-free asset is borrowed.
Optimal asset allocation will depend on three factors: the risk-free rate, the risk
and expect return of risky assets, and the covariance among the risky assets. Any change
in these three input variables results in different optimal asset allocation. Table 4-14
represents the results of optimal risky asset allocation following different investment
strategies. Without for accounting the regime switching effect, an optimal investment
should allocate about 48% percent of the risky asset fraction to bonds, 29% to GSCI, and
23% to stocks. The results are totally changed when the regime switching effect is
considered. At state 1, about 54% percent of the risky asset fraction should be allocated
to stocks, 36% to GSCI, and 10% to bonds; at state 2, an optimal risky asset investment
allocates no risky asset investment to stocks, about 80% to the bond market, and 20% to
commodity futures.

102

Not only is the set of optimal risky allocation significantly changed, but also the
fraction of the risky and risk-free asset. Figure 4-12 presents the overall optimal asset
allocation for different risk aversion levels. The overall optimal asset allocation is
changed under different investment strategies, especially when the risk aversion level is
less than 15. In general, state 1 is a relatively favorable regime for risky asset investment
and state 2 is a relatively unfavorable regime. We expect more risky asset investment in
first regime and less risky asset investment in second regime. For example, at risk
aversion level 2, the optimal risk-free asset fraction is -2.5 for the non-regime strategy,
i.e., a mean-variance investor who does not consider the regimes switching effect should
borrow 2.5 times capital to invest on risky assets. The risk-free fraction level is about -4
at state 1 and -1 at state 2. In the non-regime switching case, the optimal weight on the
risk-free asset becomes positive at risk level 8. This threshold risk aversion level is 12 in
regime 1, and 4 in regime 2. The optimal fraction of the risk-free asset is sharply
increasing with the risk aversion level.
Again, we note that the optimal weight to commodity futures is always positive
and significant under all the circumstances. Commodity futures should also be included
into financial asset portfolio when a risk-free asset exists. An investor should invest
about 35% his risky asset investment in commodity futures when the financial market is
in normal regime and 20% when the financial market is in bad regime.
The significant differences of optimal risky asset allocation and risk-free asset
fraction imply that if the true data-generating process of asses return displays regime
switching behavior, then a mean-variance investor who ignores the regime switching
effects and allocates the investment using the historical average information, in general,
will reward a poor risk-return tradeoff. We can overview the regime switching effect on
portfolio selection by examining the performance of optimal portfolio constructed with
different. Figure 4-13 presents the risk and return of different optimal portfolios. Regime
0 line denotes the performance of optimal portfolio constructed with the non-regime
strategy. Regime 1 line denotes the performance of optimal portfolio constructed with the

103

optimal weight vector in regime 1. Regime 2 line denotes the performance of optimal
portfolio constructed with the optimal weight vector in regime 2. A mean-variance
investor following the non-regime switching strategy at state 1 achieves a lower expected
return. The return has been decreased by about 50% to 10%. The lower risk aversion
level, the greater reduction in return. The investment risk has been also decreased. In
regime 2, investors following non-regime strategy will achieve higher risk and lower
returns.
Table 4-15 reports the risk-return tradeoff results. Without accounting regime
switching effect, the Sharpe ratio value of optimal portfolio is only deceased by about
11% in regime 1, but is 5 to 7 times reduced in regime 2. For example, at risk aversion 2
level, the Sharpe ratio is only improved by 10% at state 1, however, at state 2, the Sharpe
ratio is more than 6 times increased. Again, our results suggest that regime switching
effects are more important when financial markets move from the normal regime to the
bad regime.
To justify the significant effect of regime switching, again, a series of out-sample
asset allocation simulation is conducted following the same scheme as without risk-free
asset case. Figure 4-15 reports the mean and standard deviation of annualized one-period
return for different risk aversion levels. Following the regime switching strategies always
results in better one-period return. The higher risk tolerance ability, the greater
improvement in return by sticking to regime switching strategies. For example, at the
risk aversion coefficient level 2, the average one-period return is increased by. As the risk
aversion level increases, the improvement becomes less observable. The regime
switching investment strategies also slightly increase the investment risk at the same
time. The standard deviation of one-period return corresponding to the regime switching
strategies is about 10% larger than the one for the non-regime strategy. For example, at
risk aversion level 2, the risk level denoted by standard deviation is 0.14 when the
non-regime strategy is applied and is increased to 0.172 by the regime strategies. For
high risk aversion investors (risk aversion coefficient larger than 25), there is almost no

104

increase of risk level by taking regime switching strategies.


Let us examine the risk-return tradeoff results to see whether regime switching
strategies can achieve better investment performance. Table 4-16 reports the Sharpe ratio
level for different risk aversion levels. The results demonstrate that the regime switching
strategies are superior to the non-regime switching strategy since their optimal portfolios
obtain a higher Sharpe ratio level at any risk aversion levels, especially at the middle risk
aversion level. For example, at risk aversion level 8, the Sharpe ratio is increased by 12%
by taking regime switching strategies. On average, following strategy 1 can improve the
optimal portfolio performance by 8% and following strategy 2 can improve the optimal
portfolio performance by 10%.
Let us consider the results of accumulated wealth by different investment
strategies. Figure 4-16 shows how wealth is cumulated over time for three different
strategies and Table 4-17 reports the end accumulated wealth after 10 years investment
simulations. The regime switching strategies consistently gain higher accumulated
wealth than non-regime strategy, especially for high risk aversion investors. For example,
at risk aversion level 2, an investor with $100 initial capital could gain about $3000
accumulated wealth if he follows the regime switching strategies; instead, he can only
obtain $1500 accumulated wealth if he adopts the non-regime switching strategy. At a
middle risk aversion level, the accumulated wealth still can be improved by about 50%
on average. These results imply that for most investors, following regime switching
strategies to optimally allocate portfolio could substantially improve the investment
return.
Table 4-18 reports the end accumulated wealth over 10 years investment
simulation. Investors with a higher risk tolerance ability characterized by a lower risk
aversion achieve a higher accumulated wealth. For example, at risk aversion level 2, the
accumulated wealth for strategy 0, strategy 1, and strategy 2 is 1393.7, 2376.8, and
2228.1, respectively; but at risk aversion 15, the accumulated wealth for strategy 0,

105

strategy 1, and strategy 2 is 224.22, 261.98, and 252.72, respectively. An investor with
high risk tolerance ability can improve his accumulated wealth by about 100% if he
follows the regime-dependent investment strategies. For investors with middle risk
aversion level, he still can improve the accumulated wealth by about 15-30%. In
conclusion, the accumulated wealth results of investment simulation strongly support
that the regime switching strategies outperform the non-regime strategy.
What is the role of commodity futures in our investment simulation? Table 4-18
presents the average weight on commodity futures over the 10 years investment
simulation under different circumstances. Commodity futures have been showed to be a
significant part of optimal portfolio. The average level of optimal weight on commodity
futures varies with risk aversion coefficient. The higher risk aversion level, the less
capital is allocated to commodity futures. Again, all three strategies show similar patterns
of the optimal weights on commodity futures. However, this similarity in the average
weight does obscure the striking difference in the allocation to commodity futures.
Figure 4-17 shows that the optimal weight on GSCI in regime switching strategies is
highly fluctuant, especially during the period from 40th to 80th month, which corresponds
to the bubble economy period. During this time period, the financial market
experienced a transition from boom to bust, and asset returns displayed significant
variation. As a result, the optimal solution on asset weights significantly varied with the
regime switching. On the other hand, the non-regime switching strategy uses the
historical average information and has much more stable estimates of asset returns and
variance, consequently, its optimal weight solution is more stable. Regime switching
investment strategies significantly increase the asset allocation to commodity futures
when the financial market is poorly performed and lead to a higher accumulated wealth.

4.6 Conclusion

This chapter applies a single two-regime switching model to examine the

106

dynamics of asset return and to investigate the portfolio selection problem. Both mean
and variance of commodity futures are regime-dependent just as stocks and bonds.
Meanwhile, the regime switching movements of the three asset classes display
independence and uniqueness. To study the optimal asset allocation under regime
switching dynamic, it is assumed that an unobserved variable determines the integrated
financial market regime switching by following a first-order Markov chain. The
estimation results reveal two regimes existing in three risky assets dynamic: a normal
regime characterized by relatively high return and low risk, and a bad regime
characterized by relatively low return and high risk. If the regime switching model is the
true data generating process, an investment strategy ignoring the regime switching effect
will lead to inefficient capital allocation and achieve poor investment performance. To
justify this argument, a 10-year out-of-sample investment simulation is run and the
results strongly support that regime switching effect is important to optimal asset
allocation, especially at the bad market condition. Meanwhile, commodity futures are a
significant part optimal portfolio under all market conditions. Regime switching
investment strategies increase the asset allocation to commodity futures when the
financial market perform poorly and then investors can achieve better investment
performance. Moreover, the diversification benefits of commodity futures are not
identical to different investors. Investors who have higher risk tolerance ability allocate
more capital to commodity futures and benefit from commodity futures investment.

107

Table 4-1: Hansens LR tests for regime-switching GSCI indexes

Log-likelihood
Likelihood ratio
Log-likelihood
Likelihood ratio
Log-likelihood
Likelihood ratio

GSCI COMPOSITES
Null hypothesis
1025.12
GSCI ENERGY
854.21
GSCI NON-ENERGY
1004.78

Alternative 1
1147.80
107.24

Alternative 2
1254.14
287.40

Alternative 3
1307.48
308.18

985.47
94.85

1047.24
118.26

1087.54
148.53

1094.21
104.20

1187.51
147.58

1218.35
225.47

Table 4-2: Estimates of regime switching model for commodity futures index
GSCI return
GSCI energy return
GSCI non-energy return

(s=1)
0.1233
0.0318
0.1203
0.0733
0.0623
0.0201

(s=2)
0.1443
0.0654
0.4375
0.1203
0.1933
0.0850

Note: standard error is reported under the estimate.

P11
0.9810
0.0123
0.9489
0.0387
0.9971
0.0030

P22
0.9755
0.0186
0.7381
0.2414
0.9875
0.0111

(s=2)
0.1803
0.0218
0.7538
0.1486
0.1217
0.0100

(s=1)
0.7075
0.0977
3.2369
1.4382
0.7142
0.1038

108

Table 4-3: Hamiltons LM test for model misspecification


TEST
LM test AR(1) in regime 1

GSCI
0.0454
0.8312
3.2900
0.0697
1.7686
0.1835
0.6358
0.4252
0.0841
0.7716
0.0705
0.7904
0.7866
0.9402
0.1385
0.9330
0.6653
0.7165

LM test AR(1) in regime 2


LM test AR (1) across regimes
LM test ARCH(1) in regime 1
LM test ARCH(1) in regime 2
LM test ARCH(1) across regimes
White test of Markov specification
Higher-order Markov effects in regime 1
Higher-order Markov effects in regime 2

ENERGY
2.0221
0.1560
3.3139
0.0698
2.8726
0.0925
0.0055
0.9450
3.6016
0.0574
1.2845
0.2141
5.5395
0.2361
2.3346
0.3321
2.0513
0.3547

NON-ENERGY
0.0200
0.9145
1.6352
0.1941
0.5656
0.45478
1.6265
0.1974
0.1412
0.7147
0.5963
0.44875
18.860
0.0011
1.6010
0.8003
2.5647
0.3478

Note: p-value is reported under the estimated statistic.

Table 4-4: Regime switching model estimates for GSCI, Stocks and Bonds
P11
(s=1)
(s=2)
GSCI return
0.1233
0.1443
0.9810
Standard error
0.0318
0.0654
0.0123
Stock return
0.1862
-0.1176
0.9489
Standard error
0.0296
0.1514
0.0288
Bond return
0.0711
0.1223
0.9892
0.0135
0.0367
0.0069
Note: standard error is reported under the estimate.

(s=2)
0.1803
0.0218
0.1935
0.0268
0.0479
0.0052

P22
0.9755
0.0186
0.8298
0.1162
0.9817
0.0137

Table 4-5: Correlation of transition probability among three asset classes


Stock

GSCI

Bond

Stock

GSCI

0.4289

Bond

0.0481

-0.2442

(s=1)
0.7075
0.0977
0.6456
0.1463
0.1633
0.0234

109

Table 4-6: Parameters estimation of simple MRS model for three assets
Means for state 1
GSCI
0.1405
(0.0308)
Means for state 2
GSCI
0.1121
(0.0931)

Stocks
0.1719
(0.0271)

Bonds
0.0938
(0.0186)

Stocks
-0.0197
(0.0616)

Bonds
0.0733
(0.0242)

Prob from state 1 to state 1 (p)


0.973 (0.0127)
Expected duration of state 1: 38.28
Prob from state 2 to state 2 (q)
0.935(0.0280)
Expected duration of state 2: 15.55
Variance-covariance matrix for state 1
GSCI
Stocks
0.2204
0.0210
(0.0229)
(0.0085)
0.0210
0.2085
(0.0085)
(0.0131)
-0.0067
0.0602
(0.0091)
(0.0081)

Bonds
-0.0067
(0.0091)
0.0602
(0.0081)
0.0930
(0.0171)

Variance-covariance matrix for state 2


GSCI
Stocks
0.8809
-0.0524
(0.1215)
(0.0709)
-0.0524
0.5325
(0.0709)
(0.0641)
-0.0303
-0.0371
(0.0088)
(0.0175)

Bonds
-0.0303
(0.0088)
-0.0371
(0.0175)
0.0554
(0.0209)

Value of log likelihood function:

410.447

Note: All return, variance and covariance results are annualized. Standard errors are reported in the parenthesis. The
expected durations for each state is calculated as 1/(1-pii).

Table 4-7: Sharpe ratio of three asset classes


State 1
GSCI
0.1929

stocks
0.2671

bonds
0.1439

State 2
GSCI
0.0662

Stocks
-0.0955

Bonds
0.0993

110

Table 4-8: Correlation among three asset classes


State 1
1.0000
0.0984
-0.0468

0.0984
1.0000
0.4329

State 2
1.0000
-0.0766
-0.1372

-0.0468
0.4329
1.0000

-0.0766
1.0000
-0.2165

-0.1372
-0.2165
1.0000

Table 4-9: Sharpe ratio comparing under different regimes


Without weight constraint

risk aversion

non-regime

regime 1

With weight constraint


regime 2

non-regime

regime 1

regime 2

0.205

0.315

0.125

0.205

0.275

0.144

0.213

0.317

0.127

0.213

0.317

0.135

10

0.210

0.312

0.113

0.210

0.312

0.113

15

0.207

0.301

0.102

0.207

0.301

0.102

20

0.206

0.293

0.096

0.206

0.293

0.096

50

0.202

0.274

0.083

0.202

0.274

0.083

Table 4-10: Sharpe ratio comparing for optimal portfolio without risk-free asset
Sharpe ratio comparing in regime 1
with constraint
2

10

15

20

50

regime 0

0.309

0.283

0.260

0.252

0.248

0.240

regime 1

0.312

0.314

0.309

0.297

0.289

0.270

without constraint
2

10

15

20

50

regime 0

0.309

0.283

0.260

0.252

0.248

0.240

regime 1

0.274

0.314

0.309

0.297

0.289

0.270

Sharpe ratio comparing in regime 2


with constraint
2

10

15

20

50

regime 0

0.012

0.040

0.066

0.072

0.075

0.081

regime 2

0.122

0.123

0.109

0.098

0.091

0.078

without constraint
2

10

15

20

50

regime 0

0.012

0.040

0.066

0.072

0.075

0.081

regime 2

0.142

0.131

0.109

0.098

0.091

0.078

111

Table 4-11: Sharpe ratio comparing for simulated one period returns over 10 years
(without risk-free asset)
with constraint
2

10

15

20

50

strategy 0

0.158

0.144

0.162

0.161

0.160

0.160

strategy 1

0.190

0.208

0.188

0.181

0.176

0.166

strategy 2

0.186

0.188

0.184

0.177

0.173

0.166

without constraint
2

10

15

20

50

strategy 0

0.158

0.144

0.162

0.161

0.160

0.160

strategy 1

0.159

0.192

0.187

0.181

0.176

0.166

strategy 2

0.154

0.171

0.184

0.177

0.173

0.166

Table 4-12: Accumulated end wealth over 10 years simulation without risk-free
asset
Panel A: with weight constraint
Risk aversion

Strategy 0

Strategy 1

Strategy 2

259.3769

318.4708

313.0969

235.4627

306.6017

277.2393

10

236.8888

268.0083

263.5921

50

232.7163

240.2120

238.3693

Panel B: without weight constraint


Risk aversion

Strategy 0

Strategy 1

Strategy 2

259.3769

373.0666

338.9087

235.4627

296.3801

267.0174

10

236.8888

266.9812

263.0047

50

232.7163

240.2125

238.3693

Table 4-13: Average weight on GSCI over 10 years investment simulation


without risk-free asset
Risk
aversion

Strategy 0

Strategy 1

Strategy 2

With constraint

Strategy 0

Strategy 1

Strategy 2

Without constraint

0.3596

0.2705

0.2858

0.3596

0.3040

0.3247

0.2568

0.2460

0.2466

0.2568

0.2464

0.2468

10

0.2299

0.2396

0.2379

0.2299

0.2396

0.2379

15

0.2185

0.2359

0.2218

0.2185

0.2359

0.2218

20

0.2116

0.2321

0.2176

0.2116

0.2321

0.2176

50

0.1992

0.2192

0.2089

0.1992

0.2192

0.2089

112

Table 4-14: Optimal risky asset allocation when a risk-free asset existing
GSCI

Stocks

Bonds

Non-regime

0.2882

0.2308

0.4809

Regime 1

0.3647

0.5427

0.0926

Regime 2

0.1943

0.8057

Table 4-15: Sharpe ratio comparing for optimal portfolio with risk-free asset
Regime 1
Regime 2
Risk aversion Non-regime Regime 1 decrease Non-regime Regime 2 decrease
2

0.260

0.289

-0.112

0.013

0.088

-5.597

0.260

0.289

-0.114

0.013

0.090

-5.716

10

0.260

0.290

-0.117

0.013

0.093

-5.925

15

0.260

0.291

-0.120

0.013

0.096

-6.127

20

0.260

0.292

-0.123

0.013

0.098

-6.328

50

0.260

0.297

-0.142

0.013

0.115

-7.560

Table 4-16: Sharpe ratio comparing for simulated one period returns over 10
years (with risk-free asset)
2

10

15

20

50

0.22711

0.2498

0.27249

0.29518

0.31786

0.34052

0.37446

0.24218

0.26032

0.2966

0.31473

0.33285

0.36

0.40514

0.24249

0.26182

0.30048

0.3198

0.33911

0.36804

0.41614

Table 4-17: Cumulated wealth at the end of 10 years investment simulation (with
risk-free asset)
2

1393.7

2376.8

2228.1

470.81

697.74

641.22

10

274.07

343

325.92

15

224.22

261.98

252.72

20

202.01

227.55

221.32

50

166.34

174.74

172.72

Table 4-18: Average optimal weight on GSCI over 10 years investment


simulation (with risk-free asset)
2

1.6258

1.702

1.6401

0.65031

0.68082

0.65603

10

0.32515

0.34041

0.32802

15

0.21677

0.22694

0.21868

20

0.16258

0.1702

0.16401

50

0.065031

0.068082

0.065603

Feb/04

Feb/02

Feb/00

Feb/98

Feb/96

Feb/94

Feb/92

Feb/90

Feb/88

Feb/86

Feb/84

Feb/82

Feb/80

Feb/78

Feb/76

Feb/74

Feb/72

Feb/70

Feb/04

Feb/02

Feb/00

Feb/98

Feb/96

Feb/94

Feb/92

Feb/90

Feb/88

Feb/86

Feb/84

Feb/82

Feb/80

Feb/78

Feb/76

Feb/74

Feb/72

Feb/70

Feb/04

Feb/02

Feb/00

Feb/98

Feb/96

Feb/94

Feb/92

Feb/90

Feb/88

Feb/86

Feb/84

Feb/82

Feb/80

Feb/78

Feb/76

Feb/74

Feb/72

Feb/70

113

Figure 4-1: Smooth probability of state 1 for commodity future indexes


1.2

0.8

0.6
GSCI

Non-energy

Energy

0.4

0.2

Figure 4-2: Smooth probability of state 1 for three asset classes


1.2

0.8

0.6
Stock

GSCI

Bond

0.4

0.2

Figure 4-3: Smooth probability of state 1 for three asset classes MRS

1.2

0.8

0.6

0.4

0.2

114

Figure 4-4: Mean-variance efficient frontier with three asset classes

Note: the inputs of asset return and risk are historical average in non-regime case, estimated results
for regime 1 in regim1 case, and estimated results for regime 2 in regime 2 case.

115

Figure 4-5: Optimal portfolio weight on three assets

Note: the inputs of asset return and risk are historical average in non-regime case.

116

Figure 4-5: Optimal portfolio weight on three assets (continued)

Note: the inputs of asset return and risk are estimates for regime 1 in regim1 case.

117

Figure 4-5: Optimal portfolio weight on three assets (continued)

Note: the inputs of asset return and risk are estimates for regime 2 in regim2 case.

118

Figure 4-6: risk and return comparing (no risk-free asset)

Panel A: in regime 1 situation


performance comparing at regime 1 with constraint
0.035

0.03

0.025

risk0
return0
risk1
return1

0.02

0.015

0.01

0.005

0
2

10

12

15

20

50

risk aversion

performance comparing at regime 1 without constraint


0.06

0.05

0.04
risk0
return0
risk1
return1

0.03

0.02

0.01

0
2

10

12

15

20

50

risk aversion

Note: risk0 and return0 are achieved by non-regime strategy; risk1 and return1 are achieved by
regime strategy.

119

Panel B: in regime 2 situation

performance comparing at regime 2 with constraint


0.035

0.03

0.025

risk0
return0
risk2
return2

0.02

0.015

0.01

0.005

0
2

10

12

15

20

50

risk aversion

performance comparing at regime 2 without constraint


0.04

0.035

0.03

0.025
risk0
return0
risk2
return2

0.02

0.015

0.01

0.005

0
2

10

12

15

20

50

risk aversion

Note: risk0 and return0 are achieved by non-regime strategy; risk2 and return2 are
achieved by regime strategy.

120

Figure 4-7: Mean and standard deviation of simulated one-period return (no
risk-free asset)

121

Figure 4-8: Mean and standard deviation of simulated one-period return (no
risk-free asset, continued)

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

122

Figure 4-9: Accumulated wealth over 10 years investment simulation (no risk-free
asset)
Panel A: 10 years accumulated return with weight constraint

123

Note: non-regime denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

124

Panel B: 10 years accumulated return without weight constraint

125

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

126

Figure 4-10: Optimal weight on GSCI over 10 years investment simulation (no
risk-free asset)
Panel A: GSCI weight with constraint

127

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

128

Panel B: GSCI weight without constraint

129

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

130

Figure 4-11: Optimal capital allocation with three risky assets and a risk-free asset

Panel A: Low risk aversion case

Panel B: high risk aversion case

131

Figure 4-12: Optimal weight on 4 asset classes

Note: non-regime denotes investment strategy ignoring regime, regime 1 denotes regime
investment strategy in regime 1, and regime 2 denotes regime investment strategy in regime 2.

132

Figure 4-13: Risk and return comparing (with risk-free asset)

Note: non-regime denotes the mean return achieved by investment strategy ignoring
regime, regime 1 denotes the mean return achieved in regime 1 following regime
investment strategy, and regime 2 denotes the mean return achieved in regime 2
following regime investment strategy.

133

Figure 4-14: Risk and return comparing (with risk-free asset, continued)

Note: non-regime denotes the risk achieved by investment strategy ignoring regime,
regime 1 denotes the risk achieved in regime 1 following regime investment strategy,
and regime 2 denotes the risk achieved in regime 2 following regime investment
strategy.

134

Figure 4-15: Mean and standard deviation of simulated one-period return (with
risk-free asset)

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

135

Figure 4-16: Accumulate wealth over 10 years investment simulation (with


risk-free asset)

136

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

137

Figure 4-17: Optimal weight on GSCI over the 10 years investment simulation
(with risk-free asset)

138

Note: strategy 0 denotes investment strategy ignoring regime, strategy 1 denotes first regime
investment strategy, and strategy 2 denotes second regime investment strategy.

139

Chapter 5 DYNAMIC PORTFOLIO SELECTION WITH COMMODITY FUTURES

Last chapter examined the issue of static portfolio selection with commodity
futures basing on a simple regime switching model. This chapter extends the simple
regime switching model into two general directions: economic inflation regime model
and business cycle regime model. An alternative objective utility function, the constant
relative risk aversion (CRRA) utility function, is used to investigate multi-period
strategic asset allocation problem.
5.1. General Regime-Switching Model

There is a large body of literature claiming that financial asset returns depend on
the underlying economic circumstances (Fama and French 1989, Ferson and Harvey
1991, Chen 1991, Bailey and Chan 1993, Hamilton and Lin 1996, Schaller and Norden
1997, and Campbell 2000). Chow, Jacquier, Kritzman and Lowry (1999) found that
commodities perform well when the general financial market climate is negative.
Edwards and Caglayan [2001] showed that commodity funds have higher returns during
bearish stock markets, along with a lower correlation. Jensen, Johnson and Mercer (2000
and 2002) found that during restrictive monetary environments, commodity futures
tended to show strong return/risk performance as a stand-alone asset and shared a
substantial weight in the optimal portfolio. All of these studies come to these conclusions
with a static model and without accounting for the dynamic dependence between
economic conditions. This dissertation is an initial effort to use the regime switching
model to examine the diversification benefits of commodity futures.
The general RS model can be written as:

140

yt = ( st ) + ( st ) zt 1 + yt
zt = c( st ) + ( st ) zt 1 + zt

. (5.1)

t = ( yt , zt ) ' ~ N (0, ( st ))
where yt is the N 1 asset return vector. zt is the M1 predictive instruments vector, st
is a latent state variable denoting the economic regime, rt-1 is the monthly US short rate,
and other characters denote model parameters. In this general model, the distribution of
asset returns depends on current economic regime st at time t and the previous
realization of the instrumental variables zt-1. The instruments zt themselves follow an
autoregressive process and the coefficients can vary with the regime. This general RS
model is a specific version of Kims (1994) state-space models with Markov switching.
Kim (1994) and Kim and Nelson (1999) present basic filtering and smoothing
algorithms for this model along with maximum likelihood estimation. The Kalman
filter is a key part of this linear dynamic system.
Suppose there are K regimes for our dynamic linear system (st=1, 2, K)
following a first-order Markov process with transition probabilities given by :
P11 K P1K

P= M O M
P

K 1 L PKK

where Pij = prob[ st = j / st 1 = i ] with

P
j =1

ij

(5.2)

= 1 for all i=1,2,,K. This system can

be complicated further by allowing time variation in the transition probabilities. Prior


researchers have discussed the motivations for allowing the transition probabilities to
depend on the instruments (Dielbold et al. 1994, Filardo 1994, Schaller and Norden
1997). It is possible to use a general function for the transition probabilities:
P t 1ij = prob[ st = j / st 1 = i, zt 1 ] = f t 1ij ( zt 1 )

(5.3)

Usually, a logistic functional form is chosen.


The general RS switching model can capture long-horizon predictability of

141

returns by instruments. Suppose we know the regime state and the values of instrumental
variables at time period t, then one-period ahead t+1 and multi-period ahead t+m
forecasts for expected values of mean returns and instrumental variables can be derived
using the iteration algorithm developed by Hamilton (1993). Meanwhile, the probability
of regimes can also be obtained.
The set of asset returns is yt = ( ytst , ytbd , ytcf )' where st denotes stock index
return, bd is the bonds return, and cf denotes commodity futures index. Two general
models are considered: model 1 using the T-bill rate as the instrumental variable and
model 2 using the change of Industrial Production Index (IPI) as the instrument. The first
model examines the inflation hedge benefit of commodity futures, since the inflation rate
supposedly exhibits opposite effects on the commodity futures and traditional assets. The
second model studies optimal portfolio selection with the business cycle identified by the
change in the Industrial Production Index (IPI). This model provides a detailed
examination on counter-cyclical movement between commodity futures and traditional
assets and its effect on the investment opportunity.

5.2 Portfolio Selection with T-bill Model


MLE estimation results

Let us first examine the optimal portfolio selection with the model using the
T-bill as the instrumental variable. It is assumed that there are two regimes which
dominate the asset return dynamics. The conditional means of three asset return depends
on the lagged T-bill rate with the dependence varying with the regime. Another channel
for the T-bill rate as an instrumental variable to predict the expected asset return is that
transitions probability between regimes is a function of the T-bill rate such as:

142

Pt 1 = prob( st = 1/ st 1 = 1, zt 1 ) =

exp(a1 + b1 zt 1 )
1 + exp(a1 + b1 zt 1 )

exp(a2 + b2 zt 1 )
Qt 1 = prob( st = 2 / st 1 = 2, zt 1 ) =
1 + exp(a2 + b2 zt 1 )

(5.4)

Meanwhile, the specification of equation 5.1 assumes that the T-bill rate follows
an autoregressive process, which has regime-dependent constant term and autoregressive
parameter. Ang and Bekaert (2002) provided evidence to support such a model, which
one regime captures normal times in which interest rates are highly persistent and not
variable, and another regime captures times of very variable, higher interest rates which
revert quickly to lower rates. The T-bill also is considered as a risk-free asset which
could be a significant part of the optimal portfolio when a bear market regime is
expected. Moreover, the T-bill rate is highly correlated with the inflation expectation and
significantly affected by monetary policy environment. The T-bill rate tends to be high in
high inflation expectation condition and a restrictive monetary environment and to be
low in low inflation expectation condition and an expansive monetary environment. With
the T-bill model, we can examine portfolio selection with three risky assets and one
risk-free asset under different inflation conditions and monetary environments.
Although the short rate has been proved to have a significant predictive power for
asset returns, our model estimation result shows that the dependence of asset return on
the lagged short rate is weak and we fail to reject the null hypothesis of =0 for all three
risky assets. Moreover, as the asset return is always estimated with little precision, a
simpler version of equation 1 is proposed, which restricts the risk asset return to be
constant and =0. A formal likelihood ratio test fails to reject this restricted model. Both
the estimation results of the unrestricted and restricted models are reported in Table 5-1.
Let us focus on the parameters of the restricted model. First, we look at the
constant mean return of three risky assets. Again, we identify that stocks and commodity
futures have almost same expected return over the sample period. The annualized
monthly returns for these two assets are approximately 14%. Bonds have 8% average

143

monthly returns. The volatility of three risky assets displays different degree
regime-dependence. Clearly, all three risky assets have high volatility in regime 2. The
annualized standard deviations implied for the the stock

market shock are 14% in

regime 1 versus 21% in regime 2; for shocks to commodity futures the relative numbers
are 18% and 19%, and for shocks to bonds the regime dependent volatilities are 7% and
14%. Volatility of stocks and bonds both show highly strong regime-dependence while
commodity futures display relatively consistent high volatility for both regimes.
Meanwhile, as Table 5-2 and Table 5-3 show, commodity futures negatively correlate
with stocks in the low volatility regime and positively relate with stocks in the high
volatility regime.
Our estimation results support the Ang and Bekaert (2002) conclusion about the
two regimes dynamics of the short term rate. The implied shocks to the T-bill rate
measured by standard deviations are 0.13% in first regime and 0.70% in second regime.
The drifts of the AR (1) process are 0.017 and 0.20, respectively, for first regime and
second regime; and the autoregressive coefficient value is 0.95 and 0.83, respectively.
The first regime captures normal times in which interest rates are highly persistent and
not variable, and the second regime captures times of very variable, higher interest rates
which revert quickly to lower rates.
A formal likelihood ratio test strongly rejects the hypothesis of the constant
regime transition probability. The estimates of coefficients a and b are statistically
significant and display a high regime-dependence. The negative value of estimated b in
regime 1 implies that as the short rate increases, the probability staying in the low
volatility regime will decrease, i.e. the regime will be more likely to switch from regime
1 to regime 2. Figure 5-2 clearly shows this negative relation. In regime 2, the estimated
coefficient b has positive value, which implies that, the probability staying in the high
volatility regime will increase with the short term rate.
To sum up, our restricted model identifies two regimes. The first regime is the

144

low volatility regime; both risky and risk-free assets have a low volatility. The T-bill rate
is relatively low and displays high persistence. The second regime is the high volatility
regime; both risky and risk-free assets have a high volatility. The T-bill rate is relatively
high and displays high variation. As short rate increases, the regime will have a greater
chance to either switch to the high volatility regime or stay in the high volatility regime.
Figure 5-1 presents the smoother probability of regime 1 at time period t over the sample
period.
Asset allocation

Now we examine the asset allocation issue with one risk-free asset and three
risky assets. It is assumed that the data generation process follows the restricted model
and the model estimation results is used to compute optimal asset weight without
constraints on short sale. Let us first investigate the one-period ahead optimal asset
allocation problem. Figure 5-3 presents the results of optimal asset weights in regime 1
and regime 2 with different short rate conditions and different risk aversion levels (Panel
A presenting commodity futures weights, panel B presenting stocks weights, panel C
presenting bonds weights, and panel D presenting the risk-free asset weights).
There are several visible features revealed in these figures. First, a linear
relationship is identified between the optimal asset weight and the short rate; the risky
asset weights linearly decrease with the short rate level while the risk-free asset weights
linearly increase with the short rate. Second, the lower the risk aversion levels, the
steeper optimal weights line, i.e. investors with higher risk tolerance ability are more
sensitive to the short term interest rate change. The optimal asset weight lines are almost
flat when risk aversion level is extremely high. Third, when the short rate is below the
mean return of bonds, about 0.8% a month, optimal asset allocation strategy borrows a
risk-free asset to invest on risky assets. About 30-40% capital is allocated to bonds and
the rest is distributed to stocks and commodity futures. When the short rate exceeds the
mean return of bonds, optimal asset allocation strategy shorts bonds and invests in a

145

risk-free asset, and the weights on stocks and commodity futures are decreased. Since
only the volatility is regime-dependent and the asset returns are supposed to be
independent of the regime, the optimal asset allocations strategy does not significantly
change with the regime variable. The only variation of optimal asset weights in regime 1
and regime 2 is the magnitude of absolute value.
One interesting result is that the optimal weights for stocks and commodity
futures are almost the same at most of time. The estimation results of restricted model
tell us that the mean returns of stocks and commodity futures are close to each other and
only the volatility is significantly different. Corresponding to that, the optimal weight on
stocks is slightly greater in regime 1 and slightly smaller in regime 2 compared to the
weight on commodity futures. This result is consistent with the results in chapter 3 but
inconsistent with Jensen et al (2000) conclusion that commodity futures have little or no
weight in the optimal portfolio under expansive monetary policy. Our conclusion is that
under expansive monetary policy, both the short rate and the inflation rate are low,
economic system has more probability to remain regime 1, stocks outperform
commodity futures and have a larger weight; Under restricted monetary policy, the short
rate and the inflation rate are high; economic system will be more likely to stay in regime
2, commodity futures outperform stocks and have a greater weight than stocks.
Given a specified risk aversion level, one period optimal asset weights are
dependent on the current period short rate level and current regime variable. Let us
examine the multi-period optimal portfolio allocation to find out whether there exists
intertemporal hedging demand. Assume that we know the current regime and the short
rate level and the true data generating process follows the restricted model, we plan to
derive monthly optimal asset weights for the next 3 years using the dynamic
programming method. Although there is no practical meaning for a multi-period optimal
asset allocation, it is worth investigating the optimal asset allocation problem under
different economic dynamics. It is impossible to efficiently forecast long-period asset
returns, especially in our case, which has to predict both the instrumental variable and

146

asset returns; it is also unnecessary to forecast a long-period asset returns series since we
can always use updated information to forecast the next period asset returns and derive
the one-period ahead optimal asset weights.
Four kinds of dynamics are identified for the multi-period asset allocations. In the
first dynamic, the short rate is low and the economic state is in regime 1, the short rate
displays strong consistency and less variation over the next 3 years. As the short rate
remains at a low over the periods, the chance of staying in regime 1 is very high. The
optimal asset weights are consistent with one-period results, just as panel A in Figure 5-4
presents. The optimal asset weights lines are almost flat for all three risky assets and
there is no noticeable difference between the multi-period asset allocation and the
one-period asset allocation, i.e. the intertemporal hedging demand is unobservable.
The second dynamic starts in regime 1 with a high short rate. As the probability
of switching is increased with the short rate level, it quickly switches to regime 2, a state
when the short rate displays high variation and quickly inverts to a low level. After the
high short rate reverts back to a lower level, economic state is more likely switching
back to regime 1 and the short rate is eventually stable at a low and consistent state.
Corresponding to the dynamic movement of short rate and regime switching, there are
four different phases of optimal asset allocations over the all periods: optimal asset
allocation in regime 2 with high interest rate, optimal asset allocation in regime 2 with
decreasing interest rate, optimal asset allocation in regime 1 with high interest rate, and
optimal asset allocation in regime 1 with consistent low interest rate. As panel B in
Figure 5-4 presents, the optimal asset allocations under these four phases are
substantially different and there are significant intertemporal hedging demands.
The third asset allocation dynamic is the multi-period portfolio selection starting
in regime 2 with a high short rate. As the short rate remains high, the chance of switching
is low and no regime switching occurs over all periods. The optimal asset weights are
consistent with the one-period results as panel C in Figure 5-4 presents. The optimal

147

asset weight lines are almost flat for all three risky assets and there is no noticeable
difference between multi-period asset allocation and one-period asset allocation, i.e. the
intertemporal hedging demand is unobservable.
The fourth dynamic is the multi-period portfolio selection starting in regime 2
with a low short rate. As the probability of switching to regime 1 is high when the short
rate level is low, it quickly switches to regime 1, a state the short rate displays low
variation and high consistence. Corresponding to this dynamic movement of short rate
and regime switching, there are two different phases of optimal asset allocations over the
period: optimal asset allocation in regime 2 with a low interest rate, and optimal asset
allocation in regime 1 with low interest rate. As panel D in Figure 5-4 has presented, the
optimal asset allocations under these two phases are substantially different and there is
significantly intertemporal hedging demand.
Conclusion

In conclusion, the short rate plays an important role in both one-period and
multi-period asset allocation problems. The short rate level determines two input
variables for optimal asset allocating: the transition probability and the risk-free asset
return. The optimal asset weights on risky assets are negatively related with the risk-free
rate. If the short rate stays consistently low in regime 1 or consistently high in regime 2,
multi-period asset allocation is not significantly different from the one-period asset
allocation, and no intertemporal hedging demand is observed. If the initial short rate is
high in regime 1, multi-period asset allocation experiences four different phases. If the
initial short rate is low in regime 2, multi-period asset allocation experiences two
different phases. In both cases, a significant intertemporal hedging demand is found.
5.3 Portfolio Selection with IPI model
MLE estimation results

148

Chapter 3 has discussed the counter-cycle movement among stocks, bonds, and
commodity futures. Commodity futures show strong performance during the business
cycle phases when stocks and the bond markets poorly perform. This section provides a
mathematic approach to test the counter-cycle movement. The change of IPI is used as an
instrumental variable in the regime switching model. Again, it is assume that there are
two regimes which dominate the asset return dynamics. The conditional means of three
asset return depends on the lagged value of the change in IPI with the dependence
varying with the regime. Another channel for the IPI growth rate as an instrumental
variable to predict the expected asset return is that the transitions probability between
regimes is a function of IPI rate such as:

Pt 1 = prob( st = 1/ st 1 = 1, zt 1 ) =

exp(a1 + b1 zt 1 )
1 + exp(a1 + b1 zt 1 )

exp(a2 + b2 zt 1 )
Qt 1 = prob( st = 2 / st 1 = 2, zt 1 ) =
1 + exp(a2 + b2 zt 1 )

(5.5)

Meanwhile, the specification of equation 5.1 assumes that the change of IPI rate
follows an autoregressive process, which has a regime-dependent constant term and an
autoregressive parameter. Considerable literature (such as Hamilton 1996) supports the
specification of autoregressive model with regime dependent coefficients for the change
in the IPI level. Table 5-4 reports the MLE results of IPI regime switching model. Two
null hypotheses are tested. The first hypothesis is =0 and constant mean returns for all
three risky assets; the second hypothesis is the constant transition probability, i.e. a=b=0.
A formal likelihood ratio test rejects both null hypotheses.
First, we look at the dynamic of asset returns. The one-period ahead of expected
asset return and the change rate in the IPI level can be expressed as equation (5.6) in
regime 1 and equation (5.7) in regime 2:

149

IPI t +1 = 0.06 + 0.50* IPI t


rbonds = 0.91 0.64IPI t
rstocks = 0.06 1.84* IPI t

(5.6)

rGSCI = 1.56 + 1.28* IPI t


IPI t +1 = 0.38 + 0.06* IPI t
rbonds = 0.79 0.38IPI t
rstocks = 1.27 + 0.36* IPI t

(5.7)

rGSCI = 0.73 + 0.51* IPI t

Bonds expected returns are negatively correlated with the change direction of the IPI
level in both regimes. The coefficients of IPI for bonds return are -0.64 in regime 1
and -0.38 in regime 2. If IPI level is increased by 1%, bonds return expects to decrease
0.64% in regime 1 and 0.38% in regime 2. The drifts of bonds return are positive in both
regimes, 0.91in regime 1 and 0.79 in regime 2. The bond market tends to obtain higher
returns in regime 1 relative to in regime 2. Stocks expected return is negatively
correlated with the change rate of IPI in regime 1 and positively correlated with the
change rate of IPI in regime 2. If IPI level is increased by 1%, stocks mean returns
expect to decrease 1.84% in regime 1 and increase 0.36% in regime 2. The drifts of
stocks return are negative in regime 1 and positive in regime 2, -0.06 and 1.27
respectively. Stocks are more likely to obtain negative and highly varied return in regime
1 and consistently positive return in regime 2. Commodity futures expected returns are
positively correlated with the change direction of IPI level in both regimes. When IPI
level is increased by 1%, mean returns on commodity futures are expected to be
improved by 1.28% in regime 1 and 0.51% in regime 2. The drifts of commodity futures
are positive in both regimes and higher in regime 1, 1.56 in regime 1 and 0.73 in regime
2. Commodity futures market tends to obtain higher return in regime 1 relative to in
regime 2.
The volatility of three risky assets also displays strong regime-dependence.

150

Clearly, all three risky assets have high volatility in regime 1. The annualized standard
deviations implied for the stock market shock are 21% in regime 1 versus 11% in regime
2; for shocks to commodity futures return the relative numbers are 24% and 14%, and for
shocks to the bond market the regime dependent volatilities are 10% and 7%. Meanwhile,
as Table 5-5 displays, commodity futures are negatively correlated with stocks in the low
volatility regime and positively related with stocks in the high volatility regime.
The estimation results identify two-regime dynamics of IPI change. The implied
shocks to the IPI growth rate measured by standard deviations are 2.6% in regime 1 and
1.7% in regime 2. The drifts are -0.06 and 0.38, respectively in first and second regime.
The autoregressive coefficient value is 0.50 and 0.06. In the first regime, the IPI level
tends to decrease and displays high variation. However, in the second regime, the IPI
level is more likely to increase and shows strong persistence.
We can recognize regime 1 as the state of later expansion and early recession, a
time period when the output level is over the sustainable growth level. In this economic
state, the IPI level either increases in a reducing rate or decreases in a growing rate. Both
the economic system and financial markets show high volatility and are sensitive to an
unexpected shock. Traditional finance markets, especially the stock market, suffer from
the restrictive policy environment as a way of restraining inflation and reducing the
excess demand. However, commodity markets show strong performance as supply is
constrained by capacity limits and commodities prices rise up. Regime 2 is the state of
early expansion and later recession, a time period when the output level is under the
sustainable growth level. In this economic state, the IPI level either increases at a
growing rate or decreases at a reducing rate. Both the economic system and financial
markets show consistent growth. Traditional finance markets, especially the stock market,
benefit from the accommodative policy environment and low raw material prices.
However, commodity markets show relatively poor performance as excess capacity is
present.

151

A formal likelihood ratio test strongly rejects the hypothesis of constant regime
transition probability. The estimates of coefficients a and b are statistically significant
and regime-dependent. The negative value of estimated b in both regimes implies that as
the IPI change rate increases, the probability of staying at a regime will decrease and the
regime will be more likely to switch from one regime to the other regime. Figure 5-5
clearly shows this negative relation
To sum up, the IPI regime switching model recognizes two regimes
corresponding to different business cycle phases. The first regime is a high volatility
regime, economic output level is over the sustainable growth level, traditional finance
assets have a poor performance, and commodity futures achieve good returns. The
second regime is low volatility regime, during which economic output level is under the
sustainable growth, traditional finance assets have relative better performance than
commodity futures. As the IPI growth rate increases, the regime will have a greater
chance to switch to the other regime. Figure 5-6 presents the smoother probability of
regime 1 at time period t over the sample periods.
Asset allocation

Let us investigate the asset allocation issue with only three risky assets following
the regime switching dynamic portfolio selection model again. It is assumed that the data
generation process follows the IPI model and the model estimation results is used to
compute optimal asset weights without constraints on short sale. Figure 5-7 presents the
one-period optimal asset weights in regime 1 (Panel A presenting stocks weights, panel B
presenting bonds weights, and panel C presenting commodity futures weights).
Regime 1 is a state when the economic system is running above the sustainable
growth level and there exists a capacity constraint. Bonds and stocks are negatively
related with the IPI growth rate and commodity futures are positively related with the IPI
rate. Corresponding to that, when the IPI growth rate is negative in regime 1; i.e. the IPI
level is decreasing, optimal asset allocation strategy shorts commodity futures and

152

invests most capital on the stock market and a small faction capital on the bond market.
The bigger the magnitude of negative IPI growth rate is, the greater commodity futures is
shorted. The investment on bonds keeps constant most of time. When the IPI growth rate
is positive, i.e. the IPI level is increasing, optimal asset allocation problem becomes more
complicated. When the growth rate is below 0.8%, the optimal strategy is shorting at
bond and stock market and invests all the capital in the commodity futures market. When
the growth rate is over that level, the economic state is more likely to switch to regime 2,
in which stocks are expected to achieve a return increasing with the IPI growth rate.
Then the optimal weights on stocks quickly rise up and turn to be positive. The optimal
strategy is to short bond and invests about 2/3rds of capital on commodity futures and the
other third in the stock market. For different risk aversion levels, the pattern of optimal
strategy is not changed, only the magnitude of optimal asset weights is reduced. The
greater risk aversion level, the less the magnitude of asset weights, i.e. investors with less
risk tolerance will do less shorting and as a result have less capital to invest.
Regime 2 is a state when the economic system is running under the sustainable
growth level and the demand can be satisfied by excess supply. Bond and stock markets
perform well in this regime. Commodity futures market is relatively poor performed.
Bond returns are still negatively related with the IPI growth rate. When the IPI growth
rate is negative, both commodity futures and stock returns will be significantly decreased
while bond returns are increased. Optimal asset allocation strategy is shorting
commodity futures as well as stocks and investing all capital on the bond market. When
the IPI growth rate is positive, i.e. the IPI level is increasing; the optimal strategy is
shorting at the bond market and investing capital on the stock and commodity futures
market. When the growth rate is below 0.8%, most capital is allocated to the stock
market. When the growth rate exceeds 0.8%, the economic state is more likely to switch
to regime 1, in which stocks are expected to achieve a return decreasing with IPI growth
rate. Then the optimal weights on stocks starts to decrease with the IPI growth rate but
remains positive, and most capital is allocated to commodity futures markets.

153

Given a specified risk aversion level, one period ahead optimal asset weights are
dependent on current period IPI growth rate and current regime variable. The importance
of the IPI growth rate in asset allocation is due to two reasons: first, the one-period ahead
expected asset returns are dependent on current IPI growth rate; second, the regime
transition probability is a function of IPI growth rate. These two reasons explain why the
optimal weights on stocks display irregular behavior. The multi-period optimal portfolio
allocation also reveals the importance of the IPI growth rate in the dynamic of optimal
asset allocation. Assuming that we know current regime and IPI growth rate and the true
data generating process follows the IPI regime switching model, we can derive monthly
optimal asset weights for next 3 years using the dynamic programming method. These
multi-period optimal assets weights can help us understand the asset allocation problem
under different economic dynamics.
Four kinds of dynamics for the multi-period asset allocations are identified in
regime 1 and Figure 5-8 resents the assets weights. In the first case, the IPI growth rate
remains negative and the economic state stays in regime 1. As the figure of panel A
displays, asset weights show strong consistency and less variation in the next 3 years. No
intertemporal hedging demand is observed. In the second case, the negative IPI growth
rate transfers to positive growth and the economic state switches from regime 1 to
regime 2. This process could be interpreted as business phase is changing from later
recession to early expansion. Panel B figure identifies three different phases of optimal
asset allocation. First phase will short commodity futures and invest on stocks and bonds;
second phase shorts stocks and invests on bonds and commodity futures; and third phase
shorts bonds and invests on stocks and commodity futures. In the third dynamic, the
initial IPI growth rate is positive and turns negative as time goes on, a process switching
from later expansion to early recession, the economic state remains in regime 1 and
optimal asset allocation has two phases as panel C figure shows: shorting stocks in the
first phase and commodity futures in the second phase. In the last dynamic, the IPI
growth rate is consistently positive and the economic state switches from regime 1 to

154

regime 2. Correspondingly, optimal asset allocation strategy is to short stocks and invests
on commodity futures and bonds. After the regime switching, optimal asset allocation
strategy is to short bonds and invests on commodity futures and stocks
In regime 2, three asset allocation dynamics are recognized. First, the IPI growth
rate remains positive and economic state is in early expansion, the optimal asset weights
do not vary over time and no intertemporal hedging is noticed. Second, the initial IPI
growth rate is negative and becomes positive later, switching from later recession to
early expansion. Optimal asset allocation has two phases: first, it shorts both stocks and
commodity futures and invests on bonds; second, it shorts bonds and invests on stocks
and commodity futures. Lastly, the initial IPI growth is very high and economic state
quickly transfers from regime 2 to regime 1. At the same time, the IPI growth rate is
decreasing. Correspondingly, optimal asset allocation experiences three phases: first, it
shorts bonds and invests most capital on stocks; second, it shorts bonds and invests most
capital on commodity futures; and third, it shorts stocks and invests equally on bonds and
commodity futures.
Conclusion

In conclusion, the IPI growth rate plays an important role in both one-period and
multi-period asset allocation problem. The IPI growth rate level determines transition
probability and three assets expected returns, two key input variables of optimal asset
allocating problem. Bonds are negatively correlated with IPI growth rate in both regimes.
When the IPI growth is negative, bonds performance are expected to increase; and vice
verse, when IPI growth is positive, bonds performance are expected to decrease.
Correspondingly, the optimal weights on bonds are negatively related with the IPI
growth rate. Commodity futures always positively correlate with IPI growth rate and the
optimal weights on commodity futures are positive related with the IPI growth rate.
Stocks negatively correlate with the IPI growth rate in regime 1 but positively relate with
the IPI growth rate in regime 2. When the IPI is growing in regime 2, stock performance

155

is expected to increase; however, when the IPI is growing in regime 1, stock performance
is expected to decrease. Correspondingly, when the IPI growth rate is positive, the
optimal weights on stocks increase in regime 2 and decrease in regime 1; when the IPI
growth rate is negative, the optimal weights on stocks increase in regime 1 and decrease
in regime 2. Whether there is an intertemporal hedge demand depends on the process of
economic dynamic.

156

Table 5-1: T-bill model parameters estimation


Unrestricted
Regime 1

Model

Restricted Model

Regime 2

Regime 1

Regime 2

0.0129

(0.0066)

0.1928

(0.2193 )

0.0171

(0.0054 )

0.7345

(0.0768)

1.0922

(0.8242 )

0.7227

(0.1105)

1.2088

(0.3076 )

1.5303

(0.2835 )

1.2052

(0.2198 )

1.2093

(1.4609 )

1.3977

(0.8050 )

1.2042

(0.2540)

0.9643

(0.6003 )

0.7655

(0.0838 )

0.9546

(0.0105 )

0.8319

(0.0293 )

-0.0029

(1.6661 )

0.4431

(0.3369 )

-0.0232

(0.7375 )

-0.0975

(0.2284 )

0.2377

(2.2079 )

-0.1184

(0.7969 )

Rrr

0.0450

(0.0145 )

0.1940

(0.2642 )

0.0405

(0.0018 )

0.2021

(0.0169 )

Rrb

-0.0570

(0.0869)

0.0615

(0.6581 )

-0.0487

(0.1185)

0.0630

(0.4583 )

Rrs

-0.4626

(0.6645)

-0.3393

(0.0600 )

-0.4655

(0.2048 )

-0.3441

(0.8043)

Rrc

0.5361

(1.6634 )

0.2522

(0.4740 )

0.5377

(0.2190)

0.2480

(0.7630 )

Rbb

2.0776

(1.2345 )

3.5965

(0.2851 )

2.0674

(0.0913 )

3.5936

(0.1426)

Rbs

0.9632

(2.2331 )

0.5818

(0.6504 )

0.9653

(0.2702)

0.5844

(0.8011 )

Rbc

-0.4172

(1.5650 )

-0.0193

(0.0977 )

-0.4159

(0.1413 )

-0.0218

(0.6332)

Rss

3.9462

(2.5770 )

5.9081

(4.2314 )

3.9329

(0.1826)

5.9152

(0.5209)

Rsc

-0.0780

(0.0022 )

0.5228

(1.1019 )

-0.0816

(0.3204 )

0.5259

(0.1622)

Rcc

5.0912

(0.0237 )

5.5932

(1.1064 )

5.0909

(0.4384)

5.5972

(0.5043)

2.9186

(0.1136 )

1.2768

(2.5201 )

2.9657

(0.5846 )

1.2578

(0.0546 )

-2.0081

(0.4874 )

1.1348

(1.6073 )

-1.9892

(1.5396)

1.1248

(0.1038)

Notes: standard errors are reported in parenthesis.

0.1957

(0.0954)

157

Table 5-2: Variance-covariance matrix estimation from T-bill model


Panel A: Variance-covariance matrix in regime 1
T-Bill

Bonds

Stocks

GSCI

T-Bill

0.0016

-0.0020

-0.0188

0.0218

Bonds

-0.0020

4.2766

2.0183

-0.8861

Stocks

-0.0188

2.0183

16.6163

-0.9727

GSCI

0.0218

-0.8861

-0.9727

26.3861

Panel B: Variance-covariance matrix in regime 2


T-Bill

Bonds

Stocks

GSCI

T-Bill

0.0408

0.0127

-0.0695

0.0501

Bonds

0.0127

12.9180

2.0785

-0.0628

Stocks

-0.0695

2.0785

35.4491

3.0125

GSCI

0.0501

-0.0628

3.0125

31.6672

Table 5-3: Correlation matrix estimation from T-bill model


Panel A: correlation matrix in regime 1
T-Bill

Bonds

Stocks

GSCI

T-Bill

1.0000

-0.0242

-0.1153

0.1061

Bonds

-0.0242

1.0000

0.2394

-0.0834

Stocks

-0.1153

0.2394

1.0000

-0.0465

GSCI

0.1061

-0.0834

-0.0465

1.0000

T-Bill

Bonds

Stocks

GSCI

T-Bill

1.0000

0.0175

-0.0578

0.0441

Bonds

0.0175

1.0000

0.0971

-0.0031

Stocks

-0.0578

0.0971

1.0000

0.0899

GSCI

0.0441

-0.0031

0.0899

1.0000

Panel B: correlation matrix in regime 2

158

Table 5-4: Parameters estimations of Industrial production index model


Regime 1

Regime 2

-0.0615

(0.0483 )

0.3842

(0.3439 )

0.9125

(0.0408 )

0.7682

(0.1681 )

-0.0550

(0.2687 )

1.2708

(0.5938 )

1.5638

(0.1724 )

0.7291

(0.2117 )

0.4861

(0.2729 )

0.0645

(0.1816 )

-0.6389

(0.2818 )

-0.3816

(0.0967 )

-1.8409

(0.5834 )

0.3596

(0.4972 )

1.2834

(0.3395 )

0.5144

(0.2250 )

Rrr

0.7592

(0.0541 )

0.5359

(0.5904 )

Rrb

-0.1731

(0.0232 )

-0.2658

(0.2768 )

Rrs

0.2978

(0.3085 )

-0.0279

(0.3582 )

Rrc

0.9066

(0.2391 )

0.3238

(0.0523 )

Rbb

2.9019

(0.3331 )

1.9852

(0.5837 )

Rbs

0.4403

(0.4069 )

1.3697

(0.2891 )

Rbc

-0.1625

(0.7172 )

-0.3080

(0.3718 )

Rss

5.8961

(0.4792 )

3.1346

(0.2096 )

Rsc

0.4153

(0.0504 )

-0.0958

(0.7247 )

Rcc

6.8757

(0.0166 )

4.1546

(0.4743 )

2.6436

(0.2484 )

3.3811

(0.7484 )

-2.4531

(0.1264 )

-1.1893

(0.1146 )

Notes: standard errors are reported in parenthesis.

159

Table 5-5: Variance-covariance matrix estimation for IPI model


Panel A: Variance-covariance matrix in regime 1
IPI
Bonds
IPI
0.5763
-0.1314
Bonds
-0.1314
8.4507
Stocks
0.2261
1.2261
GSCI
0.6883
-0.6283
Panel B: Variance-covariance matrix in regime 2
IPI
Bonds
IPI
0.2872
-0.1424
Bonds
-0.1424
4.0115
Stocks
-0.0149
2.7264
GSCI
0.1735
-0.6974
Panel C: correlation matrix in regime 1
IPI
Bonds
IPI
1.0000
-0.0595
Bonds
-0.0595
1.0000
Stocks
0.0503
0.0712
GSCI
0.1305
-0.0311
Panel D: correlation matrix in regime 2
IPI
Bonds
IPI
1.0000
-0.1327
Bonds
-0.1327
1.0000
Stocks
-0.0081
0.3979
GSCI
0.0775
-0.0833

Stocks
0.2261
1.2261
35.0463
2.6468

GSCI
0.6883
-0.6283
2.6468
48.2955

Stocks
-0.0149
2.7264
11.7023
-0.7312

GSCI
0.1735
-0.6974
-0.7312
17.4695

Stocks
0.0503
0.0712
1.0000
0.0643

GSCI
0.1305
-0.0311
0.0643
1.0000

Stocks
-0.0081
0.3979
1.0000
-0.0511

GSCI
0.0775
-0.0833
-0.0511
1.0000

Feb/04

Feb/02

Feb/00

Feb/98

Feb/96

Feb/94

Feb/92

Feb/90

Feb/88

Feb/86

Feb/84

Feb/82

Feb/80

Feb/78

Feb/76

Feb/74

Feb/72

Feb/70

160

Figure 5-1: Smoother probability of regime 1 at period t for T-bill model

1.2

0.8

0.6

0.4

0.2

161

Figure 5-2: Transition probability for T-bill model


Panel A: Transition probability for from regime 1 to regime 1

Panel B: Transition probability for from regime 2 to regime 2

162

Figure 5-3: One-period ahead optimal asset allocation (T-bill model)


Panel A: One-period ahead optimal weights on commodity futures

163

Panel B: One-period ahead optimal weights on stocks

164

Panel C: One-period ahead optimal weights on bonds

165

Panel D: One-period ahead optimal weights on T-bill

166

Figure 5-4: Multi-period ahead optimal asset allocation (T-bill model)


Panel A: asset weights dynamic 1

Panel B: asset weights dynamic 2

167

Panel C: asset weights dynamic 3

Panel D: asset weights dynamic 4

168

Figure 5-5: Transition probability for IPI model


Panel A: Transition probability from regime 1 to regime 1

Panel B: Transition probability from regime 2 to regime 2

Feb-04

Feb-02

Feb-00

Feb-98

Feb-96

Feb-94

Feb-92

Feb-90

Feb-88

Feb-86

Feb-84

Feb-82

Feb-80

Feb-78

Feb-76

Feb-74

Feb-72

Feb-70

169

Figure 5-6: Smoother probability of regime 1 at period t for IPI model

1.2

0.8

0.6

0.4

0.2

170

Figure 5-7: One-period ahead optimal asset allocation (IPI model)


Panel A: One-period ahead optimal weights on commodity futures

171

Panel B: One-period ahead optimal weights on stocks

172

Panel C: One-period ahead optimal weights on bonds

173

Figure 5-8: Multi-period ahead optimal asset allocation (IPI model)


Panel A: asset weights dynamic 1 in regime 1

Panel B: asset weights dynamic 2 in regime 1

174

Panel C: asset weights dynamic 3 in regime 1

Panel D: asset weights dynamic 4 in regime 1

175

Panel E: asset weights dynamic 1 in regime 2

Panel F: asset weights dynamic 2 in regime 2

176

Panel G: asset weights dynamic 3 in regime 2

177

Chapter 6 PORTFOLIO SELECTION WITH THE VALUE AT RISK

Prior chapters provide empirical evidence to support the diversification benefits


of commodity futures investment using some optimization models. However, as
Bacmann and Gawron (2004) have pointed out, any optimization framework relies on the
definition of expected returns, which is particularly prone to errors. As a consequence,
the choice of expected returns coming from a model or from a historical perspective
influences the optimal weights (Bacmann and Gawron [2004]). Moreover, optimization
methods are very sensitive to estimation errors and tend to exacerbate the impact of the
errors on the optimal weights (see Michaud [1998] for discussion about the
mean-variance model). In fact, the behavior of institutional investors is not well captured
by an optimization framework. When adding a new asset class into their portfolio, the
institutional investor is inclined to limit its investment portion to 5-10%. The
optimization framework usually suggests a much higher level of commodity futures
investment. In order to further investigate the portfolio role of commodity future, this
chapter provides a different approach using VaR as a risk measure. Instead of running an
optimization process, a number of portfolio sets are constructed by including different
levels of commodity futures into stocks and bonds portfolio. The performance of these
portfolios is examined to find out the risk reduction effect of commodity futures.

6.1 Why Use the Value at Risk

The mean-variance model measures investment risk based on the variation of the
return distribution. Balzer (1994) has shown that variance is only a suitable measure of
risk when the assets returns are normally distributed. Higher moments should be
considered when the distribution of asset returns is non-normal. In addition, investors are

178

mainly concerned about downside risk caused by adverse market movement, instead of
general volatility. Value at Risk (VaR), measuring the left (right) tail risk for holding a
long (short) market position at a given confidence level, can be obtained for general
distribution and is consistent with investors intuition. While VaR has become a popular
risk measurement tool in the investment world, no literature appears to examine the risk
reduction benefit of commodity futures based on the VaR risk measure.
Empirical evidence has shown that the returns of commodity futures and financial
assets are not normally distributed and exhibit unusual levels of skewness and kurtosis
(Hudson, Leuthold, and Sarassoro [1987], Hall, Brorsen, and Irwin [1989], Deaton and
Laroque [1992], Yang and Brorsen [1992], Myers [994], Vercammen [1995], Hilliard and
Reis [1999], Wei and Leuthold [2000], Roberts [2001], Kraus and Litzenberger [1976],
Duffie and Pan [1997], Timmermann [2000]). As a consequence, portfolio analysis based
solely on mean and variance may be leading to wrong conclusions and decisions. Amin
and Kat (2003), Bacmann and Scholz (2003) and Bacmann and Pache (2003) examined
hedge funds diversification effect and showed that while hedge funds combine well with
stocks and bonds in the mean-variance framework, this is no longer the case when
skewness is considered. Alexander and Baptista (2002) developed a mean-VaR model for
portfolio selection instead of the traditional mean-variance model and concluded that the
optimal portfolio solution with a mean-VaR model converges to the solution of a
mean-variance model only when the asset returns distribution is normal and the
confidence level for the VaR measure is very high. In 2003, they proposed a new
investment performance measure called reward-to-VaR ratio as a complement to the
traditional Sharpe ratio. The reward-to-VaR measure ranks portfolio performance
differently from the Sharpe ratio under non-normality. All of above discussed studies
encourage us to examine the risk reduction effects of commodity futures using VaR as
risk measurement.
6.2 Traditional VaR Analysis

179

Portfolio constructing

In order to examine the risk reduction benefit of commodity futures using VaR as
risk measure, a number of portfolio sets are constructed by adding different levels of
commodity futures into the initial portfolio with only stocks and bonds. The primary
object is to investigate whether adding commodity futures into the traditional asset
portfolio can significantly reduce the portfolio VaR under different economic conditions.
The different sets of portfolios are built by choosing the initial composition between
stocks and bonds. Eleven sets are defined, where the allocation to stocks (bonds) is
ranging from 0% (100%) to 100% (0%) with a step of 10%. In each of the sets, I add
different levels of commodity futures (0%, 5%, 10%, 15%, and 20% up to 100% with a
step of 5%). When commodity futures are added to the portfolio, the proportion of stocks
(or bonds) is kept constant in the traditional part of the portfolio. For example, if a set is
built with 20% stocks and 80% bonds, adding 20% commodity futures will decrease the
Weight of stocks to 20% * 80%= 16% and the Weight of bonds to 80% *80%=64%.
Portfolio VaR with regime switching

We learned from chapter 4 that the two-regime switching model fits well with the
three asset classes returns. The result of 10-year investment simulation has strongly
supports that portfolio strategy accounting regime switching effect. Let us assume that
the two-regime switching model is the true data generation of three asset classes return
and use this model to estimate the portfolios VaR. Given the value of state variable, the
conditional distribution of three asset returns at time period t is defined as a multi-normal
distribution with the estimated mean and covariance matrix. Then the 5% VaR can be
estimated

by

VaR = Wt0 *1.65 * h

and

1%

VaR

can

be

estimated

by

VaR = Wt0 * 2.33 * h . 13 As the regime switching model identifies two distinctive
market regimes for asset returns, we can examine the risk reduction effect of commodity

13

Chapter 2 has a detail instruction about the estimation of portfolio VaR

180

futures under both a normal and a bad market condition.


Figure 6-1 presents the results of both 1% VaR and 5% VaR for different
portfolios in a normal regime. A visible feature in this figure is that adding commodity
futures at first substantially reduces investment risk measured by the VaR. When the
weight on commodity futures increases to about 30%, the VaR level of the portfolio with
most bonds starts to increase with the weight of commodity futures; for the traditional
portfolios containing mostly stocks, portfolio VaR level goes up with commodity futures
when the weight on commodity futures increases to about 45%. Figure 6-2 provides both
results of the Reward-to-VaR for 1% VaR and 5% VaR in regime 1. In contrast to the
results of VaR, the reward-to-VaR is a concave function on the commodity futures weight.
Adding commodity futures can significantly increase the portfolios reward-to-VaR ratio
at first, as the commodity futures weight reaches 30% (for initial traditional portfolio
with most bonds) or 45%

(for initial traditional portfolio with most stocks), portfolios

reward-to-VaR ratio decreases when more commodity futures is included.


In practice, most of institutional investors only allocate about 5-10% investment
in commodity futures. Table 6-1 presents the increase of portfolios reward-to-VaR for
5%, 10%, and 15% commodity futures weight. The result shows that the benefit of
reward-to-VaR increase is more observable in the initial portfolio with most bonds. This
is could be due to the feature that commodity futures have comparable investment
performance to stocks in this market regime and adding commodity futures to traditional
portfolio with most bonds can increase expected return. On average, adding 5%
commodity futures will increase 5% portfolios reward-to-VaR, adding 10% commodity
futures will increase 10% portfolios reward-to-VaR, and adding 15% commodity futures
will increase 15% portfolios reward-to-VaR.
Figure 6-3 presents the results of both 1% VaR and 5% VaR for different
portfolios when market is in bad regime. Recall that regime 2 is a bad market regime,
in which stocks have poor investment performance characterized by negative return and

181

high price variance. Commodity futures achieve better performance and can provide
better diversification benefits for initial traditional portfolio containing most stocks.
Adding commodity futures at first can reduce the portfolio VaR and the risk reduction
benefit is very significant for the initial traditional portfolio with most stocks. For the
traditional portfolios containing mostly bonds, as the weight on commodity futures
increases to about 10%, portfolio VaR level starts to rise up. However, for the traditional
portfolios with most stocks, more commodity futures can be included. Portfolio VaR
increases with commodity futures only when the weight on commodity futures is up to
35% and adding commodity futures can reduce 25% of the portfolio VaR level.
Figure 6-4 provides the results for 1% VaR and 5% VaR in regime 2. For the
initial traditional portfolio with most bonds, adding commodity futures can only slightly
increase portfolio performance measured by the reward-to-VaR ratio. As the commodity
futures weight reaches 10%, portfolios reward-to-VaR starts to decrease with
commodity futures. For the initial traditional portfolio with most stocks, portfolios
reward-to-VaR ratio will always increase by including commodity futures. As the mean
return on stock is negative in this regime, adding commodity futures can turn the
negative portfolios mean return to the positive and significantly increase the portfolios
reward-to-VaR. The lower half of Table 6-1 reports the increase of reward-to-VaR ratio.
The results demonstrate that the increase of reward-to-VaR ratio is significantly observed
in the initial traditional portfolios with most stocks. The reason could be that commodity
futures have much better investment performance than stocks and at the same time shows
negative correlation with stocks. On average, adding 5% commodity futures will increase
portfolios reward-to-VaR by 95%, adding 10% commodity futures will increase
portfolios reward-to-VaR by 200%, and adding 15% commodity futures will increase
portfolios reward-to-VaR by 300%.
To sum up, portfolio VaR analysis with regime switching model again provides
evidence to support the diversification benefit of adding commodity futures into the
traditional portfolio. In most cases, adding 10% to 30% commodity futures can

182

significantly reduce portfolio VaR and provide a good downside risk protection. This
benefit is especially significant for the initial traditional portfolio with most stocks and
when market is in a bad regime.
Portfolio VaR with historical simulation

As discussed in chapter 2, the true data generating process for three asset returns
may not follow our regime switching assumption, and moreover, the result of our
portfolio VaR analysis with the regime switching model can not be guaranteed because
of the estimation error and model misspecification problem. An alternative VaR approach
called historical simulation (HS) is applied to re-estimate the portfolio VaR. The
historical simulation method needs few assumptions and past returns are used to predict
future returns. 14 In order to improve our estimation, 1000 times of the simulation are run
using the Bootstrap method developed by Efron (1982). As the choice of sample size has
a large impact on the value predicted by historical simulation, a variation of HS proposed
by Butler and Schachter (1996) is used to resolve this problem. The advantage of this
variation method is that a properly constructed kernel distribution provides a smooth
sampling distribution, which is not as sensitive to the sample length as HS.
Figure 6-5 presents the HR measurement of portfolios VaR at the 1% and 5%
levels. Figure 6-6 reports the portfolios reward-to-VaR. The results display a similar
pattern to the regime switching portfolio analysis. Adding commodity futures at first
significantly reduces investment risk measured by the VaR and improves the portfolio
performance by reward-to-VaR measure. For the traditional portfolios containing more
bonds, as the weight on commodity futures increases to about 20%, the portfolio VaR
level starts to increase and the portfolio reward-to-VaR measure decreases as more
commodity futures are included. For the traditional portfolios containing mostly stocks,
the portfolio VaR level goes up with commodity futures when the weight on
commodity futures increases to about 45%, while the portfolio reward-to-VaR measure
14

Mahoney (1996) provided a detail discussion on the historical simulation method to estimate VaR.

183

decreases. The portfolio with stocks to bonds ratio as 20:80 achieves the lowest
investment risk and best portfolio performance.
The percentage of portfolios VaR decrease is reported in Table 6-2 and the
percentage of portfolios Reward-to-VaR increase is presented in Table 6-3 for 5% to
30% commodity futures weight. Again, these results demonstrate that the
diversification benefit of commodity futures is different for the distinct traditional
portfolio. In the case of traditional portfolio with more bonds, adding commodity
futures can rapidly reduce portfolio risk and improve the performance. The optimal
weight of commodity futures is about 20-25%. While in the case of traditional portfolio
with more stocks, adding commodity futures only slowly reduces portfolio risk and
improves the performance, and the optimal weight of commodity futures is about
45-50%. On average, adding 5% commodity futures will reduce the portfolio VaR level
by about 6% and increase the portfolios reward-to-VaR by around 8%; adding 10%
commodity futures will reduce the portfolio VaR level by about 10% and increase the
portfolios reward-to-VaR by nearly 15%; adding 15% commodity futures will reduce
the portfolio VaR level by about 13% and increase the portfolios reward-to-VaR by
around 20%.
At this point, we arrive at the following conclusions:
1.

Adding commodity futures into traditional portfolio can reduce the


portfolio VaR and improve the portfolio performance measure by
Reward-to-VaR level.

2.

The diversification benefit of commodity futures is not identical for


heterogeneous traditional portfolios.

3.

For the traditional portfolio with most bonds, the optimal weight for
commodity futures which helps obtain best portfolio performance is about
20%; for the traditional portfolio with most stocks, the optimal level of
commodity futures is around 45%.

184

4.

Given that intuitional investors mostly allocate 5-10% capital on


commodity futures investment, our results suggest that the weight on
commodity futures should be increased, especially for equity investors.

6.3 Portfolio VaR with the Extreme Value Theory

Traditional risk measures including the VaR have difficulties in estimating the
extreme event risk existing in financial markets. However, extraordinary events such as
the oil crises in the early of 1970s, the stock market crash of October 1987, the
breakdown of the European Monetary System in September 1992, the turmoil in the
bond market in February 1994s and the emerging markets crisis in the early part of this
century are a central issue in finance, particularly in risk management. In order to
improve the estimation of such extreme events risk, Longin (1995) and Diebold et al.
(1998) suggest the use of the VaR measure based on the Extreme Value Theory (EVT).
VaR estimates are calculated from the lower extreme of a portfolio forecast
distribution therefore, accurate estimation of the lower tail of portfolio returns is of
primary importance in any VaR application. However, traditional methods are designed
to predict common volatilities, and therefore have poor tail properties. Even historical
simulation (HS) has less than desirable sampling properties out in the tails 15. The EVT
provides statistical tools to estimate the tails of the probability distributions. It
considers the distribution of extreme returns instead of the distribution of all returns.
The link between VaR and extreme value theory has been well established and the
literature on this field is plentiful. However, there are few papers to apply extreme
value VaR on commodity futures research. Longin (1999) discussed the optimal setting
of the margin level for both long and short positions on futures contract using the VaR
estimation based on extreme value theory. Odening and Hinrichs (2002) used the
extreme value theory to estimate VaR in order to quantify the market risk of feeder pigs,
finished hogs, and hog finishing margin. In this dissertation, I make use of this
15

Danielsson and de Vries (1997) had a through discussion about the VaR and extreme asset return issue.

185

powerful tool to study the benefit of commodity futures investment under extreme
market condition. Some basic concepts are briefly addressed below. A much more
comprehensive treatment can be found in Embrechts, Klppelberg, and Mikosch (1997)
and Kellezi and Gilli (2000).
The EVT plays a fundamental role in modeling the maxima or minimum of a
random variable. It tells us the limiting distribution of the extreme value of random
variables, just as the Central Limit theorem tells us the limiting distribution of the
sample mean of random variables. Let us consider an independent, identically
distributed random variable X with common distribution F, which may represent daily
losses or returns. Extreme value theory offers some interesting results about the
statistical distribution of extreme returns. In particular, the limiting distribution of
extreme returns observed over a long time period is largely independent of the
distribution of returns itself.
Introduction of the EVT

Generally there are two related ways of identifying extremes in real data: the
block maxima method and the peak over the threshold method. The first approach then
considers the maximum (or minimum) the variable takes in successive periods. These
selected observations constitute the extreme events, also called block maxima. From
the so-called extremal type theorem, the limit distribution for the block maxima is a
non-degenerate distribution function H, which belongs to one of the three standard
extreme value distributions:
x

Type I (Gumbel or thin-tailed class):

( x ) = e e , x

Type II (Frchet or heavy-tailed class):

x0
0,
>0
( x ) = x
e
,
x>0

Type III (Weibull or short-tailed class):

( x) , x 0
>0
( x ) = e
x>0
0,

(6.1)

186

Jenkinson (1955) suggested the following one-parameter representation of these


three standard distributions, with x such that1 + x > 0 . This generalization is known as
the generalized extreme value (GEV) distribution:

(1+ x )
H ( x ) = e e x
e

1/

if 0
if =0

. (6.2)

The second approach focuses on the realizations which exceed a given (high)
threshold; all exceed the threshold u and constitute extreme events. We consider an
(unknown) distribution function F of a random variable X. We are interested in
estimating the distribution function Fu of values of x above a certain threshold u and the
excesses are denoted by y. The distribution function Fu is called the conditional excess
distribution function (CEDF) and is formally defined as:
Fu ( y ) = P( X u y / X > u ), 0 y xF u where xF is the right endpoint of F.
Fu can be written in terms of F,
Fu ( y ) =

F (u + y ) F (u ) F ( x) F (u )
=
1 F (u )
1 F (u )

(6.3)

Pickands (1975) proved that generalized Pareto distribution (GPD) is the


limiting distribution for the CEDF Fu, Fu ( y ) = GPD , ( y ), for u . The GDP is
defined as:

1 (1 + y ) 1/
GPD , ( y ) =

1 e y /

if 0
if =0

(6.4)

where is the scaling parameter and is the shape parameter or tail index. The mean
of this distribution exists if < 1 and the variance if < 1/ 2 ; more generally, the
k th moment exists if < 1/ k . If x is defined as x=u+y, the GPD can be expressed as a

function of x:
GPD ,u , ( x) = 1 (1 + ( x u ) / ) 1/

for 0

(6.5)

187

since
F ( x) = (1 F (u )) Fu ( x) + F (u )

(6.6)

the a tail estimate of F(x) can be obtained by replacing Fu by the GPD and F(u) by the
estimate (n N u ) / n , where n is the total sample observations and Nu is the number of
observations above the threshold level u.
)
N
N
F ( x) = u (1 (1 + ( x u ) / ) 1/ ) + (1 u )
n
n

(6.7)

This is simplified to:


)
N
F ( x) = 1 u (1 + ( x u ) / ) 1/
n

(6.8)

Now the following is the explicit formula for extreme VaR estimate for a given
probability :

n
=u+
(1 ) 1

N u

VaR1

(6.9)

we can also give the estimate of another tail related risk measures, expected shortfall
(ES), which is defined as the expected size of a loss that exceed VaR .
ES1 = E[ X / X > VaR1 ]

(6.10)

let us rewrite the ES as:


ES1 = VaR1 + E[ X VaR1 / X > VaR1 ]

(6.11)

The second term on the right is the mean of the excess distribution FVaR1 ( y ) , which is
given as:
e(u ) = E ( X u / X > u ) =

+ u
1

for <1

(6.12)

Then the explicit formula for extreme VaR estimate for a given probability :

188

+ (VaR1 u )
ES1 = VaR1 +
1

VaR1 u
=
) +
1
1

(6.13)

The block maxima method is the traditional method used to analyze data with
seasonality such as hydrological data. In practice, the block maxima method suffers
from an important drawback. There is no consensus on the size of the blocks and the
estimation results are strongly influenced by the choice of sub-period. Moreover, this
method is very demanding in terms of data. Threshold methods use data more
efficiently and, for that reason, seem to have become the choice method in recent
applications.
Model the fat tail of three asset returns

Before starting a portfolio analysis using the extreme value theory, let us
examine the negative movement of three asset indexes by estimating their extreme
quantile. The data are the daily returns of three assets indexes from January 1977 to
November 2004: GSCI total return index, CRSP value weight stock index combining
NYSE/AMEX/NASDAQ, including dividends, and Ryan 10-year T-bond index. 16 6836
observations are included. Table 6-4 presents the descriptive statistics analysis. Stocks
have highest average daily return while bonds have lowest return. GSCIs expected
daily return is comparable with that of stocks but more volatile. The maximum loss of
GSCI is 18.43%, much larger than that of bonds and also greater than stocks. Stocks
and GSCI returns show negative skewness and bonds returns display positive skewness.
All of three assets indexes display fat tail behavior.
Maximum likelihood estimation

The MLE is used to estimate the parameters of the GPD, which fits the extreme
16

GSCI data comes from the DataStream international, stock index is provided by CRSP, and the bond index is
downloaded at Ryan web site.

189

negative return of three assets. Table 6-5 reports the MLE estimation results and the
computation of VaR, ES, and Reward-to-VaR for all three assets. The tail index
estimates for GSCI, stocks, and bonds are 0.1738, 0.1525, and 0.0262, respectively. The
tail index for bonds is not statistically significant. Commodity futures have the largest
tail index value and the highest daily VaR and ES at both 5% and 1% critical level, i.e.
the investment risk of commodity futures is higher than stocks and bonds. The 1% VaR
and ES for GSCI are 2.82% and 3.82%, which indicates that with probability 0.01, the
tomorrows loss of investing in commodity futures will exceed the value 2.82% and
that the corresponding expected loss, that is the average loss in situations where the
losses exceed 2.82%, is about 4%. These point estimates are completed with 95%
confidence intervals reported in the parentheses. Thus the expected loss will, in 95 out
of 100 cases, lie between 3.81% and 3.84%. The 1% VaR and ES for stocks are 2.62%
and 3.57%, both are slightly less than the corresponding value of GSCI. With
probability 0.01, the tomorrows loss of investing in stocks will exceed the value by
2.62% and that the corresponding expected loss is about 3.57%. With 95% confidence
intervals reported in the parentheses, the expected loss of stocks investment will, in 95
out of 100 cases, lie between 3.55% and 3.58%. The 1% VaR and ES for bonds are
1.39% and 1.77%. Both of them are considerably lower than the corresponding value of
GSCI and stocks.
As the choice of the threshold is critical in our EVT framework, the impact of
this choice is presented by creating a plot which shows how the estimate of a high
quantile in the tail of data based on GPD estimation varies with the number of
exceedances. Figure 6-7 presents the plots for both 99% and 95% quantile of negative
return of three assets .We do not find significant differences between the results of the
various thresholds when the number of exceedances is over 100. These results strongly
support my conclusion on stand-alone investment risk of three assets.
The Reward-to-VaR values assets investment performance by using VaR as the
risk measure. The last two rows in Table 6-5 report the Reward-to-VaR measure for

190

three asset classes using both 1% and 5% VaR. Stocks present the best investment
performance as they achieve the highest Reward-to-VaR ratio. The stand-alone
performance of commodity futures is poorer than that of stocks and bonds. The
Reward-to-VaR at 99% confidence level for GSCI is about 50% lower than that of
stocks. Commodity futures display much more extreme negative return than stocks and
bonds. The extreme negative movement of commodity futures markets can negatively
relate with downside movement of traditional financial markets. Adding commodity
futures might reduce the VaR of traditional portfolio and increase the overall portfolio
performance.
Portfolio VaR analysis

This part examines the benefit of commodity investment under extreme market
conditions. The extreme value VaR and the expected shortfall are estimated for of a
number of different portfolio sets to see whether adding commodity futures can
significantly reduce the extreme risk of the portfolio. The portfolios Reward-to-VaR
ratio is used study the portfolio performance. In order to improve the estimation results,
the bootstrap method has been applied here again and the mean value of 1000
bootstraps estimation results is used as the final estimates.
There are two technique problems with portfolio EVT empirical analysis. First,
how the EVT, which is designed for the estimation of univariate distribution, can be
adapted to analyze the tail behavior of a portfolio return composed by multiple asset.
Danielsson and De Vries (1996) describe two different approaches to apply EVT to a
portfolio, namely post fitting and presampling. Post fitting is similar to Historical
Simulation insofar as the returns of the different portfolio components are aggregated to
give a univariate series of portfolio return to which EVT can be applied. Correlations
need not be estimated explicitly, but are implicitly assumed to be constant. Presampling,
in contrast, is a multivariate approach. A tail estimation is carried out for each portfolio
component and samples are drawn from the fitted distributions. A vector of portfolio

191

returns is then calculated taking into account the sample covariance. Obviously post
fitting is computationally much simpler and is preferred used here.
The second problem is about that of transforming a short-term VaR to a
long-term VaR. As we are discussing strategy asset allocation problem, a long-term
VaR and ES might be more appropriate in our framework. Since the observation
frequency of the data is on a daily base, a short-term VaR (daily) needs to be
extrapolated to the long-term VaR (monthly). In practice the time-scaling is conducted
by means of the square-root-rule VaR(h) = h1/ 2VaR(1) . Let VaR(1) and VaR(h) denote
the one-period VaR and the h-period VaR, respectively, then according the
square-root-rule, VaR(h) is equal to the VaR(1) multiplied by the square root of h.
Danielsson and de Vries (2000) showed that square-root-rule can not apply to the EVT
framework and provided an alternative and more efficient method to convert one period
VaR into h-period VaR. They showed that VaR (h) = h1/ VaR(1) , where is the Hill
estimator of the tail distribution.
Figure 6-9 presents 1% VaR and 5% VaR for different portfolio, Figure 6-10
shows the ES results, and Figure 6-11 reports the Reward-to-VaR ratios. Instead of
showing any disparity, all of these results are very consistent with the traditional
portfolio VaR analysis. The only difference we can observe is that Reward-to-VaR ratio
displays irregular behavior and the Reward-to-VaR curve is not as smooth as we
observe in traditional VaR case. But overall, the pattern of Reward-to-VaR movement is
the same. VaR and ES curves are concave lines while Reward-to-VaR curves display
convex property. Adding commodity futures firstly significantly reduces investment
risk and improves portfolio performance. The rate of risk reduction and performance
improvement decreases with commodity futures weight. After the weight on
commodity future is over a threshold level, including more commodity futures will lead
to a VaR risk measure increase and overall portfolio performance decreases as well.
This threshold level is different for various traditional portfolios.

192

The optimal weight on commodity futures for the risk reduction increases with
the portion of stocks in the portfolio. It is about 20% for the traditional portfolios with
most bonds and about 45% for portfolio with most stocks. For traditional portfolio with
only stocks, allocating 55% to commodity futures will result in the lowest investment
risk. The optimal weight on commodity futures for Reward-to-VaR increase is identical
for all traditional portfolios. No matter the composition of traditional portfolio, 25-30%
investment on commodity futures is always doing the best to improve overall portfolio
performance.
To measure the magnitude of commodity futures benefit for risk reducing and
performance improvement, the percentage change of portfolio VaR and Reward-to-VaR
ratio are estimated. Only the 1% VaR results are reported. Table 6-6 shows the
percentage change of various traditional portfolios VaR when different levels of
commodity futures from 5% up to 50% are included, and Table 6-7 presents the results
for Reward-to-VaR ratio. The results again confirm our conclusion that optimal level of
commodity futures for VaR reduction is an increasing function of the stocks portion.
Allocating 15% to 20% capital to commodity futures can decrease 10-15% VaR for
traditional portfolio with most bonds. For traditional portfolio with more stocks than
bonds, the optimal commodity futures weight is 40%, which leads to a 20-25%
decrease in portfolio risk. For initial traditional portfolio with 90% or more stocks,
more than 50% investment in commodity futures can reduce the portfolios VaR by
40%. An allocation of 25-30% investment on commodity futures is the optimal strategy
for various traditional portfolios when we consider the Reward-to-VaR improvement.
For traditional portfolio with more bonds than stocks, 25-30% investment on
commodity futures can improve overall portfolios performance by about 30%; for
traditional portfolio with bonds equal to stocks, 25-30% investment on commodity
futures can improve overall portfolios performance by about 15%.
The difference of optimal weight on commodity futures between risk reducing
and portfolio performance could be due to the relative risk/return characteristics of the

193

three asset classes. As we have identified, the expected return of commodity futures is
much better than that of bonds but poorer than that of stocks. Commodity futures
volatility is higher than both stocks and bonds. When the traditional portfolio includes
mostly bonds, the main benefit of adding commodity futures could be expected return
improvement and as a result, the Reward-to-VaR ratio is significantly increased.
However, when traditional portfolio includes mostly stocks, the main benefit of adding
commodity futures could be risk reducing because of the negative correlation between
stocks and commodity futures market movement. Instead of increasing Reward-to-VaR
ratio, portfolio VaR is significantly decreased.
6.4 Conclusion

This chapter presents an alternative approach to study the risk reduction benefit
of commodity futures. Instead of using variation or volatility as investment risk
measure, a new risk measure, Value at Risk, which accounts the third and fourth
moments of asset returns, is applied here to measure the risk and overall performance
of different portfolios. Three different estimation of VaR is provided: parametric VaR
with regime switching model, historical simulation, and the EVT model. Although there
are some inconsistent implications on the portfolio performance for these three VaR
estimators, we can made following conclusions based on all three methods. First,
adding commodity futures into traditional portfolio can reduce the portfolio VaR and
improve the portfolio performance measure by Reward-to-VaR ratio. Second, the
diversification benefit of commodity futures is not identical for different initial
traditional portfolios, the more stocks are included in initial traditional portfolio, the
greater commodity futures can be added into the portfolio to reduce VaR and increase
Reward-to-VaR ratio. Third, allocating 5-10% capital on commodity futures investment,
as most intuitional investors have done, is not the optimal investment strategy,
especially for equity investors. Finally, extreme negative market movement for the
traditional asset does not occur synchronously with that for commodity futures; hence,
commodity futures do provide good downside protection.

194

Table 6-1: Percentage of Reward-to-VaR increase in different regime


Regime 1
GSCI

0-100

10-90

20-80

50-50

80-20

90-10

100-0

7.93%

7.15%

6.46%

4.89%

3.98%

3.78%

3.60%

10

15.87%

14.17%

12.68%

9.53%

7.88%

7.51%

7.20%

15

23.47%

20.69%

18.34%
13.70%
Regime 2

11.54%

11.09%

10.69%

GSCI

0-100

10-90

20-80

50-50

80-20

90-10

100-0

8.79%

10.47%

12.31%

22.94%

516.66%

56.27%

29.64%

10

12.84%

14.74%

18.66%

45.54%

1086.99%

119.11%

62.98%

15
11.27%
11.97%
17.70%
65.76%
1701.71% 188.16%
100.06%
Note: the first column is the weight on commodity futures and the first raw denotes the stocks and
bonds ratio, for example, 80-20 means the proportion of stocks and bonds is 8:2.

Table 6-2: Historical simulation result of portfolio VaR decrease


1% VaR
GSCI

0-100

10-90

20-80

50-50

80-20

90-10

100-0

-6.10%

-6.82%

-7.16%

-4.40%

-4.51%

-4.32%

-4.27%

10

-11.88%

-13.49%

-13.03%

-7.61%

-7.66%

-7.65%

-8.06%

15

-16.21%

-17.28%

-16.26%

-7.68%

-9.81%

-10.69%

-10.86%

20

-18.01%

-18.75%

-16.07%

-8.38%

-11.73%

-12.26%

-13.69%

25

-19.13%

-19.39%

-15.09%

-7.90%

-13.25%

-15.16%

-16.79%

30

-13.75%

-15.36%

-10.56%
-7.97%
5% VaR

-15.83%

-17.13%

-19.13%

0-100

10-90

20-80

80-20

90-10

100-0

GSCI

50-50

-7.02%

-6.68%

-3.73%

-6.69%

-7.58%

-5.76%

-6.37%

10

-10.34%

-8.84%

-8.55%

-10.76%

-12.90%

-11.77%

-12.19%

15

-12.65%

-11.69%

-11.09%

-12.84%

-18.17%

-17.74%

-17.57%

20

-12.63%

-12.88%

-10.77%

-13.59%

-22.73%

-21.84%

-22.51%

25

-9.59%

-7.80%

-2.89%

-11.57%

-26.56%

-25.54%

-25.87%

30
-4.91%
-0.49%
-4.42%
-8.54%
-27.74%
-28.80%
-29.10%
Note: the first column is the weight on commodity futures and the first raw denotes the stocks and
bonds ratio, for example, 80-20 means the proportion of stocks and bonds is 8:2.

195

Table 6-3: Historical simulation result of portfolio Reward-to-VaR increase


1% VaR
GSCI

0-100

10-90

20-80

50-50

80-20

90-10

100-0

10.91%

9.68%

9.47%

6.53%

5.21%

6.17%

6.43%

10

20.13%

21.07%

19.94%

12.16%

10.14%

9.82%

10.86%

15

29.65%

30.41%

26.92%

13.94%

14.19%

15.18%

14.53%

20

35.33%

34.15%

28.65%

16.34%

17.40%

17.68%

19.97%

25

40.79%

39.50%

30.34%

15.79%

20.63%

22.28%

25.00%

30

35.37%

34.72%

25.49%
18.31%
5% VaR

24.76%

26.63%

28.54%

GSCI

0-100

10-90

20-80

50-50

80-20

90-10

100-0

9.56%

9.98%

5.04%

9.19%

8.32%

7.63%

7.37%

10

16.78%

16.16%

13.27%

15.40%

16.23%

16.03%

16.62%

15

21.83%

20.92%

18.78%

20.53%

25.28%

25.44%

24.89%

20

25.79%

25.80%

20.82%

23.83%

33.04%

31.41%

33.19%

25

23.27%

20.83%

12.81%

22.27%

42.50%

39.78%

39.67%

30

20.19%

14.81%

6.42%

19.75%

45.80%

48.51%

45.92%

Note: the first column is the weight on commodity futures and the first raw denotes the stocks and
bonds ratio, for example, 80-20 means the proportion of stocks and bonds is 8:2.

Table 6-4: Descriptive statistics of three assets daily returns


Stocks

T-bonds

GSCI

Mean

0.054%

0.023%

0.040%

Standard Deviation

0.963%

0.511%

1.061%

Kurtosis

19.737

5.235

16.599

Skewness

-0.979

0.178

-0.888

Range

25.797%

8.731%

26.031%

Minimum

-17.135%

-3.800%

-18.431%

Maximum

8.662%

4.931%

7.600%

Count

6836

6836

6836

196

Table 6-5: Maximum estimation results of left tail on three risky asset returns
GSCI

Stocks

T-bonds

0.1738
(0.0525)

0.1525
(0.0510)

0.0262
(0.04571)

0.6032
(0.0413)

0.5999
(0.0402)

0.3572
(0.0221)

VaR0.05

1.6108
(1.6088

1.6128)

1.4553
(1.4532

1.4573)

0.7932
(0.7921

0.7943)

VaR0.01

2.8138
(2.8083

2.8192)

2.6181
(2.6131

2.6232)

1.3873
(1.3849

1.3896)

ES0.05

2.3869
(2.3826

2.3913)

2.2016
(2.1974

2.2058)

1.1635
(1.1618

1.1653)

ES0.01

3.8286
(3.8137

3.8435)

3.5724
(3.5589

3.5860)

1.7668
(1.7627

1.7710)

RTV0.05

0.0251
(0.0246

0.0256)

0.0375
(0.0370

0.0380)

0.0297
(0.0292

0.0302)

RTV0.01

0.0143
(0.0140

0.0146)

0.0205
(0.0202

0.0208)

0.0169
(0.0166

0.0172)

Note: for the estimates of

, , the value reported in the parentheses is estimates standard error;

for the estimate of VaR, ES, and Reward-to-VaR (RTV), the value reported in the parentheses is the
95% confidence interval.

Table 6-6: Percentage of portfolio EVT VaR decrease


5

10

15

20

30

40

50

0-100

-4.13%

-8.86%

-13.05%

-14.08%

-12.32%

-2.72%

9.61%

10-90

-4.45%

-8.94%

-11.85%

-13.61%

-10.34%

0.99%

13.79%

20-80

-3.55%

-7.56%

-10.83%

-12.96%

-8.38%

2.92%

16.25%

30-70

-4.92%

-7.71%

-12.06%

-13.51%

-8.01%

1.64%

15.05%

40-60

-3.63%

-6.91%

-10.71%

-11.53%

-7.62%

1.34%

13.04%

50-50

-3.94%

-8.06%

-11.05%

-12.56%

-12.24%

-4.85%

5.19%

60-40

-3.56%

-7.79%

-11.78%

-14.07%

-16.20%

-10.90%

-2.77%

70-30

-4.22%

-8.50%

-13.30%

-17.42%

-21.37%

-19.16%

-12.27%

80-20

-5.13%

-9.73%

-14.82%

-19.26%

-24.98%

-24.79%

-21.78%

90-10

-5.01%

-9.49%

-14.24%

-18.97%

-26.77%

-31.19%

-30.09%

100-0
-5.61%
-10.43%
-15.83%
-21.50%
-30.05%
-36.78%
-39.05%
Note: the first row is the weight on commodity futures and the first column denotes the stocks and
bonds ratio, for example, 60-40 means the proportion of stocks and bonds is 6:4.

197

Table 6-7: Percentage of portfolio EVT Reward-to-VaR increase


5

10

15

20

30

40

50

0-100

5.86%

14.21%

23.76%

28.66%

32.05%

25.43%

19.19%

10-90

3.73%

12.04%

17.23%

23.91%

20.93%

12.62%

3.31%

20-80

10.61%

15.61%

22.53%

24.30%

23.83%

12.96%

1.33%

30-70

8.21%

15.81%

16.35%

18.81%

14.66%

3.42%

-5.12%

40-60

6.30%

11.21%

14.74%

17.32%

15.21%

4.36%

-3.99%

50-50

9.71%

8.00%

15.88%

18.56%

14.93%

9.12%

-2.79%

60-40

6.55%

9.33%

13.81%

16.07%

16.54%

7.82%

0.51%

70-30

2.56%

5.64%

12.58%

13.44%

15.00%

13.24%

4.60%

80-20

1.43%

6.18%

7.19%

13.24%

18.31%

15.54%

8.38%

90-10

5.68%

7.01%

9.84%

16.29%

16.85%

20.59%

15.94%

100-0
6.24%
11.72%
12.84%
18.24%
23.90%
24.52%
18.46%
Note: the first row is the weight on commodity futures and the first column denotes the stocks and
bonds ratio, for example, 60-40 means the proportion of stocks and bonds is 6:4.

198

Figure 6-1: VaR in regime 1 for different portfolios

199

Figure 6-2: Reward-to-VaR in regime 1 for different portfolios

200

Figure 6-3: VaR in regime 2 for different portfolios

201

Figure 6-4: Reward-to-VaR in regime 2 for different portfolios

202

Figure 6-5: Historical simulation VaR for different portfolios

203

Figure 6-6: Historical simulation Reward-to-VaR for different portfolios

204

Figure 6-7: VaR estimates and the number of exceedances

205

Figure 6-8: VaR estimates and the number of exceedances (continued)

206

Figure 6-9: the EVT VaR for different portfolios

207

Figure 6-10: the EVT ES for different portfolios

208

Figure 6-11: the EVT Reward-to-VaR for different portfolios

209

Chapter 7 CONCLUSIONS AND IMPLICATIONS

This study was primarily conducted to investigate the portfolio diversification


benefit of commodity futures investment. For traditional institutional investors who
initially invest in stocks and bonds, allocating part of the portfolio to commodity futures
can significantly improve portfolio performance. The optimal portfolio weight on
commodity futures changes with market regime, economic condition, and the component
of initial traditional portfolio. This study develops several approaches to investigate the
issue of optimal allocation to commodity futures.
First, a general study is provided to investigate the stand-alone and portfolio
performance of commodity futures. The long historical data supports that commodity
futures can supply good investment return and the risk/return character is comparable
with that of stocks and bonds. On historical average, the total return from a commodity
futures index is comparable to stocks and outperforms bonds. The relative performance
between commodity futures and traditional financial assets varies with the business and
market conditions. In general, commodity futures returns are negatively correlated with
stocks and bonds and can supply efficient diversification benefits. When the stock and
bond market displays poor performance or significant downside movement, commodity
futures market usually presents a strong performance and provides good return.
The negative correlation between commodity futures and stocks and bonds comes
from the nature of commodity futures market and three fundamental differences between
them. First, commodity futures prices are expected to rise when expected or unexpected
inflation rate increases. Commodity futures investment can provide an efficient way to
hedge against inflation risk. Traditional financial markets (stock and bond) have been
shown to be vulnerable to inflation shocks. Second, commodity futures display
counter-cycle behavior with stocks and bonds. When the business cycle moves to the
stage of output growth over sustainable growth level, the stocks and the bond markets are

210

more likely to suffer from the price of raw materials rising and the increased interest rate.
However, commodity markets show strong performance as supply is constrained by
capacity limits. Third, commodity futures returns tend to have positive exposure to event
risk since the event shocks unexpectedly cut the supply of the commodity and move the
commodities prices up, causing a significant gain to long futures investors. These event
shocks are expected to be negatively related to the financial markets as their expected
returns will be reduced by the higher cost of raw material inputs.
Both commodity futures and traditional financial markets display structure break
behavior and their risk/return characters will change under different market and
economic conditions. In order to address this regime switching issue and its relevance
with optimal asset allocation, a Markov regime-switching model is introduced.
First, a simple two-regime switching model is examined. It is found that both the
mean and variance of commodity futures return are regime-dependent just like stocks
and bonds. Meanwhile, the regime switching movements of these three asset classes
display independence and uniqueness. In order to study the optimal asset allocation
under different market condition, it is assumed that an unobserved variable determines
the regime switching of the integrated market with commodity futures and traditional
financial asset. The estimation results show that there is a normal regime characterized
by relatively high return and low risk and a bad regime characterized by relatively low
return and high risk. These estimated results are used to examine the one-period optimal
portfolio allocation issue. It is found that if the regime switching model is the true data
generating process, an investment strategy ignoring the regime switching effect leads to
an inefficient asset allocation and a poorer investment performance.
In order to justify this argument, a 10-year out-sample investment simulation is
conducted and the results strongly support that the regime switching effect is important
to optimal asset allocation, especially in the bad market regime when stocks and bonds
are poorly performed. Commodity futures have been proved to be a significant part

211

optimal portfolio. Optimal allocation to commodity futures depends on investors risk


aversion: the higher risk tolerance, the more commodity futures should be included in the
portfolio. Since commodity futures are relatively stronger than stocks and bonds during
the bad market regime, they can provide a significant diversification benefit. The
investment simulation results also show that the regime-switching model can well predict
the break change in the market and sends signal to investors to increase exposure to
commodity futures.
Although the simple regime switching model can well address the structural
break behavior of asset returns and has been proved to be successful in portfolio
allocation strategy, it can not relate asset return dynamics and optimal asset allocation to
an economic variable. A general regime-switching model which extends the simple
model by using an economic variable as the instrument to predict asset return dynamics
can help us achieve this goal. Two different general regime switching models are
addressed.
The first general regime switching model uses the short rate as the instrumental
variable to predict the expected asset returns and regime switching probability. As
expected, the model shows that the short rate plays an important role in asset return
dynamics and asset allocation strategy. Although it fails to identify a prediction ability of
the short rate for three risky asset returns, the short rate does determine the regime
transition probability and the risk-free asset expected returns, which are the two key
input variables for asset allocating optimization. This model identifies two regimes for
asset return dynamics. The first regime is a low volatility regime, during which both the
risky and risk-free assets have low volatility. The second regime is a high volatility
regime, during which both the risky and risk-free assets have high volatility. As the short
rate increases, the regime will have a greater chance to either switch to the high volatility
regime or stay at the high volatility regime. The one-period ahead optimal asset weights
on the risky assets are negatively related with the short rate. This is not surprising since
the short rate determines the return of a risk-free asset. As the risk-free asset return

212

increases, investors will expected greater returns from the risky asset and have more
opportunities to switch investment to the risk-free asset if the expectation is not met.
However, the fraction of commodity futures investment in risky asset portfolio increases
with the short rate level.
Regarding the multi-period asset allocation problem, several different dynamics
are identified. If the short rate remains low in regime 1 or high in regime 2, multi-period
asset allocation is not significantly different from the one-period asset allocation. No
intertemporal hedging demand is observed. If the initial short rate is high in regime 1, the
multi-period asset allocation experiences four different phases. If the initial short rate is
low in regime 2, multi-period asset allocation experiences two different phases. In both
cases, there exists a significant intertemporal hedging demand.
The second general regime switching model is called the IPI model which uses
the monthly change of industrial production index as the instrumental variable to model
three risky asset returns. This model investigates the portfolio selection issue under
different business cycle phases. The estimation results of this model support that the IPI
growth rate determines the regime transition probability and three risky assets expected
returns, two key input variables of optimal asset allocating problem. The IPI regime
switching model recognizes two regimes corresponding to different business cycle
phases. The first regime is a high volatility regime, during which economic output level
is over the sustainable growth level, traditional finance assets obtain poor performance
but commodity futures achieve good return. The second regime is a low volatility regime,
during which economic output level is under the sustainable growth level, traditional
finance assets have a relative better performance than commodity futures. As the IPI
growth rate increases, the regime will have more chance to switch.
Commodity futures returns are always positively correlated with the IPI growth
rate. When the IPI growth rate increases, commodity futures performance is expected to
be strong; when the IPI growth rate decreases, commodity futures performance is

213

expected to be weak. Correspondingly, the optimal weights on commodity futures are


positively related with the IPI growth rate. Stock returns negatively correlate with the IPI
growth rate in regime 1 but positively relate to it in regime 2. When the IPI grows in
regime 2, stock performance is expected to increase; however, if the IPI grows in regime
1, stock performance is expected to decrease. Corresponding to that, if the IPI growth
rate is positive, the optimal weights on stocks increase in regime 2 and decrease in
regime 1. Bonds are negatively correlated with the IPI growth rate in both regimes.
When the IPI growth rate is negative, bond performance is expected to increase; and
when the IPI growth rate is positive, bond performance is expected to decrease.
Correspondingly, the optimal weights on bonds are positive when IPI growth is negative
and negative when IPI growth is positive. The multi-period optimal asset weights are
highly path-dependent, different IPI changing paths will lead to distinct multi-period
optimal asset allocations.
There are several limitations existing in this dynamic portfolio selection study.
First, the short rate is not equal to the inflation rate; it could be a good indicator of the
expected inflation rate, but it fails to reflect the unexpected inflation shocks. Similarly,
the IPI change rate is only one of the key indicators of business cycle and can not be
identical to business cycle movement. Second, as an instrumental variable is added into
regime switching model, the model becomes over-parameterized and estimation errors
must be considered. Lastly, the optimization process becomes very complicated and
highly sensitive to estimation errors, computation method, and other technical issues.
Correspondingly, the optimal portfolio selection could be also sensitive to these
problems. In fact, how to develop an efficient model to examine optimal portfolio
selection under different economic conditions is a challenging topic for further research.
Since the optimization methods to study portfolio allocation can lead to incorrect
conclusions, an alternative approach to study the portfolio construction problem is
followed. A set of portfolios is constructed by including commodity futures to the initial
portfolio with various stocks and bonds components and the VaR of these portfolios is

214

measured. The risk reduction benefit of commodity futures can be investigated by


comparing the portfolios VaR value. VaR accounts for the third and fourth moments of
asset returns and has been popular in the investment world. Three different estimations of
portfolios VaR are presented: the parametric VaR with a regime switching model, the
historical simulation VaR, and the EVT VaR. Although there are some inconsistent
implications on the portfolio performance from these three VaR estimators, the following
conclusions can be reached. First, adding commodity futures into traditional portfolio
can reduce the portfolio VaR and improve the portfolio performance measure by
Reward-to-VaR ratio. Second, the diversification benefit of commodity futures is not
identical for different initial traditional portfolios, the more stocks are included in initial
traditional portfolio, the greater commodity futures can be added into the portfolio to
reduce VaR and increase Reward-to-VaR ratio. Third, allocating 5-10% capital on
commodity futures investment, as most intuitional investors have done, is not the optimal
investment strategy, especially for equity investors. Finally, extreme negative market
movement for the traditional asset does not occur synchronously with that for
commodity futures; hence, commodity futures do provide good downside protection.
Although this dissertation has provided some good results about the
diversification benefit of commodity futures investment, the robustness of these results
has not been well tested. As the collateral return from the deposited T-bill is the
significant part of an unleveraged commodity futures index, the diversification benefit of
commodity futures might be overstated because the T-bill makes the return from
commodity futures more consistent. Energy products have more than 70% weight in
GSCI and play an important role in GSCI. The strong performance from GSCI could be
only caused by the rally of energy price rising and the role of other commodity products
in GSCI is not clearly examined. Energy products are the most actively traded
commodities; however the trading on other commodities in GSCI is not active. Energy
products and non-energy products might have different diversification performance.
Moreover, since the asset returns are highly varied under different time periods, using

215

data from different sample periods to examine the commodity futures investment might
lead to distinct results. Last but not least, this study assumes there is no tax and
transaction cost. Accounting these market factors might compromise the good results
developed in this dissertation.

216

REFERENCES

Abanomey, W.S. and Mathur, I. (2001), International Portfolios with Commodity


Futures and Currency Forward Contracts, Journal of Investing Fall, 61-68
Alexander, C. and Leigh, C. (1997), On the Covariance Matrices Used In VaR
Models, Journal of Derivatives Spring, 4, 50-62.
Alexander, G. J. and Baptista, A. M. (2002). Economic implication of using a
Mean-VaR model for portfolio selection: A comparison with mean-variance analysis,
Journal of Economic Dynamic and Control, July, 1159-1193.
Alexander, G. J. and Baptista, A. M. (2003). Portfolio performance evaluation using
value at risk, Journal of Portfolio Management Vol 29, Iss 4, 93-102.
Alizadeh, A. and Nomikos, N. (2004) A Markov regime switching approach for
hedging stock indices, Journal of Futures Markets, Vol 24, Iss 7 , 649 - 674
Anderson, R. (1985) Some determinants of the volatility of futures prices, Journal of
Futures Markets 5: 331348.
Ang, A. and Bekaert, G. (1999) International Asset Allocation with Time-Varying
Correlations. National Bureau of Economic Research. Working Paper 7056.
Ang, A. and Bekaert, G. (2002a) Regime Switches in Interest Rates, Journal of
Business and Economic Statistics, 20 (2), 163182.
Ang, A. and Bekaert, G. (2002b) Short Rate Nonlinearities and Regime Switches,
Journal of Economic Dynamics and Control, 26 (7-8), 12431274.
Ankrim, E. M., and Hensel, C. R. (1993). Commodities in Asset Allocation: A
Real-Asset Alternative to Real Estate? Financial Analysts Journal May/June 49(3)
20-29.
Anson, M.J.P. (1999), Maximizing Utility with Commodity Futures Diversification,
Journal of Portfolio Management, Summer, 86-94
Anson, M. J. P. (2002). Handbook of Alternative Assets, John Wiley & Sons Inc.
Bacmann, J. F. and Pache, S.2003, Optimal hedge fund style allocation under higher

217

moments, RMF Research Paper.


Bacmann, J.F. and Scholz, S. (2003), Alternative performance measures for hedge
funds, AIMA Journal (June).
Bacmann, J. F. and Gawron, G. 2004, Fat tail risk in portfolios of hedge funds and
traditional investments, RMF Research Paper.
Bailey W, Chan K C. (1993). Macroeconomic influences and the variability of the
commodity futures basis. Journal of Finance 45: 555573.
Bansal, Ravi and Hao Zhou, (2002) Term Structure of Interest Rates with Regime
Shifts, Journal of Finance, October, 221-247.
Barberis, N. (2000) Investing for the Long Run when Returns are Predictable.
Journal of Finance, 55(1), 225-264.
Becker, K.G. and Finnerty, J.E. (1994), Indexed Commodity Futures and the Risk of
Institutional Portfolios, OFOR Working Paper, February, no. 94-02.
Beder, T. S. (1995). VaR: seductive but dangerous. Financial Analysts Journal 51,
1224.
Bergman, M. and Hansson, J. (1997). Real Exchange Rates and Switching Regimes.
Working Paper 1999:4, Department of Economics, Lund University.
Best, M. J. and Grauer, R. R. (1991), on the sensitivity of Mean-variance-efficient
portfolios to changes in asset means: Some analytical and computational results,
Review of Financial Studies 4, 315-342.
Bjornson, B. and Carter. C. (1997) New Evidence on Agricultural Commodity Return
Performance under Time-Varying Risk. American Journal of Agricultural Economics,
August, 918-930.
Black, F. (1976). The pricing of commodity contracts. Journal of Financial
Economics, 3, 167179.
Black, F., and Scholes, M. (1973). The pricing of options and corporate liabilities,
Journal of Political Economy, 81, 637659.
Boehlje, M. D. and Lins, D. A. (1998): Risks and Risk Management in an
Industrialized Agriculture, Agricultural Finance Review, 58: 1-16.

218

Bollerslev, T. (1986). Generalized autoregressive conditional heteroscedasticity.


Journal of Econometrics, 31, 307327.
Bollerslev T. (1987). A conditionally heteroscedastic time series model for securities
prices and rates of returns data, Review of Economics and Statistics 59: 542547.
Bollerslev T, Chou RY, Kroner KF. (1992). ARCH modelling in finance. Journal of
Econometrics 52: 559.
Bollerslev T, Wooldridge JM. (1992). Quasi-maximum likelihood estimation and
inference in dynamic models with time-varying covariances. Econometric Reviews 11:
143172.
Brandt, M. (1999) Estimating Portfolio and Consumption Choice: A Conditional Euler
Equations Approach. Journal of Finance, 64(5), 1609-1645.
Brennan, M. J. (1958). The supply of storage, The American Economic Review, 48,
5072.
Brennan, M. J., Schwartz, E. S. and Lagnado, R. (1997) Strategic Asset Allocation.
Journal of
Economic Dynamics and Control, 21, 1377-1403.
Brorsen, B. W., and Irwin, S. T. (1987) Futures funds and price volatility, The Review
of Futures Markets, 6: 118-135.
Brooks, C., Clare, A. D. and Persand, G. (2002) A note on estimating market-based
minimum capital risk requirements: A multivariate GARCH approach, The
Manchester School Vol 70 No. 5: 666-681
Bodie, Z. (1983), Commodity Futures as a Hedge Against Inflation, Journal of
Portfolio Management, Spring: 12-17
Cai J. (1994). A Markov model of regime-switching ARCH, Journal of Business and
Economic Statistics 12: 309316.
Campbell, J. Y. (1998). Consumption-Based Asset Pricing , Harvard Institute
Research Working Paper No. 1974
Campbell, J. Y. and Viceira L.M. (1999) Consumption and Portfolio Decisions When
Expected Returns are Time Varying. The Quarterly Journal of Economics, May
433-495.

219

Campbell, J. Y. and Viceira L.M. (2000) Consumption and Portfolio Decisions When
Expected
Returns
are
Time
Varying:
Erratum.
on
http://kuznets.fas.harvard.edu/~campbell/papers.html.
Campbell, J. Y. and Viceira L. M. (2002) Strategic Asset Allocation: Portfolio Choice
for Long-Term Investors. Oxford University Press Inc., New York.
Campbell, S., and Li, C. (2001), Option Pricing with Unobserved and
Regime-Switching Volatility, Manuscript, University of Pennsylvania.
Cechetti, S., Lam, P. and Mark, N. (1990), Mean Reversion in Equilibrium Asset
Prices, American Economic Review, 80, 398-418.
Chen, N.-F. (1991), Financial Investment Opportunities and the Macroeconomy,
Journal of Finance, vol. 46, p. 529-554
Chow, Y-F. (1998) Regime Switching and Co-integration tests of the efficiency of
futures, The Journal of Futures Markets 18: 871-901
Chow, G., Jacquier, E., Kritzman, M. and Lowry, K (1999), Optimal Portfolios in
Good Times and Bad, Financial Analysts Journal, May/June: 65 - 73
Christoffersen, P. F., and Diebold, F.X. (2000), How Relevant is Volatility
Forecasting for Financial Risk Management? Review of Economics and Statistics 82:
1-11.
Dahlquist, M. and Gray, S. F. (2000), Regime-Switching and Interest Rates in the
European
Monetary System, Journal of International Economics, 50, 399-419.
Danielsson, J., and C.G. de Vries (1997), Beyond the Sample: Extreme Quantile and
Probability Estimation, Working Paper, London School of Economics, University of
Iceland.
Danielsson, J., L. de Haan, L. Peng, and C.G. de Vries (2001), Using a Bootstrap
Method to Choose the Sample Fraction in Tail Index Estimation, Journal of
Multivariate Analysis 76: 226-248.
Dave, R. D. and Stahl, G. (1997). On the accuracy of VaR estimates based on the
VarianceCovariance approach, Working paper, Olshen & Associates.
Deaton, A., & Laroque, G. (1992). On the behavior of commodity prices. Review of
Economic Studies, 59, 123.

220

Deaton, A., & Laroque, G. (1996). Competitive storage and commodity price
dynamics. Journal of Political Economy, 104, 896923.
Dewachter, H. (2001). Can Markov Switching Models Replicate Chartist Profits in the
Foreign Exchange Market? Journal of International Money and Finance, 20(1):2541.
Diebold F.X, Lee J.H., Weinbach G. (1994), Regime Switching with Time-Varying
Transition Probabilities, in C. Hargreaves (ed.) Nonstationary Time-Series analysis
and Cointegration, 283-302, New York: Oxford University Press (Reprinted in F.X.
Diebold and G.D. Rudebusch Business Cycles: Durations, Dynamics, and Forecasting
(1999), Princeton: Princeton University Press).
Diebold, F.X. and Rudebusch, G.D. (1996), Measuring Business Cycles: A Modern
Perspective, Review of Economics and Statistics, 78, 67-77.
Diebold, F.X., Hickman, A., Inoue, A., and Schuermann, T., (1997), Converting
1-Day Volatility to h-Day Volatility: Scaling by Root-h is Worse than You Think,.
Wharton Financial Institutions Center, Working Paper 97-34. Published in condensed
form as "Scale Models", Risk, 11, 104-107 (1998).
Diebold, F.X., Schuerman, T., and Stroughair, J.D., (1998), Pitfalls and Opportunities
in the Use of Extreme Value Theory in Risk Management, Working Paper 98-10, The
Wharton School, University of Pennsylvania.
Dowd, K., (1998), Beyond Value at Risk: the New Science of Risk Management,
Chichester and New York: John Wiley & Sons,
Dowd, K. (1999), A Value at Risk approach to risk-return analysis, Journal of
Portfolio Management, Summer, 25, 4: 60-67
Duffie D, and Gray S. (1995). Volatility in energy prices. In Managing Energy Price
Risk. Risk Publications: London.
Duffie, D. and Pan, J., (1997). An Overview of Value at Risk. The Journal of
Derivatives 4 (3), 7-49.
Edwin J. E., Martin J. G., (2003), Modern portfolio theory and investment analysis New
York : J. Wiley & Sons
Edwards, F. and Ma, C., (1988), Commodity Pool Performance: Is the Information
Contained in Pool Prospectuses Useful? The Journal of Futures Markets, Vol. 8, No.
5.

221

Edwards, F., and Park, J., (1996), "Do Managed Futures Make Good Investments? The
Journal of Futures Markets, Vol. 16, No. 5.
Edwards, F.R. and Caglayan, M.O., (2001), Hedge Fund and Commodity Investments
in Bull and Bear Markets, Journal of Portfolio Management, Summer: 97-108
Efron, B. (1982). The Jackknife, the bootstrap, and other resampling plans,
Philadelphia, PA, Society for Industrial and Applied Mathematics.
Efron, B. and Tibshirani, R. (1993). An Introduction to the Bootstrap, London,
Chapman and Hall.
Elton, Edwin J., Martin J. Gruber, and Joel Rentzler. (1987) Professionally Managed,
Publicly Traded Commodity Funds. Journal of Business, 60, 175-199.
Embrechts, P., Klppelberg, C., and Mikosch, T., (1997), Modeling Extremal Events
for Insurance and Finance, Springer, Berlin.
Engel, C. and Hamilton, J. D. (1990). Long Swings in the Dollar: Are They in the
Data and Do Markets Know It? American Economic Review, 80:689713.
Engel, C. (1994). Can the Markov Switching Model Forecast Exchange Rates?
Journal of International Economics, 36:151165.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticty with estimates of the
variance of U.K. inflation. Econometrica, 50, 9871008.
Engle R. F., Bollerslev T. (1986). Modelling the persistence of conditional variances.
Econometric Review 5: 120.
Engle, R. F. and Manganelli, S. (2000). CAViaR: conditional autoregressive value at
risk by regression quantile, NBER. paper 0841.
Fama E. F., French K. R. 1987a. Permanent and transitory components of stock
prices. Journal of Political Economy 96: 246273.
Fama, E. F., & French, K. R. (1987b). Commodity futures prices: Some evidence on
forecast power, premiums, and the theory of storage. Journal of Business, 60, 5573.
Fama E. F., French K. R. (1988a), Business cycles and the behavior of metals prices,
Journal of Finance, 43: 10751093.

222

Fama, E. F. and K.R. French (1988b), "Dividend Yields and Expected Stock Returns,"
Journal of Financial Economics, 22, 3-25.
Fama E. F. and French K. R (1989). Business conditions and expected returns on
stocks and bonds, Journal of Financial Economics, 25: 23-50.
Fan, J. Q. and Gu, J. (2003), Semiparametric estimation of Value at Risk
Econometrics Journal, 6: 261290.

Ferson, W. E., and Harvey, C. R., (1991), The variation of economic risk premiums,
Journal of Political Economy 99: 285-315.
Ferson, W. E., and Harvey, C. R., (1993), The risk and predictability of international
equity returns, Review of Financial Studies 6, 527--566.
Filardo, A. J., (1994), Business-Cycle Phases and Their Transitional Dynamics.
Journal of Business and Economic Statistics, 12, 3: 299-308.
Fong, W. M. and See, K. H. (2001) Modeling conditional volatility of commodity
index futures as a regime switching process, Journal of Applied Econometrics, 16:
133163.
Fong, W. M. and See, K. H. (2003) Basis variations and regime shifts in the oil futures
market The European Journal of Finance 9: 499513
French K. R., (1986), Detecting spot price forecasts in futures prices. Journal of
Business 59: 3954.
French K. R, Schwert G. W., and Stambaugh, R. F. (1987). Expected stock returns and
volatility. Journal of Financial Economics 19: 330.

Froot, K. A. (1995). Hedging portfolios with real assets, Journal of Portfolio


Management 21(4): 60-77.
Georgiev, G. (2001). Benefits of Commodity Investment, Journal of Alternative
Investments, Summer, 40-48
Gibson, R. (1999). The Rewards of Multiple Asset Class Investing, Journal of
Financial Planning (March), 50-59
Gilli, M., and Kellezi, E. (2000), Portfolio selection, tail risk and heuristic

223

optimization, working paper, University of Geneva


Giot, P. (2003): The information content of implied volatility in agricultural
commodity markets, The Journal of Futures Markets, 23, 5: 441-454.
Giot, P., and Laurent, S., (2002): Market risk in commodity markets: A VaR
approach, working paper University of Namur
Gloy, B.A., and Baker, T.G. (2001): A Comparison of Criteria for Evaluating Risk
Management Strategies, Agricultural Finance Review 61: 36-56.
Gorton, G. and Rouwenhorst, K. G. (2004). Facts and Fantasies about commodity
Futures, NBER Working Paper, No.10595
Graflund, A. and Nilsson, B. (2002). Dynamic Portfolio Selection: The Relevance of
Switching Regimes and Investment Horizon working paper, Lund University
Granger, C.W. J. and Terasvirta, T., (1993), Modeling Nonlinear Economic
Relationships. Oxford University Press.
Gray S. F. (1995). Volatility in energy prices. In Managing Energy Risk, Risk
Publication: London; Chapter 2.
Gray S. F. (1996a). Modelling the conditional distribution of interest rates as a
regime-switching process. Journal of Financial Economics 42: 2762.
Gray S. F. (1996b). An analysis of conditional regime-switching models. Working
paper, Duke University.
Greene, W.H. (2000): Econometric Analysis, Fourth Edition. Prentice-Hall, New
Jersey.
Greer, R. J. (1978), Conservative Commodities: A Key Inflation Hedge. Journal of
Portfolio Management, Summer, pp. 26-29.
Hall, J. A., Brorsen, B. W., and Irwin, S. H. (1989), The Distribution of Futures Prices:
A test of the stable distribution and mixture of normal hypothesis, Journal of
Financial and Quantitative Analysis, 24: 105-116
Hamilton, D. J. (1989), A New Approach to the Economic Analysis of Nonstationary
Time Series and the Business Cycle. Econometrica, 57(2), 357-384.
Hamilton, D. J. (1990) Analysis of Time Series Subject to Changes in Regime.

224

Journal of Econometrics, 45, 39-70


Hamilton, D. J. (1993) Specification Testing in Markov-Switching Time Series
Models, Mimeographed, University of California, San Diego.
Hamilton, D. J. (1994), Time Series Analysis, Princeton University Press.
Hamilton, D. J., and Susmel, R., (1994). Autoregressive conditional heteroscedasticity
and changes in regime. Journal of Econometrics 64: 307333.
Hamilton, D. J. (1996a), Specification Testing in Markov -Switching Models,
Journal of Econometrics, 70, 127-157.
Hamilton, D. J., and Lin, G., (1996a), Stock Market Volatility and the Business
Cycle, Journal of Applied Econometrics, 11, 573-593.
Hammersley J.M., and Handscomb D.C., (1964), Monte Carlo Methods. New York:
John Wiley and Sons Inc.
Hansen B. 1992. The likelihood ratio test under nonstandard conditions: testing the
Markov switching model of GNP. Journal of Applied Econometrics 7: 6182.
Hansen, B. (1993) Inference When a Nuisance Parameter is Not Identified under the
Null Hypothesis, Mimeographed, University of Rochester.
Hansson, B. and Persson, M., (2000), Time Diversification and Estimation Risk.
Financial Analysts Journal 56(5), 55-62.
Harlow, W. V. (1991): Asset Allocation in a Downside-Risk Framework. Financial
Analysts Journal 47: 28-40.
Hartzmark, M. L. (1991), Luck versus forecast ability: determinants of trader
performance in futures markets, Journal of Business, 64, 1: 49-74
Harvey, A. C. (1987) Applications of the Kalman Filter in Econometrics, in: T.F.
Bewley, ed., Advances in Econometrics, Fifth World Congress, Volume I. Cambridge,
England: Cambridge University Press.
Harvey, A. C. (1989) Forecasting, Structural Time Series Models and the Kalman
Filter. Cambridge, England: Cambridge University Press.
Hendricks, D. (1996). Evaluation of value-at-risk models using historical data.
Federal Reserve Bank of New York Economic Policy Review, 3969.

225

Hilliard, J. E., & Reis, J. A. (1999). Jump processes in commodity futures prices and
options pricing. American Journal of Agricultural Economics, 81,273286.
Hilary, T. (2001), Taking Full Advantage of the Statistical Properties of Commodity
Investments. The Journal of Alternative Investments, Summer, pp. 63-66.
Hopper, G.P. (1996): Value-at-Risk: A New Methodology for Measuring Portfolio
Risk, Business Review Federal Reserve Bank of Philadelphia, 19-31.
Huang, C. F., and Litzenberger, R. H., (1988), Foundations for Financial economics.
Englewood Cliffs, NJ: Prentice Hall.
Huberman, G. and Kandel, S. (1987). Mean-variance spanning, Journal of Finance
42, 4: 873-888
Hudson, M. A., Leuthold, R. M., and Sarassoro, G. F. (1987). Commodity futures
price changes: Recent evidence for wheat, soybeans, and live cattle. Journal of
Futures Markets, 7, 287301.
Humphreys, Brett and David Shimko, (1997). Commodity Risk Management and the
Corporate Treasury. Financial Risk and the Corporate Treasury, Risk Publications, p.
115.
Irwin, S. H. and Landa, D. (1987). Real estate, futures, and gold as portfolio assets.
The journal of portfolio management, Fall, 29-34
Irwin, S. H., Krukemyer, T. R., and Zulauf, C. R., (1993), Investment Performance of
Public Commodity Pools: 1979-1990. The Journal of Futures markets, 13, 799-820.
Irwin, S. H., Zulauf, C. R., and Ward., B. (1994), The Predictability of Managed
Futures Returns. The Journal of Derivatives, winter, 20-27.
Irwin, S. H. and Yoshimaru, S. (1999), Managed Futures, Positive Feedback Trading,
and Futures Price Volatility, The Journal of Futures markets, 19 , 759-776.
Jarrow, R. A., D. Lando, and S. M. Turnbull (1997). A Markov model for the term
structure of credit risk spreads. Review of Financial Studies 10(2), 481-523.
Jenkinson, A. F. (1955) the frequency distribution of the annual maximum (or
minimum) values of meteorological elements, Quarterly Journal of the Royal
Meteorology Society, 87, 145-158

226

Jenson G. R., Johnson R. R., and Mercer J. M, (2000) Efficient Use of Commodity
Futures in Diversified Portfolio, Journal of Futures Markets, May, 489-506
Jenson G. R., Johnson R. R., and Mercer J. M, (2002), Tactical Asset Allocation and
Commodity Futures, Journal of Portfolio Management, Summer, p. 100-111
Johnson, L. L. (1960) The theory of hedging and speculation in commodity futures.
Review of Economic Studies, 27, 139151.
Johnson, R. R. and Jensen, G.R. (2001), The Diversification Benefits of Commodities
and Real Estate in Alternative Monetary Conditions, Journal of Alternative
Investments, Spring, p. 53-61
Jorion, P. (2000): Value at Risk The New Benchmark for Controlling Market Risk.
McGraw-Hill, New York.
Kaplan, P.D. and Lummer, S. L. (1998), GSCI Collateralized Futures as a Hedging
and Diversification Tool for Institutional Portfolios: An Update, Journal of Investing,
Winter, p. 11-18
Kim, C. J. (1994) Dynamic Linear Models with Markov Switching. Journal of
Econometrics, 60, 1-22.
Kim, C. J. and Nelson, C. (1999) State-Space Models with Regime Switching. MIT
Press.
Kraus, A., and Litzenberger. R. (1976). Skewness preference and the Valuation of
Risky Assets. Journal of Finance .31, 1085-1100.
Levy, H., & Markowitz, H. M. (1979). Approximating expected utility by a function
of mean and variance. American Economic Review, 69, 308317.
Linsmeier, T .J. and Pearson, N. D. (1997): Quantitative Disclosures of Market Risk in
the SEC Release, Accounting Horizons, 11: 107-135.
Linsmeier, T. J. and Pearson, N. D. (1996): Risk Measurement: An Introduction to
Value at Risk, Office for Futures and Options Research Working Paper #96-04, pp.
1-44.
Longin, F. M., (1995). Optimal margins in futures markets: A parametric
extreme-based method. Paper presented at the 7th Chicago Board of Trade Conference
on Futures and Options. September 1994, Bonn, Germany.

227

Longin, F. M., (1996). The asymptotic distribution of extreme stock market returns.
Journal of Business 63, 383-408.
Longin, F. M., (1999), Optimal margin level in futures markets: Extreme price
movements The Journal of Futures Markets; April; 19, pp 127-152
Mahoney, J. M. (1996): Empirical-based versus Model-based Approaches to
Value-at-Risk: An Examination of Foreign Exchange and Global Equity Portfolios,
Proceedings of a Joint Central Bank Research Conference, Board of Governors of the
Federal Reserve System, pp. 199-217.
Maheu, J. M. and McCurdy, T. H. (2000) Identifying Bull and Bear Markets in Stock
Returns, Journal of Business & Economic Statistics (January) 18, 1: 100-112.
Manfredo, M .R., and Leuthold, R. M. (1999): Value-at-Risk Analysis: A Review and
the Potential for Agricultural Applications. Review of Agricultural Economics 21:
99-111.
Manfredo, M .R., and Leuthold, R. M. (2001): Market Risk and the Cattle Feeding
Margin: An Application of Value-at-Risk. Agribusiness 17: 333-353.
Markowitz, H. M. (1952). Portfolio selection. Journal of Finance, 7, 7791.
Markowitz, H.M. (1959) Portfolio Selection: Efficient Diversification of Investments,
New York: John Wiley & Sons, 1959.
Marsh, I. W. (2000). High-Frequency Markov Switching Models in the Foreign
Exchange Market. Journal of Forecasting, 19:123134.
Merton, R. (1969) Lifetime Portfolio Selection: The Continuous-Time Case. Review
of Economics and Statistics, 51, 247-257.
Merton, R. (1971) Optimal Consumption and Portfolio Rules in a Continuous-Time
Model. Journal of Economic Theory, 3, 373-413.
Merton, R., (1973), An intertemporal capital asset pricing model, Econometrica, 41,
867-888.
Merton, R., (1980), On estimating the expected return on the market: an exploratory
investigation, Journal of Financial Economics, 8, 323-362.
Michael, R. S., (2004), commodities enter investment mainstream, Wall Street
Journal. (2004, September 9). News item, p. A1.

228

Michaud, R., (1989), The Markowitz optimization enigma: Is optimized optimal?


Fianncial Analysts Journal 5, 31-42
Michaud, R., (1998), Efficient asset management, Harvard Business School Press.
Muth, J. F. (1961). Rational expectations and the theory of price movements.
Econometrica, 29, 315335.
Myers, R. J. (1991). Estimating time-varying optimal hedge ratios on futures market.
The Journal of Futures Market, 11, 3953.
Myers, R. J. (1994). Time series econometrics and commodity price analysis: A
review. Review of Marketing and Agricultural Economics, 62, 167181.
Myers, R. J., and Hanson, S. D. (1993). Pricing commodity options when the
underlying futures price exhibits time-varying volatility. American Journal of
Agricultural Economics, 75, 121130.
Nelson, D. B. 1992, Filtering and forecasting with misspecified ARCH models I:
Getting the right variance with the wrong model, Journal of Econometrics 52, 61-90.
Nelson, D. B. and Foster, D. P. (1994). Asymptotic filtering theory for univariate
ARCH models. Econometrica 62, 1-41.
Ng V. K, Pirrong S. C. (1994). Fundamentals and volatility: storage, spreads and the
dynamics of metals prices. Journal of Business 67: 203230.
Nijman, T. E. and Swinkels, L. A. P. (2003), Strategic and Tactical Allocation to
Commodities for Retirement Savings Schemes, CentER Discussion Paper, no. 20
Odening, M., and Hinrichs, J. (2002): Using Extreme Value Theory to Estimate
Value-at-Risk, working paper, University of Berlin.
Phillips, P. C. B. and Perron, P. (1988). Testing Unit Roots in Time Series
Regression. Biometrika. 75:335-346.
Pickands J., 1975, Statistical inference using extreme order statistics, Annals of
Statistics 3. 119131.
Pratt, J. and Zeckhauser, R. (1987), Proper risk aversion, Eoconmetrica 55 (1).
Press, W. H., Tekolsky, S. A., Vetterling, W. T., and Flannery, B. P., (1992),

229

Numerical Recipes in C: The Art of Scientific Computing, Cambridge University Press.


Roberts, M. C., (2001). Specification and Estimation of Econometric Models of
Asset Prices. PhD dissertation, North Carolina State University, Raleigh, NC
December
Robert, P. C. and Casella, G. (1999), Monte Carlo Statistical Methods, Springer-Verlag,
New York,Inc.
Roll, R. (1992), A mean-variance analysis of tracking error, Journal of Portfolio
Management 18, 13-22.
Samuelson P., (1967), General Proof that diversification pays, Journal of Financial
and Quantitative Analysis, 2, 1-13
Samuelson, P., (1969) Lifetime Portfolio Selection by Dynamic Stochastic
Programming. Review of Economics and Statistics 51, 239-246.
Satyanarayan, S. and Varangis, P. (1996). Diversification Benefits of Commodity
Assets in Global Portfolio, The Journal of Investing (Spring), pp. 69-78.
Schaller, H., and Van Norden S. (1997), Regime Switching in Stock Market Returns,
Applied Financial Economics, 7, 177-91.
Schneeweis, T., Savanayana, U. and McCarthy, D. (1991), Alternative Commodity
Trading Vehicles: A Performance Analysis, The Journal of Futures Markets, August,
475-490.
Sharpe, W. F. (1963) A Simplified Model for Portfolio Analysis. Management
Science 9: 277-293.
Sharpe W. F. (1964), Capital Asset Prices: A Theory of Market Equilibrium Under
Conditions of Risk, The Journal of Finance, 19, 425-442.
Sharpe, W. F. (1970). Portfolio Theory and Capital Markets, New York, McGraw Hill.
Sortino, F.A. and Satchell, S.E. (2001) Managing downside risk in financial markets:
theory, practice and implementation, Butterworth-Heinemann
Sortino, F.A. and Forsey, H. J. (1996) On the use and misuse of downside risk, The
Journal of Portfolio Management Winter pp 35-42
Tauchen, G. and Hussey, R. (1991) Quadrature-Based Methods for Obtaining

230

Approximate Solutions to Nonlinear Asset Pricing Models. Econometrica, 59(2),


371-396.
Taylor, S. 1986. Modeling Financial Time Series. John Wiley & Sons, Inc. New York.
Timmermann, A. (2000), Moments of Markov Switching Models, Journal of
Econometrics, 96, 75-111.
Tobin, J. (1958), Liquidity Preference as Behavior toward Risk, Review of Economic
Studies, 67, pp.65-86.
Tomek, W. G., and Myers, R. J. (1993). Empirical analysis of agricultural commodity
prices: A viewpoint. Review of Agricultural Economics, 15, 181202.
Tomek, W. G. (1997). Commodity futures prices as forecasts. Review of Agricultural
Economics, 19, 2344.
Tomek, W. G. and Peterson, H. H., (2001) Risk Management in Agricultural Markets:
A Review. Journal of Futures Markets, 21:853-985.
Tsonis, A.A. (1992). Chaos: From Theory to Applications. Plenum Press. New York
and London.
Tvede, L. (1992). What Chaos Really Means in Financial Markets. Futures.
21:34-37.
Vaidynanathan, R. and Krehbiel,T. (1992). Does the S&P 500 Futures Mispricing
Series Exhibit Nonlinear Dependence across Time? The Journal of Futures Markets.
12:659-677.
Vercammen, J. (1995), Hedging with commodity options when price distributions are
skewed, American Journal of Agricultural Economics, 77, 4: 935-945
Venkataraman, S. (1997), Value at risk for a mixture of normal distributions: the use of
quasi-Bayesian estimation techniques, Economic Perspectives: A Review form the
Federal Reserve Bank of Chicago, 21: 2-13
Ward, R. W., and Dasse, F. A. (1977). Empirical contributions to basis theory: The
case of citrus fruit. American Journal of Agricultural Economics, 59, 7190.
Wei, A. (1997). Long Agricultural Futures Price Series: ARCH, Long Memory, or
Chaos Processes? Ph.D. dissertation. University of Illinois at Urbana-Champaign,
1997.

231

Wei, A. and Leuthold, R. M.. (2000), Agricultural Futures Prices and Long Memory
Processes, Office for Futures and Options Research Working Paper 98-03. University
of Illinois at Urbana-Champaign.
Westcott, P. C., and Hoffman, L. A. (1999). Price determination for corn and wheat:
The role of market factors and government programs. Economic Research Service,
U.S. Department of Agriculture, Technical Bulletin 1878.
Williams, J. C., and Wright, B. D. (1991). Storage and commodity markets. Cambridge,
MA: Cambridge University Press.
Winston, K. (1993), The efficient index and prediction of portfolio variance,
journal of portfolio management 19, 27-34.
Working, H. (1949). The theory of price of storage. American Economic Review, 39,
12541262.
Yang, S. R., and Brorsen, B. W. (1992). Nonlinear dynamics of daily cash prices.
American Journal of Agricultural Economics, 74, 706715.
Yang, S. R., and Brorsen, B. W. (1993). "Nonlinear Dynamics of Daily Futures Prices:
Conditional Heteroscedasticity or Chaos?" The Journal of Futures Markets.
13:175-191.

VITA
Shengwu Du
308 Armsby Building
University Park, PA 16802

EDUCATION
PENNSYLVANIA STATE UNIVERSITY (2001-2005, expected in Nov 2005)

Ph.D., Agricultural, Environmental and Regional


(Dual-tile program)

BEIJING UNIVERSITY (1997 2000)

Economics and Operation Research

Beijing, China

M.S., Economics

ZHONGNAN UNIVERSITY OF FINANCE & ECONOMICS (1993-1997) Wuhan, China

B.A., Economics and Statistics

AREA OF INTERETES

Commodity futures and options,

Applied Econometrics, and Industrial Organization

PROFESSIONAL EXPERIENCE

PENNSYLVANIA STATE UNIVERSITY (2001-2005)

Research Assistant, Department of Agricultural Economics and Rural Sociology

CHANGSHENG FUND MANAGEMENT CO., Ltd. (1999-2001)

Senior Financial Analyst, Research & Development Department

Beijing, China

Das könnte Ihnen auch gefallen