Sie sind auf Seite 1von 18

Development of Algorithmic Volatility Trading

Strategies for Equity Options


MS&E 448 Project Report
Gilad Ashpis
Gino Rooney
Ian Schultz
Zach Skokan
Department of Management Science & Engineering, Stanford University
June 9, 2014

Project Introduction

olatility is an increasingly important asset class in finance and trading, and therefore an
understanding of the dynamics and idiosyncrasies of volatility markets is important to
holistic market modeling. Throughout the course of the project, we explored several different
volatility models in hopes of better understanding the dynamics of volatility and finding
profitable trading strategies based off of them. Our modeling techniques can be broken down
into two categories: fitting the implied volatility surface and time-series volatility modeling.
For both categories of models, we formulated and back-tested trading strategies. For several
of the strategies tested based off of our models, we found desirable returns compared to that
of the market portfolio. The paper is organized as follows: in Section 2.1, we provide a brief
overview of the necessary finance needed to understand the project. Sections 2.2 - 2.3 review
the data collection and visualization process. The remaining sections of part 2 go over the
theory behind the volatility models we explored. In part 3, we explain and analyze backtested strategies, which includes a section on portfolio construction and risk management.
Lastly, we conclude with some final thoughts and ideas for future work.

2
2.1

Volatility, Data, and Modeling


Background
n equity option is a derivative contract that gives the holder the right, but not the
obligation, to trade the underlying equity under specified terms. Typically, an option
1

contract specifies the strike price at which the equity can be traded, and during which time
period it can be exercised. Furthermore, a call option is the right to buy the asset, while a
put option is the right to sell the asset. Two primary styles of options are commonly traded:
American and European. An American option allows the holder to exercise the option at
any time prior to the contract expiration, while a European option only allows for exercise
on the expiration date. The option premium refers to the observable price of the option as
traded in the financial markets. The option will be exercised if the payoff exceeds zero; for a
call this occurs if the stock price is greater than the strike price, while for a put this occurs
if the stock price is less than the strike price.
The Black-Scholes model (Black & Scholes 1973) is the classic framework for option
pricing, and one of the most widely adopted options pricing methods in the financial industry.
Assuming the price dynamics of a stock are described by geometric Brownian motion, the
option premium f (S, t) satisfies the partial differential equation:
f
1 2f 2 2
f
+
rS +
S = rf
t
S
2 S 2
where S is the underlying equity price, r is the risk-free rate, and is the equity volatility.
Typically no analytical solution to this equation exists, but in the specific case of a European
call option, the analytical solution C(S, t) exists and can be expressed as:
C(S, t) = SN (d1 ) Ker(T t) N (d2 )
where
ln(S/K) + (r + 2 /2)(T t)

T t
ln(S/K) + (r 2 /2)(T t)

d2 =
T t
d1 =

K is the strike price, (T t) is the time until expiration, and N (.) is the standard normal
cumulative distribution function. Black-Scholes pricing relies on several key assumptions.
Stock prices at any given future date are assumed to take a log-normal distribution. In reality,
prices take on a more fat-tailed distribution (see Borland 2002, for instance). Additionally,
it is assumed that the risk-free rate is known and constant, the stock pays no dividends, and
there are no market frictions such as transaction costs or penalties for short selling. Despite
its shortcomings, the model provides a reasonable and convenient pricing formula. Although
the formula is derived for a European call, it can also be applied to an American call, since
it can be shown that, in theory, for a non-dividend paying stock, the option should never be
exercised prior to the expiration date. The trading strategies developed later in this paper
will focus on call options because of computational ease.
In the Black-Scholes formula, S, K, and (T t) are readily observed in the market, while
r and must be computed from observable market data. The risk-free rate may be estimated
in a number of ways, however, the formula is not particularly sensitive to this input. Rather,
2

the volatility is of much greater significance. In short, higher volatility of the underlying
equity translates into a more valuable option. Although volatility, like the risk-free rate, is
a theoretical quantity, if one assumes a certain risk-free rate, the volatility can be obtained
from the formula because all other inputs are known. This quantity is referred to as the
implied volatility, since rather than being directly observable, it is the value that is implied
by Black-Scholes. The implied volatility can be interpreted as the market expectations for
volatility between the current date and expiration. If one believes that the actual expected
volatility deviates from the implied volatility, a trading opportunity presents itself. When
looking at a curve of implied volatilities plotted against different strike prices for a given
expiration date, one might observe a volatility smile. The implied volatility of an option
with strike price far from the current stock price is often higher than for one with a strike
price near the current price.
The holder of an option is exposed to the risk associated with the price movements of the
underlying equity. This risk can be hedged away by constructing a delta neutral portfolio,
which allows a trader to bet on the price of the option only without betting on the direction
of the equity. Delta is defined as:
=

f (S, t)
S

This measures the sensitivity of the option premium to changes in stock price. Delta is
equivalent to the quantity N (d1 ) from the Black-Scholes formula. The delta of a portfolio is
simply the sum of the deltas of the individual assets. Moreover, the delta of a unit of stock
is equal to 1. A portfolio with delta equal to 0, or delta neutral portfolio, can be constructed
by offsetting the negative delta exposure from call positions with an equity position in
proportionate weighting. However, delta changes constantly as the time to expiration and
the stock price change. In order to maintain a delta neutral portfolio, regular rebalancing
of the positions is required. Trading costs must be taken into account when deciding on the
frequency of rebalancing.

2.2

Data Procurement and Cleaning

aily historical options data were sourced from the OptionMetrics database. The options we examine are the SPDR S&P500 exchange-traded fund (Ticker: SPY), Google
(Ticker: GOOG), and Apple (Ticker: AAPL). These were selected for being amongst the
most actively traded names. Managing and cleaning the data was a non-trivial task; for instance, over the period 2005-2013, 4 million daily quotes were available for all options written
on the SPY index. The dataset exceeded the capacity of a Microsoft Excel spreadsheet, so
by necessity the data pre-processing was performed using python. Each record consists of
the best bid quote and ask quote for a specific date, for each option with a unique strike
price, expiration date, and exercise style. Option contracts are written on 100 underlying
stocks and are priced accordingly. The majority of the records consisted of options with
no trading volume; these were removed from the dataset. To approximate realistic trading
execution prices, the mid price was computed from the bid and ask spread. Each option
3

record had to be matched with the equity price for the current date. Daily historical equity
prices were sourced from the Center for Research in Security Prices (CRSP), which provides
a clean price series adjusted for dividends, splits, and other corporate actions. Similarly,
the risk-free rate had to be merged into the dataset. The zero-coupon U.S. Treasury curve,
sourced from OptionMetrics, was used as a proxy for the risk-free rate. Since rates are only
available for a limited number of maturities, the complete term-structure was obtained by
linear interpolation and matched to each option by expiration date. Finally, the implied
volatility was extracted from the Black-Scholes formula using the bisection method.

2.3

Data Visualization

fter the necessary data had been procured and appropriately filtered, our next objective
was to visualize the data to make sure that it was consistent and identify any potential
sources of error. To do this, we wrote several python scripts to plot features of the data.
Since our project largely relied on spline fitting the implied volatility smile, visualizing the
volatility smile and surface was the central focus of this step. Using our implied volatility
calculator, we wrote a method to plot the volatility smile for any set of expiries on a given
day. Furthermore, we also created a 3D plotting method that would graph the entire implied
volatility surface on a given day, a fitted thin-plate spline, and contours of the resulting
implied volatility smiles from the fit. Figure 1 shows an example output from these methods.

Figure 1: Implied volatility smile and surface plots from the data visualization step
Now that we had taken measures to enable visualization and verify the general consistency of the data we collected, we were ready to start modeling the data and brainstorming
potential trading strategies. We found that this data visualization step and the methods we
created to help with it were particularly useful in identifying anomalies in the data that we
encountered upstream in analyzing our trading strategy.

2.4

Arbitrage-Free Smoothing of Volatility Surfaces

nce the datas cleanliness and consistency had been observed from our visualizations,
our next step goal was to find tradeable opportunities in the data. To find such opportunities, we employed several modeling techniques, one of which sought to create an
arbitrage-free spline fit to the implied volatility smile. It was our hope that we could then
find large enough, tradeable residuals between our arbitrage-free price and the market price
of an option. The approach we took to create arbitrage-free spline fits was first described by
Matthias Fengler and is summarized below (Fengler, 2005).
The problem of fitting a cubic smoothing spline to option prices can be solved with the
following convex optimization problem:
Z b
n
X
2
[g 00 (v)]2 dv
min.
i {yi g(ui )} +
a

i=1

given strictly positive weights 1 , 2 , ... n and smoothing parameter , where (yi , ui )
represent a single options market and exercise price respectively over the set [a,b] of exercise
prices with the same days to maturity on a given day. This minimizer g provides us with a call
price function for a volatility smile whose smoothness is determined by . This minimization
is carried out with chosen linear constraints to ensure our fit is a natural cubic spline.
Furthermore, we can preclude calendar arbitrage by iteratively solving this optimization
problem for each set of expiries on a given day and imposing additional constraints based
on the previous fits results. Constraints are also added to guarantee that the call price
function is non-increasing in the exercise price. Including such conditions results in a convex
quadratic program whose solution provides us with an arbitrage-free call price function
and, consequently, an arbitrage-free implied volatility smile. Formulating the spline fitting
problem in this way provides us with several benefits. Firstly, because the problem is cast
as a convex quadratic program (QP), it can be solved quickly, in polynomial time (Flouda
and Viswewaran, 1995). This speed proved useful to us in back-testing our strategy as we
needed to solve this QP multiple times a day (for each set of expiries) for 8 years. There are
also plenty of available solvers for QPs. We utilized the OpenOpt framework for python, a
universal numerical optimization package that contains several of its own solvers as well as
other popular open source solvers. In consequence, the spline fitting for all of our arbitragefree smiles took only around 10 minutes to complete. Secondly, the minimizer returned is
unique, which guaranteed our results were repeatable. Finally, because a natural cubic spline
can be completely defined by its set of function values and second-derivatives at the knots,
our spline fits could be efficiently stored and evaluated. Efficient storage was useful for us
since the options data we were dealing with was already large. We wrote python scripts
to transform this definition of a cubic spline into a piecewise polynomial that we could use
to evaluate and plot our spline fits. Figure 2 shows an example of such plots from our
implementation.
While the residual between the market and arbitrage-free price is difficult to see in the
entire fit, the zoomed in fit shows more visually noticeable residuals. The date of this fit is
08/15/2007 on the set of options with 66 days to maturity. As seen and further discussed
5

Figure 2: Resulting arbitrage-free spline fit for options with 66 DTM on 08/15/2007 and
underlying of $141.04.
in section 3.1, this day occurred during a time of high market volatility according to the
VIX index, and our strategy seemed to perform best during these time periods for the SPY.
Figure 3 shows a plot of the percent residuals from the same fit versus the strike price.

Figure 3: Residuals between arbitrage-free spline fit and market prices for options with 66
DTM on 08/15/2007 and underlying of $141.04.
For this fit, the residual plot shows differences of up to 1% between the market and
arbitrage-free price. Fig. 3 displays a typical range of residuals found in our spline fits. Once
we had successfully constructed our arbitrage-free fits, we sought to identify potential trading
strategies. The motivation for the strategy we ultimately tested came from examining an
options residuals and the progression of them over time. When there is a positive residual
for a given option, this signals a lower market price than what our arbitrage-free fit suggests.
Thus, our fit suggests that this option is currently undervalued. Figure 4 shows how the
6

residual of an option changes over the life of the option. From this plot, we can see a trend
that the options residual reverts back to zero (or some small value), and often quite quickly.
We found this to be the case for all of the options we examined. Thus, when options become
under or over valued according to the arbitrage free price, they tend to converge back to
an appropriate value in the ensuing days. This inspired our strategy of buying or selling
under or over valued options and realizing their profit on the following day. The decision
parameters and more details of this strategy are described in Section 3.1.1.

Figure 4: An options day-to-day residual over its life

2.5

Time-Series Volatility Prediction Modeling

The other strategy we sought to investigate was using time series models to trade on volatility
forecasts. This differs fundamentally from the first strategy in that it attempts to speculate
on volatility movements, whereas the other looks for arbitrage opportunities based on current
prices. If we predict that the volatility over the term of an option will be different than the
implied volatility based on the market, a trading opportunity is presented. The first type
of volatility modeling we attempted was a linear regression over different macroeconomic
factors. Among these were interest rates, GDP growth, and consumer sentiment. Our linear
regressions excluded a constant term and therefore were of the form:
yj =

n
X

i xji

i=1

While testing different regressions based on macroeconomic factors, we did come up with
some models that seemed promising by strictly looking at their R-squared and p-values. Despite these favorable values, there are some flaws in these models that make them less than
desirable. One of the most important flaws in these models is that our dependent and independent variables do not always have the same time horizon (i.e one month volatilities based
on quarterly delinquency rates). Thus, making an actionable model that uses real-time data
7

would thus be a challenge. Furthermore, if one looks at past volatility along with the economic factors, one can see there is significant collinearity. These flaws (along with the advice
garnered from our talk with SpiderRock Holdings, LLC) led us away from using a linear regression on macroeconomic factors and towards our next type of regression model. The next
type of time series models we tried were different models from the ARCH (Autoregressive
Conditional Heteroscedastic) family of models. Specifically, we decided to model five-day
volatility using the GARCH and EGARCH models. Our models regress on a time series of
five-day log returns. The volatility for these two respective models take on the following form:
GARCH:
t2

=+

p
X

2
i t1

q
X

j 2tj

j=1

i=1

EGARCH:
log(t2 )

=+

p
X

2
i log(ti
)

i=1

q
X

j (|tj | + i tj )

j=1

One must specify p and q, which affect the number of lag coefficients to be estimated by
the model. ARCH models can capture how market volatility changes over time. This would
be useful because the S&P 500 returns exhibit a clustering behavior, which is illustrated by
Figure 5 . There are obvious periods of lower volatility and others of higher volatility.

Figure 5: Weekly Log Returns of S&P 500. Note the clustering behavior.
The main difference between GARCH and EGARCH is an assumption about the effect
of lagged error. GARCH uses the squared error, canceling out the effect of sign. EGARCH,
on the other hand, uses the magnitude of the error as well as the actual error. Thus,
EGARCH can treat positive and negative past returns differently in making its forecasts.
8

Before proceeding with implementing these models, we tested the data to ensure these models
were appropriate. The data tested positive for skew and kurtosis, as well as for ARCH
effects (using the Ljung-Box Q-test). The results demonstrated that our data does have the
characteristics of something that can be modeled using the ARCH family of models. We will
go more into the estimation of our models and results in Section 3.3.

3
3.1

Strategy Back-Testing and Results


Volatility Skew Surface Arbitrage

By comparing arbitrage-free spline fit volatility surfaces to market prices, tradeable mispricings and inefficiencies in the options market can be identified and traded upon. This concept
is at the core of several different back-tested trading strategies presented here. The daily
closing mid price (the average of the bid and ask prices) on each call option was compared
with the arbitrage free spline corresponding to the same strike and time to maturity and
the residual between the two prices was computed. When the magnitude of the residual
between the two prices was larger than a threshold value (in the examples here, the residual was required to be larger than 0.5% of the option price, with an absolute magnitude
greater than $0.01 per share), the contract would either be bought or sold, depending on
the sign of the residual. Because these criteria were chosen arbitrarily, the entire data set
was considered out-of-sample and was used for back-testing. Each position was delta-hedged
with the underlying security at the market mid price in order to minimize sensitivity of the
portfolio to movements in the underlying asset. However, no other efforts to hedge sensitives
were made, and therefore the returns here are potentially partially affected by other Greeks.
Contracts within four days of expiration were not considered in order to avoid the need to
exercise options. Transaction costs were assumed to be $0.005 per share for both equity
and options. After holding the portfolio for one day, the entire portfolio was sold at market
prices and a new portfolio was purchased using the updated arbitrage-free spline fit and
market price residuals. In this way, back-test results here represent a worst-case scenario
as the algorithm potentially trades out of and into the same position each day, which adds
unnecessary transaction costs.
In many cases, hundreds of different option contracts are traded each day for a single
underlying instrument. Therefore, there are potentially many trading opportunities each day,
and rules must be implemented in order to divide assets allocated to this strategy among
the many different opportunities. Here, we explore two different strategies: uniform and
residual-based position weighting. In the uniform weighting case, each trading opportunity
on a given day is assigned an equal share of the total portfolio value. Conversely, in the
residual-weighted case, each position was sized according to the size of its corresponding
residual. In other words, each position was sized according to the its residual normalized to
the total residual of all positions entered on that day.
These volatility skew-arbitrage trading strategies were explored for options on three underlying assets: The State Street Global Advisors S&P 500 exchange-traded fund (SPY),
9

Google Inc. common stock (GOOG), and Apple Computers Inc. common stock (AAPL).
Options markets on these three names have varying degrees of liquidity, for example, in May
2014 over two million AAPL contracts were exchanged, while only 200,000 GOOG contracts
changed hands 1 . Large index-tracking ETFs such as SPY typically have significantly deeper
options markets than individual stocks. In the same one-month period, for SPY there were
nearly 12 million contracts exchanged 2 . Therefore, this study provides insight into how the
strategy scales with the depth of the options market.
3.1.1

Performance of Volatility Skew Arbitrage on the SPY Options Market

Back-test equity curve results (net of costs) for the volatility skew-arbitrage strategies with
uniform weighting and residual-based weighting are shown in Fig. 6. For the uniform
weighting case (left plot), there is fairly consistent initial growth, followed by a rapid, large
spike around mid-2007. This is followed by relatively stagnated performance and increased
volatility in the equity curve. At the end of 2009, the equity curve suddenly spikes upward
and then rapidly shoots down to its previous level again, which corresponds to the largest
drawdown experienced during this test. Subsequently, the strategy settles down and is
profitable, but not spectacularly so. The right hand plot in Fig. 6 shows the same back-test
period results, this time using the residual-based weighting scheme. Here, there is noticeably
less volatility in the equity curve, and the large spike near the middle of 2007 is reduced to
a significant but small jump in the equity curve. However, there is an enormous spike in
the equity value during the period of late-2008 to early-2009. This period corresponds to
the same period of rapid gains and subsequent losses in the uniform-weighted case. After
around mid-2009 the strategy equity shows no drastic jumps or dips and there are only
modest positive returns.
Table 1 contains relevant descriptive statistics for the back-tests of skew-arbitrage trading
on the S&P 500 options. Both strategies use the same trading rules, and therefore the fraction
of days traded, hit rate (defined as the percentage of total trades that are profitable), average
return on winning trades, and average loss on losing trades are all the same. These results
show that the strategy trades on about one-third of all days, with an average return of about
0.55% on both wins and losses. The hit rate for this strategy is 56.2%, which is quite strong
for an inter-day strategy. Statistics unique to each implementation of the strategy include
annual return, information ratio, and max drawdown, and these are also shown in Table
1. Although the uniform-weighted strategy has a higher annualized return, the information
ratio of this strategy is much smaller because the variance of returns is very large. It also
suffers from significant drawdowns. Nevertheless, this information ratio is actually quite
strong when compared to a static position in SPY (IR = 0.14)3 . A particularly strong
information ratio is achieved in the residual-weighting case. This is because the standard
deviation of returns is around 5% on an annualized basis for this strategy, which smooths the
returns significantly. Another important consideration is that the maximum drawdown of
1

Data collected from http://www.cboe.com/data/AvgDailyVolArchive.aspx


Data collected from http://www.cboe.com/data/monthlyvolume.aspx
3
From http://www.macroaxis.com/invest/technicalIndicator/SPYInformation-Ratio
2

10

Figure 6: Equity growth curves for skew-arbitrage trading strategies on S&P 500 options.
Left: Positions taken each day are weighted equally. Right: Positions taken each day are
weighted according to residual magnitude.
Table 1: Performance statistics for volatility skew-arbitrage trading strategy on SPY.
Max.
Days
Hit
Avg.
Avg.
Annual Information
Strategy
Drawdown Traded
Rate
Win
Loss
Return
Ratio
Uniform
22.3%
0.84
55.5%
32.0%
56.2%
0.55%
0.57%
Weighted
6.27%
1.23
5.30%
32.0%
56.2%
0.55%
0.57%
the residual-weighted strategy is more than an order of magnitude smaller than the uniform
weighting case. This characteristic is particularly desirable because many funds take fees
based on a so-called high-water mark, and therefore significant drawdowns make trading
strategies particularly undesirable.
Because a large portion of the total positive returns of each strategy are generated during relatively short time periods, a more detailed study was performed in order to better
understand how these returns are generated. Figure 7 shows the one-year portion of the
residual-weighted S&P 500 curve with the largest gains, from about July 2008 through July
2009. We see that from about October of 2008 through February of 2009, the total equity
grew by more than 30% over a series of profitable trading days. It was suspected that this
high-return period corresponded to stressed, dislocated markets in periods of high volatility.
In these market conditions it is more difficult for market participants to arbitrage away small
inefficiencies in the volatility skew surface, and many participants are likely more concerned
with hedging delta risk on the options market, whatever the price. To confirm these suspicions, the equity curve was qualitatively compared with the S&P 500 volatility index (VIX),
which is a measure of market expected 30-day volatility on an annualized basis. It is evident
from this comparison that the increased returns and performance of the skew-arbitrage strategy corresponds nearly directly with increased volatility beginning around October 2008 and
continuing through about April 2009. This result suggests that volatility skew arbitrage is a
11

particularly attractive strategy during periods of elevated volatility. Additional work could
incorporate regime modeling in order to enact this strategy only when market conditions are
favorable.

Figure 7: Single-year equity growth for skew-arbitrage trading strategy using residualmagnitude weighting on S&P 500 options.

Figure 8: Single-year history of the VIX, which measures market volatility based on S&P
500 options.

3.1.2

Performance of Volatility Skew Arbitrage on the AAPL Options Market

Figure 9 shows equity curve growth for equally-weighted (left) and residual-weighted (right)
positions on AAPL options. As before with SPY options, the equally-weighted strategy
12

Table 2: Performance statistics for volatility skew-arbitrage


Max.
Days
Annual Information
Strategy
Drawdown Traded
Return
Ratio
Uniform
-7.23%
-0.31
75.2%
22.9%
Weighted
1.46%
0.13
31.4%
22.9%

trading strategy on
Hit
Avg.
Rate
Win
48.2%
1.48%
48.2%
1.48%

AAPL.
Avg.
Loss
1.41%
1.41%

under-performs the residual-weighted strategy, and actually provides negative returns over
the back-test period. However, in this case, the residual-weighted strategy provides only very
small total returns over the back-test and is highly volatile compared to returns with the same
strategy on SPY. Strategy performance statistics for the volatility skew-arbitrage strategies
on AAPL are summarized in Table 2. Because the equal-weighted strategy provided negative
returns, it has a negative information ratio. The information ratio of the residual-weighted
strategy corresponds approximately with a static position in the S&P 500. We also observe
that AAPL options provide fewer trading opportunities than SPY, as both strategies trade
on only 22.9% of days. However, when the strategy does trade the average win and loss
of profitable and unprofitable trades, respectively, is near three times larger than the SPY
results. This suggests that when mispricings exist, they are larger in this less liquid option
market.

Figure 9: Equity growth curves for skew-arbitrage trading strategies on AAPL options. Left:
Positions taken each day are weighted equally. Right: Positions taken each day are weighted
according to residual magnitude.

3.1.3

Performance of Volatility Skew Arbitrage on the GOOG Options Market

In a third study, the volatility skew-arbitrage trading strategy was back-tested on GOOG
options. Equity growth time-histories are shown in Fig. 10 for both equally-weighted (left)
and residual-weighted (right) strategies. Here, we see that the same high-volatility period
13

Table 3: Performance statistics for volatility skew-arbitrage


Max.
Days
Annual Information
Strategy
Drawdown Traded
Return
Ratio
Uniform
-30.8%
-0.43
100%
32.0%
Weighted
5.44%
0.65
5.30%
28.5%

trading strategy on GOOG.


Hit
Avg.
Avg.
Rate
Win
Loss
52.2%
2.68%
2.88%
52.2%
2.68%
2.88%

in late 2008 bankrupts the equally-weighted strategy, while the residual-weighted strategy
actually produces strong positive returns over the same period. It is evident from this case in
particular, but also the other two options markets presented, that there is considerable strategy performance improvements to be gained by implementing the residual-weighted strategy
over the simple equal-weighting scheme. Table 3 summarizes the performance statistics for
the volatility skew-arbitrage strategies on GOOG. As expected based on the equity curve,
the weighted strategy provides both a high information ratio and minimal maximum drawdowns over the back-test. This strategy also enjoys a reasonably strong hit-rate. Finally,
note that for this market, which trades the smallest options volumes of the three studied,
the average win and loss are also considerably larger, once again suggesting there exists correlation between the depth of the market and the extent to which inefficiencies in volatility
skew are tolerated by market participants.

Figure 10: Equity growth curves for skew-arbitrage trading strategies on GOOG options.
Left: Positions taken each day are weighted equally. Right: Positions taken each day are
weighted according to residual magnitude.

3.1.4

Distribution of Returns & Understanding Skew Arbitrage Performance

The distribution of residual-weighted volatility skew-arbitrage trading for each of the three
options markets discussed are shown in Fig. 11. These plots show a histogram of returns
for each individual position taken by the algorithm (rather than the distribution of daily
14

returns). The returns from SPY, shown on the left, are narrowly distributed between 2%,
with a visible bias of the mode towards the positive direction. The other two plots in Fig. 11,
for AAPL and GOOG options in the middle and right, respectively, show that the mode is
either near zero or perhaps slightly biased in the negative direction. The AAPL and GOOG
distributions are also significantly wider indicating the larger variance of returns for these
two cases, which leads to lower information ratios. As mentioned previously, we can also
see that the returns are distributed over a wider range of values inversely proportional to
the depth of the options market in a particular security. In Fig. 11, the returns are shown
in order of decreasing depth of market, based on total volume. Proceeding from left to
right, the return distribution is spread out over a larger range corresponding to these market
depths.

Figure 11: Distributions of returns for residual-weighted volatility skew arbitrage strategy.
Left: SPY, Middle: AAPL, Right: GOOG.

3.2

Portfolio Construction and Risk Management

In order to reduce the overall risk of the volatility skew-arbitrage trading strategy, portfolios
of combinations of securities were studied in order to investigate whether there would be
significant diversification benefit from investing the strategy in multiple markets simultaneously. Two different configurations were tested: 1) A portfolio that invests assets equally
in SPY and GOOG, and 2) A portfolio that invests assets equally across SPY, GOOG, and
AAPL. Results from the back-tests of these two strategies are shown in Fig. 14. In both
cases, strong positive returns are observed throughout the test time, although the strongest
trends, occurring during periods of market turmoil, are still plainly evident. These results
suggest that, indeed, diversification substantially improves the performance of this strategy.
Moreover, by diversifying the strategy among many different securities, the overall capacity
will increase substantially.
Table 4 shows statistics on each of these portfolios. Initially it was suspected that the
first configuration would produce superior results due to the lackluster performance of the
volatility skew arbitrage strategy on AAPL options. However, this assumption is incorrect
as evidenced by this data. Although including AAPL options in the portfolio reduced the
15

Figure 12: Back-test returns of volatility skew arbitrage strategy on a portfolio of different
securities (equally-weighted between each security). Left: SPY and GOOG, Right: SPY,
GOOG, and AAPL.
Table 4: Performance statistics for volatility skew-arbitrage trading strategy portfolio of
multiple securities.
Hit Rate
Portfolio
Total Return
Information Ratio Days Traded
SPY + GOOG
147%
0.52
40%
53.9%
SPY + GOOG + AAPL
134%
0.91
58%
54.2%
total return, it dramatically increased the information ratio, and also produced a significant
improvement in the number of days traded, which will help to smooth returns over a shorter
period of time. Finally, it also improved the hit rate to over 54%, which is a strong enough
result to be considered for practical use.

3.3

Time-Series Predictors

There were multiple different specifications on the GARCH and EGARCH models that we
tested when trying to determine which would best be able to predict future volatilities. To
test how well each model specification worked, we took each type and, over each of the last
200 observations, refit the model by each period. After each fit, we used the new model to
predict the next periods volatility. We then checked to see if the predicted volatility was
higher or lower than the actual volatility on the last period, and compared that to how the
volatility actually moved going into the next period. If the predicted volatility was a shift in
the same direction as the actual volatility movement, we considered that a hit, and if they
didnt both move in the same direction it was a miss. The hit percentage for a handful of
GARCH and EGARCH models that we tested can be seen in Figure 13.
From this figure, you can see that none of the model specifications we ran had a hit
percentage over 50% which does not seem promising. They also almost all had the same
16

Figure 13: GARCH and EGARCH models and their Hit Percentages
hit percentage, so it is hard to tell which specifications might be better than the others and
merit further research more than the others. On a brighter note, even though none of the
models have a hit percentage over 50%, when we ran the EGARCH model with specification
(1,1) and return ARMA specification of (2,2) through the same trading strategy that was
used to determine the profitability of the skew arbitrage strategy, we did make money. We
only made a little under 5% over the course of about two and a half years, but it did make
some money. A graph of the cumulative returns, along with a histogram of returns, are
shown in Figure 14.

Figure 14: Histogram of returns and graph of cumulative returns for EGARCH(1,1) with
ARMA(2,2)
Although returns were low and hitting percentages slightly below 50%, these models did
show promise in being able to model and predict future volatilities. With this in mind, we
do think that they merit a deeper looking into, possibly using a different time scale than 5
17

days for the volatility or on a series of data that has more than 536 points on it.

Final Thoughts and Future Work

In all, we found varied success in the multitude of strategies we explored. Our residuallyweighted arbitrage-free spline fitting strategy on the SPY provided us with the most promise
as a potential participant in a regime-switching strategy. While this strategy provided less
impressive results on GOOG and AAPL, we saw some diversification benefits from their
inclusion in a blended portfolio with SPY. In the future, we hope to test this strategy on
more underlying securities and perhaps use portfolio optimization or dynamic programming
techniques to more concretely extract diversification benefits in constructing our blended
portfolio. Although our time-series modeling back-tests were not compelling enough to
make the strategy actionable currently, we believe with more analysis and refinement we
could improve upon them in hopes of including them in our blended portfolio.
In summation, we were able to procure, filter, and visualize options data for several
different underlying securities and then back-test three different options trading strategies
on them. Going all of the way from data procurement to analyzing strategy results gave us
valuable experience in forging and executing a full cycle of a quantitative trading strategy.
While many other classes probably focus on specific parts of this cycle, this class provided us
with a unique opportunity to go through an iteration of the entire process, which seems to be
a more accurate representation of industry practice. Furthermore, we were given the freedom
to pursue newer ideas that we found intriguing while also learning more about commonplace
volatility modeling techniques.

References
Black, F., and M. Scholes (1973), The Pricing of Options and Corporate Liabilities, Journal
of Political Economy, 81, pp. 637-654.
Borland, L. (2002), Option Pricing Formulas Based on a Non-Gaussian Stock Price Model,
Physical Review Letters, 89.
Fengler, M. (2005a). Arbitrage-free smoothing of the implied volatility surface. SFB 649
Discussion Paper No. 2005019, SFB 649, Humboldt-Universitt zu Berlin.
Floudas, C. A. and Viswewaran, V. (1995). Quadratic optimization, in R. Horst and P. M.
Pardalos (eds), Handbook of global optimization, Kluwer Academic Publishers, Dordrecht,
pp. 217270.
Luenberger, D. (2014), Investment Science. Oxford University Press. pp. 374-442.

18

Das könnte Ihnen auch gefallen