AnintroductiontoModernPortfolioTheory: Markowitz,CAP-M, APTandBlack-Litterman

Dr. Graeme West Financial Modelling Agency www.nmod.co.za graeme@nmod.co.za June 26, 2006

Contents
1 Markowitz portfolio theory 1.1 Axioms of the theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 The expected return and risk of a portfolio of assets . . . . . . . . . . . . . . . . . . . . . . . . 1.3 The benets of diversication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 1.4 Delineating ecient portfolios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 1.4.1 Only long positions allowed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 1.4.2 Long and short positions allowed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.3 Risk free lending and borrowing allowed . . . . . . . . . . . . . . . . . . . . . . . . . .. 1.5 Finding the Ecient frontier and in particular the OPRA . . . . . . . . . . . . . . . . . . . . . 1.6 Calibrating the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Capital Asset Pricing Model 2.1 The single index model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 2.2 CAP-M and diversication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 2.3 The Sharpe-Lintner-Mossin CAP-M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 The intuition of CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 A formal proof of CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6 CAPM and prices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Arbitrage Pricing Theory 3.1 The Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 3.2 Consistency of APT and CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 The Black-Litterman model 4.1 Another look at Markowitz and CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Expected returns under Black-Litterman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1 3 3 4 6 6 7 9 9 11 13 15 15 17 17 18 18 19 20 20 22 24 24 25

Some of these notes are based on parts of (Elton & Gruber 1995) and (Harrington 1987). However, the intention here has been to provide as much mathematical rigour as possible, which we believe these two texts avoid as much as possible.

Chapter 1

Markowitz portfolio theory

Denition 1 The simple return on a nancial instrument P is This denition has a number of caveats: The time from 1 to t one business day. Thus it is the daily return. We could also be interested in t is monthly returns, annual returns, etc. P t price. Sometimes a conversion needs to be made to the raw data in order to achieve this. For is a example,ifwestartwithbondyields t ,itdoesn'tmakemuchsensetofocusonthereturnasformulated y above. Why? We need to worry about other income sources, such as dividends, coupons, etc.
P The continuous return is lnt t 1 . This has better mathematical and modelling properties than the P simple return above. For example, it is what occurs in all nancial modelling and (hence) is what we used for calibrating a historical volatility calculator. However, it is bad for portfolios - see Tutorial. See (J.P.Morgan & Reuters December 18, 1996, TD4ePt2.pdf,for additional information and clues. 4.1) Pt Pt Pt 1
1

TheMarkowitzframework(Markowitz1952)isoftengenericallyknownasthemean-varianceframework. The assumptions (axioms) of this model are 1.Investorsbasetheirdecisionsonexpectedreturnandrisk, asmeasuredbythemeanandvarianceofthe returns on various assets. All investors have the same time horizon. In other words, they are concerned only with the utility of 2. theirterminalwealth,andnotwiththestateoftheirportfolio beforehand, andthisterminaltimeisthe same for all investors.
1 Allinvestorsareinagreementastotheparametersnecessary,andtheirvalues ,intheinvestmentdecision 3.
1 This

1.1 Axioms of the theory

means that information is freely and simultaneously available to all market participants.

Figure 1.1: An individual will prefer B, C and D to A - but which one is most preferable? making process, namely, the means, variances and correlations of returns on various investments. (We say the investors are homogeneous.) 4. Financial assets are arbitrarily fungible. The basic tenant of the Markowitz theory is that knowing the mean and standard deviation of the returns on theportfolioissucient,andthatourdesireistomaximisetheexpectedreturnandtominimisethestandard deviation of the return. The standard deviation is the measure of riskiness of the portfolio. Onethingthatisobvious(becauseindividualsareutilitymaximisers),isthattheywillalwaysswitchfrom one investment to another which has the same expected return but less risk, or one which has the same risk but greater expected return, or one which has both greater expected return and less risk. See Figure 1.1.

1.2 The expected return and risk of a portfolio of assets th

Suppose we have a portfolio with n assets, the of which delivers a return at timet. This return has a i R t;i 2 mean t;i andavariance . Suppose the proportionof thevalue ofthe portfolio that iasset makesup is i w t;i P n (so i =1 wi =1). What is the mean and standard deviation of the return portfolio? All known values are assumed R of the to be known at time and thet will be implicit in what follows. We can suppress the subscript t, t as long as we understand that all of the parameters are dynamic and we need to refresh the estimates on a daily basis. " := E [R]= E
n X i =1

# wi Ri =

n X i =1

wi E [R i ]=

n X i =1

wi i

(1.1)

and 2 (R) = E (R )2 " n # X =E ( wi (R i i ))2 2 =E4 = = =

i =1 n n X X i =1 j =1 n n X X i =1 j =1 n n X X i =1 j =1 n n X X i =1 j =1

3 wi wj (R i i )(R j j )5

E [wi wj (R i i )(R j j )] wi wj covar( i ;R j ) R wi wj i;j

= w0 w 2 0 B B wherew = B B @ w1 w2 . . . wn 1 6 6 C 6 C 6 C and = [ ] = 6 i;j C 6 6 A 6 4 11 . .. . . . . .. . . . . . . n1 1n . . . . . . . .. . . . nn 3 7 7 7 7 7. This is called the covariance matrix. 7 7 7 5

So, the return on the portfolio has

0 E [R] = w p (R) = w0 w:

Note that is the covariance between R i the return on asset i andR j the return on asset j. ij
2 = is the variance ofi . R ii i

= ij

ij i j

is the correlation ofi andR j . R

We will denote the covariance matrix by ; 2 6 6 6 6 the correlation matrix [ 6 ]= 6 ij 6 6 4

11 . .. . . . . .. . . . . . . n1

1n . . . . . . . .. . . . nn 5

3 7 7 7 7 7 by P; 7 7 7 5

 0 6 . 6 0 . . 2 6 6 . . .. 6 . . the diagonal matrix of standard deviations . . 6 . 6 . .. 6 . . 0 4 . 0 0 n Then = DPD and so = (R) p w0DPDw

3 7 7 7 7 7 by D. 7 7 7 5

(1.2) (1.3)

Letusconsidersomespecialcases. Supposetheassetsareallindependent,inparticular,theyareuncorrelated, P n 2 2 so = . ( is the indicator function.) Then(R) = i =1 wi2 . Suppose further that the portfolio is ij ij ij i 1 equally weighted, wi = n for every Then so i.
2 (R) = n X i =1 2 1 2 1X i = ! n2 i n i =1 n n

1.3 The benets of diversication

as n ! 1If we accept that variance is a measure of risk, then the risk goes to 0 as we obtain more and . more assets. Suppose now that the portfolio is equally weighted, but that the assets are not necessarily uncorrelated. Then
2 (R)

= =

n n X X

1 ij n2 i =1 j =1
n n n 2 X 1X n 1X i + n i =1 n n i =1 j =1;j

6 = 6 = 6 =

ij n(n 1)

= !

1 2 n 1 + ij;i n i n j asn !1 ij;i

The limit is the average covariance, which is a measure of the undiversiable market risk.

1.4 Delineating ecient portfolios

Remember that in the theory we are dealing with now, the mean and standard deviation of the return on an assetorportfolioisallthatisrequiredforanalysis. Weplotcompetingportfolioswiththestandarddeviation on the horizontal axis and expected return on the vertical axis (risk/return space).

Figure 1.2: Two risky assets held long; diering correlations.

1.4.1 Only long positions allowed

Consider the case where we have a portfolio of two risky assets, both held long. Then w1; w 2 0

w1 + w2 = 1 E [R] = w1 1 + w2 2 " #" #" h i 0 1 1 1 2 (R) = w1 w2 0 1 0 2

2 2 2 2 = w1 + 2w1w2 2 + w2 1 1 2

0 2

#"

w1 w2

We can consider some special cases:

2 If = 1, then (R) = (w1 + w2 )2, so (R) = w1 + w2 . Then we have a straight line in 1 2 1 2 risk/return space. 2 2 2 2 2 If =0, then (R) = w1 + w2 . We have part of a hyperbola in risk/return space. 1 2 2 If = 1, then (R) = (w1 w2 )2, so (R) = jw1 w2 j. Then we have a \hooked line" in 1 2 1 2 risk/return space.

These possibilities are always paths connecting) (to ( ; 2). The portfolio which has the least risk ; 1 1 2 2 iscalledtheminimumrisk/varianceportfolio. Since=1 w1, (R) isalwaysanupwardsfacingquadratic w2 in w1, so the value of1 where this minimum occurs can be found by writing down the axis of symmetry of w 2 the parabola.One gets 2 2 1 w1 = 2 2 2 : (1.4) + 2 2 1 1 2
2

This might lead to a value of 1 outside [01], in which case the minimum occurs at one of the endpoints. w ;

Figure 1.3: Three risky assets held long.

2 Since (R) is greatest when 1 (all other factors xed) the path always lies to the left of a straight line = between two points on the path. This means that the curve is convex.

Now we suppose there are possibly more than two assets. The set of attainable points in risk/return space is no longer a line segment or curve but a \cloud" of points. Given any two points in the set, a convex path within the set connects them. The minimum risk portfolio exists, being the leftmost point on this set. The maximumreturn portfolio exists, being theportfolio whichconsists only of theasset with the highest expected return. Since any rational investor will Prefer a greater expected return to a smaller, risk being xed, Prefer a smaller risk to a greater, expected return being xed, the set of points that we need consider (and the corresponding portfolios) form the so called ecient frontier. Denition The ecient frontier is the set of portfolios on the upper left boundary of the attainable set, 2 between the minimum variance portfolio and the maximum return portfolio. Which one of these portfolios will any investor choose? They will choose the one that maximises their utility. Thus, they will choose the portfolio marked in the diagram. Everybody's utility curves are dierent (and unquantiable, but that is another matter). Thus everybody will choose dierent portfolios. 8

Figure 1.4: The portfolio chosen to maximise the utility of the individual investor

How to nd the ecient frontier here is problematic. Optimisers are required. In the next section there will be no constraints on short selling, and the problem becomes one of pure linear algebra, and so can be solved in an understandable way.

1.4.2 Long and short positions allowed

P n Ifshortsalesareallowed,theconditionthat i 0 disappears,whiletheconditionthat i =1 wi =1remains. w Most parts of the above analysis remain or generalise easily. Thereisnownoupperboundtotheecientfrontier: itdoesnotendwiththe\maximumreturn"portfolio as dened above. This is because we can short sell the asset with low expected returns (or generally, some combination of assets with low expected returns) and use the funds to go long assets with higher returns. Since this can be done to an arbitrary level, the ecient frontier continues without bound. Denition 3 ecient frontier with shorts allowed is the set of portfolios on the upper left boundary of The the attainable set, from the minimum variance portfolio and increasing without bound.

1.4.3 Risk free lending and borrowing allowed

This is simply another asset with and known return say. Lending is long and borrowing is short. =0, r Let w be the proportion of the portfolio invested in risky assets, with risk statistics p, , and 1 w in p

Figure 1.5: Three risky assets held long or short

the risk free account. Then E [R] = (1 w)r + w p = r + w( p r) (R) = w p ) E [R] = r + p r (R) p
r

which is a straight line. When=0, we get the vertical intercept at ). The slope is p p . The rational w ;r (0 investor, preferring higher returns to fewer returns at a xed level of risk, will want this line to have as steep a gradient as possible. Thus, we start with a very steep downward sloping line going through (0rotate it anti-clockwise ;r ) and while we have an attainable portfolio. The `last' portfolio is called the Optimum Portfolio of Risky Assets (OPRA).Wewillquantifyitbyobservingthatofallsuchlinesgoingthroughtheattainableregionitistheone with maximal gradient. The investor can now place themselves anywhere on this line through an appropriate amountoflendingorborrowingandusingtheremainder(whichcouldbegreaterthanone,ifborrowingoccurs - this is known as gearing) to buy the OPRA. The ecient frontier is no longer the best we can do for portfolio selection, because we can always do better: we can place ourselves on the ray diagrammed, which is called the capital market line. Finding the capital market line becomes merely a function of knowing the risk free return and nding the OPRA.

10

Figure 1.6: Finding the OPRA

1.5 Finding the Ecient frontier and in particular the OPRA

We only consider the case where short sales are allowed. Inwe saw that even with 3 assets it is not obvious computationally how to nd the ecient frontier. 1.4 In fact we can nd the ecient frontier by manipulating the concept of OPRA. What we do is hypothetically vary the risk free rate. For each risk free rate an OPRA . All of r we get r these OPRA 's form the curved ecient frontier. r The problem thus reduces to nding the OPRA for any risk free We have r. rate
n X @ 0 w w =2 wj ij @w i j =1

(1.5)

Let =

p r p

. is known as the market price of risk of the portfolio; more on this later (in this course and

11

Figure 1.7: Finding the Ecient Frontier P n elsewhere!). We need to maximise constraint is j =1 wj =1. . The P n j =1 wj ( j r) p = w0 w P p P n 2 n=1 wj ij j p w0 w( i r) @ j =1 wj ( j r) 2 w 0 w ) = @w w0 w i P n n X j =1 wj ( j r) ) 0 = i r wj ij w0 w j =1 0 1 2 3 0 1 1 1 w1 B C 6 7 B C B 2 C 6 17 B w2 C B . C r 6 . 7 B . C ) 0 = B C 6 . 7 B . C @ . A 4 . 5 @ . A . n 1 wn where it happens that = P
n j =1 wj ( j w0 w

r)

(1.6)

is known as the Merton proportion. Thus 0 1 w1 B C B w2 C B . C = B C @ . A . wn

3 1 r 6 7 6 2 r 7 16 7 (1.7) . 6 7 . 4 5 . n r P n and we can solve for1; w 2; :::; w n , using the fact that j =1 wj = 1. Having done so, the point ( w ;) calculated in the usual way with these weights is on the ecient frontier. 12

To summarise By varying we can nd the ecient frontier, r, By xingr (to be what it really is) we can nd the OPRA, The OPRA and r together give us the capital market line.

1.6 Calibrating the model

Nobody uses the Markowitz model anymore. However, if we were to use a toy model with real data, the parameters that are required might be found using the EWMA method. Likewise for CAP-M, which features in the next chapter. For convenience, these methods are repeated below: The data available x0; x 1 :::; x is ; pi
t:

xi = ln (1 i t) xi 1 v u 25 u X t 10 (0) = p2 i q (i) =
i =1 2 (i 1)+ (1 )p2250 (1 i t) i

(1.8) (1.9) (1.10)

The rolling calculator for covolatility (i.e. annualised covariance) - see (Hull 2002, 17.7) - is 25 ! X covol ( pi (x)pi (y) 10 0 x;y) =
i =1

(1.11) (1.12)

covol x;y) = covol 1(x;y) + (1 )pi (x)pi (y)250 (1 i t) i ( i Following on from this, the derived calculators are (x;y) = i i (x;y) = covol x;y) i ( (x)(y) i i covol x;y) ( i (x)2 i

(1.13) (1.14)

the latter since the CAP-M is the linear coecient in the regression equation iny is the dependent which variable and is the independent variable. The CAP-M intercept coecient to be found via rolling x has calculators. Thus
25 X p1(x) = 10 pi (x) i =1

(1.15) (1.16) (1.17)

pi (x) = pi 1(x) + (1 )pi (x)250 (1 i t) and likewise for p(y). Then (x;y) = pi (y) i (x;y)pi (x) i

However, the historical approach would not be used for nding the expected return: the above measure is historical, and while one can claim that history will provide a good estimate for the other parameters, it 13

is unlikely to provide a good measure for expected returns. Here, we might be more reliant on subjective or econometriccriteria. Noinformationinthisregardisprovidedby(Elton&Gruber1995)or(Harrington1987). All that is provided in (Elton & Gruber 1995) are some toy examples with annual (historical) returns being used as predictors of drift. This is indeed a serious shortcoming of the model. We will see a partial answer to this question in Chapter 4.

14

Chapter 2

Capital Asset Pricing Model

2.1 The single index model

The fundamental objection to the Markowitz theory is the need for 2 n + n parameters. While typically in 2 this course we use values derived from historical data analysis, for an institution to add value they will need to forecast parameters, which is impractical if the number of required parameters is large. The model of Sharpe (Sharpe 1964) is the rst simplied model - the simplication is in the data requirements - and led to Markowitz and Sharpe winning the Nobel Prize in 1990. All that is required is parameter estimation of how the security will behave relative to the market. Estimation of pairwise behaviour is not required. The model starts with a regression equation R i (t) = + i R (t) + ei (t) i where R(t) i R i (t) i i ei (t) the return at timefor the market t the index for a single security the return on the single security timet i at the -parameter of security i the -parameter of security i a random variable, with expectation 0, and independent from R(t). (2.1)

Thisisallpurelyregressionanalysis: i (t) (dependentvariable)isregressedin R t against R(t) (independent variable), with linear termand constantcoecient. Insample, regressionanalysis ensures that (t) i ei the i have sample mean 0 and thatR(t) are uncorrelated ei (t). the to Our regression analysis will be performed using Exponential Weighted Moving Averages. 6 = The fundamental assumption of a single index model is that the ei are independent iin in other words, ei and ej are independent for jHence equities move together systematically only because of market i . movement. Thereisnothingintheregressionanalysisthatensuresthatthiswillbetrueinsample. Itdoesn't even make sense practically. After all, resource stocks might move together in some statistically signicant 15

sense, nance stocks likewise, etc. Thus, the major generalisations of the CAP-M model were multifactor models. There is even a concept of multi-linear regression in the EWMA scheme, which can be used here for calibration. We won't consider these multifactor models in this course. See (Elton & Gruber 1995, Chapter 8), for example. Note that E [R i (t)] = + i E [R(t)] i and so i = + i i It follows that and so
2 2 = i2 2(R) + (ei ): i

(2.2)

2 E (R i R i )2 = E ( i (R R) + ei )2 = i2 2(R) + (ei ) (2.3)

Similarly we have E (R i R i )(R j R j ) = E ( i (R R) + ei )( j (R R) + ej ) = i j2(R) and so = i j2(R): ij Hence, the covariance matrix of these assets is given by n 2 2 2 2 1 (R) + (e1) 1 2 2(R) 1 n 2(R) 6 . 6 2(R) . 2 2 22(R) + (e2) . 6 2 1 = 6 . . .. 6 . . . . 4 . 2 2 n 2(R) ::: ::: n2(R) + (en ) 1 2 2 3 (e1) 0 0 6 7 . 6 . 7 2 0 (e2) . 6 7 6 7 . . 0 2 .. 6 7 . . = (R) + 6 . . . 7 6 7 . .. 6 7 . . 4 5 . 0 2 0 0 (en )
0 := 2(R) + e

(2.4) 3 7 7 7 7 7 5

(2.5)
n,

We are now in a situation to consider portfolios. Suppose we have a portfolio with ;weights w w1 w 2; :::; and model parameters; ; :::; n and 1; ; :::; n . Let 2 1 2 P P This denition is motivated by the fact that E [R P (t)] = + P E [R(t)]: P 16 = =
n X i =1 n X i =1

wi i wi i

(2.6) (2.7)

(2.8)

As a crucial special case of this we can suppose market portfolio ie. the portfolio that makes up P is the M the market that is being used in this analysis. Then and P =1. Thus an equity with > 1 is to M =0 P be thought of as being more risky than the market and one with to be thought of as being less risky < 1 is than the market. TheecientfrontiercannowbefoundusingthesametechniquesasthatfortheMarkowitzmodel. However thenumberofparametersnowrequiredforestimationis3 the 's, n+2: 's, (e)'s,thevolatilityofthereturn on the market and the expected return on the market.

2.2 CAP-M and diversication

For any portfolio P
2 (P ) = w0 w 0 = w0( 2(R) + e w ) n !2 n X X 2 2 = wi i (R) + wi2 (ei ) i =1 1 Now letwi = n . Then i =1

2 1 2 2 (P ) = 2(R) + (e) n

and so ) ! (P (R): This rearms thati is a measure of the contribution of th security to the risk of the portfolio. (R) i the i is called the market, or undiversiable, risk of security i ) is called the non-market risk, unsystematic i. (e risk, unique risk or residual risk of equity risk is diversiable. i. This

The CAP-M is what is known as an equilibrium model. The market participants as a whole act to put the market into equilibrium. A number of additional simplifying assumptions (over and above those of Markowitz) are made in the CAP-M which are thought to be not too far removed from reality, yet are useful in order to simplify (or even makepossible)thederivationofthemodel. Ofcourse, asetofsuchassumptions isnecessaryinanyeconomic model. In this model, they are: 1. Short sales are allowed. 2. Thereisariskfreerateforlendingandborrowingmoney. Therateisthesameforlendingandborrowing, and investors have any amount of credit. 3. There are no transaction costs in the buying and selling of capital assets.

2.3 The Sharpe-Lintner-Mossin CAPM

17

4. Similarly, there are no income or capital gains taxes. 5. The market consists of all assets. (No assets are exclusively private property.)

2.4 The intuition of CAPM

By homogeneity, everybody has the r and the same OPRA. This is then the market. same Now accept that is the appropriate measure of risk of a security. This is intuitive because all investors areholdingsome amountofthemarket,and thenon-systematic risk hasbeen diversied away. Byclassic no1 arbitrage considerations, all securities lieonastraightline when plotted in space. Twosuchportfolios are the riskless asset alone and the market alone. Thus and (1 M ) are points on this line in ;r ) (0 ; space, so the line has equation i r = i ( M r) (2.9) which is known as the equation of the security market line, and `is' CAP-M.

2.5 A formal proof of CAPM

Recall from 1.5that when nding the optimal portfolio, w = r1 and so w0 w = w0( r1): (2.11) Sincewehavehomogeneity,allinvestorsselectthesameoptimumportfolio. Thusinequilibriumthisportfolio is the market i.e. all securities are represented and their weights are the weights they have in the market. 2 Thus,w0 w = (M ) and so 2 (M ) = M r (2.12) and we have once again solved for also, . But
n X j =1 n X j =1

(2.10)

wj ij

wj E [(R i i )(R j j )]
n X j =1

3 wj (R i i )(R j j )5

= E4

= E [(R i i )(R M M )] = : iM We now substitute into the line of (2.10): i th M r =i r iM 2 (M )

1 When plotting these, if we have a security say above the line, we buy it and go short a security on or below the line. We can arrange to have a portfolio with no risk (zero no investment required, and positive return. ),

18

or i =r + which is the same as (2.9).

M r = r + i ( M r) iM 2 (M )

(2.13)

HerewemanipulatetheCapitalAssetPricingModelintoanequivalentequationofprices,ratherthanreturns. P P Suppose we are at time t, and the horizon is at 1. Let us denote prices P . Thus R = t +1 t t . t+ by P ( +1 (P ) covarP t +1 ;i ;P t +1 ;M ) It follows from this that = and = Ptt;M ;M . Substituting into (2.13) we iM M P t;i P t;M get P t+1;i P t;i P t+1;M P t;M covar( t+1;i ;P t+1;M ) P t;M P =r + r : 2 P t;i P t;M (P t+1;M ) P t;i Multiplying through byt;i , we have P covar( t+1;i ;P t+1;M ) P P t+1;i P i (t) = P i (t)r + P t+1;M P t;M rP t;M : 2 (P t+1;M ) Rearranging, we get P t+1;i (P t+1;M P t;M rP t;M ) or P t;i = covar( t+1;i ;P t+1;M ) P = Pt;i + P t;i r; 2 (P t+1;M ) (2.14)

2.6 CAPM and prices

1 covar( t+1;i ;P t+1;M ) P P t+1;i (P t+1;M (1+ r)P t;M ) 2 1+ r (P t+1;M )

which is the price equilibrium equation.

1 Such equations are common inMathematics of Finance. 1+r represents the risk-free discounting function, isthe expected value of the asset, and the remaining negative term indicates a compensating factor for P t+1;i the investor's willingness to take on risk. The term in square brackets is known as the certainty equivalent.

P t+1;M (1+ r)P t;M t+1;M ) (P is the market price of risk, while covar( t+1;i ;P t+1;M ) P t+1;M ) (P

(2.15)

(2.16)

is the price of risk associated with the individual security.

19

Chapter 3

Arbitrage Pricing Theory

The rst generalisations of CAP-M were multi-index models. Later Steven Ross (Ross 1976) developed a dierent approach to explain the pricing of assets: pricing can be inuenced by any `abstract' factors.

3.1 The Theory

The axioms of the Arbitrage Pricing Theory are Investorsseekreturntempered byrisk; theyarerisk-averseand seektomaximise theirterminal wealth. 1. 2. There is a risk free rate for lending and borrowing money. 3. There are no market frictions. Investors agree on the number and identity of the factors that are systematically important in pricing 4. assets. 5. There are no riskless arbitrage pricing opportunities. The model starts with a linear equation R i (t) = ai + where I j (t) i R i (t) ai b ij ei (t)
J X j =1

b I j (t) + ei (t) ij

(3.1)

the value at timeof index , J n 2. t j the index for a single security the return on the single security timet i at thea-parameter of security i the sensitivity of the return on securitythe level of index i to j a random variable, with expectation 0, and ....

theei are assumed independent of each other, and independent of Ij . 20

Thisappearstobeamulti-factorversionoftheCAP-Mmodelwherethereturnsaresensitivetothelevels of indices, rather thantothe returns of thesingle index (the market). However, we nowtake aquite dierent turn, because typical APT indices include (Roll & Ross Fall 1983): unanticipated changes in ination unanticipated changed in industrial production unanticipated changes in risk premia, as measured in corporate bond spreads unanticipated changes in the slope and level of the term structure of interest rates The APT theory involves a derivation of an equilibrium model, via an assumption of homogeneous expectations. Suppose we have a portfolio P with weights 1; w 2; :::; w w RP =
n X i =1 n.

Using the model, we have

wi Ri 0 wi @ai +
J X j =1

n X i =1

1 b I j + ei A ij
n X i =1

= As a consequence, we dene/have

n X i =1

wi ai +

n J X X i =1 j =1

wi b I j + ij

wi e i

aP b Pj eP RP RP

= = =

n X i =1 n X i =1 n X i =1

wi ai wi b ij wi e i
J X j =1

= aP + = aP +

b jI + eP Pj b jI : Pj

J X j =1

Sinceei canbediversiedawayintheportfolio,theinvestormustnotexpectrewardfor Thustheinvestor ei . isonly concerned about the risk parameters the valuesof . The b arethe appropriate measure of b and Ij Pj Pj risk for a portfolio, and if two portfolio's have theb then they have the same risk. same 's

21

3 b 1J 6 7 6 7 6 7 6 7 6 7 6 b 7 2J 7, b = 6 7, ::: , b = 6 . 7. This forms a set of vectors in n . Append Consider 1 = 6 b J R J 6 7 2 6 7 6 . 7 4 5 4 5 4 . 5 b nJ 2 3 2 3 1 R1 6 7 6 7 6 17 6 R2 7 1 = 6 . 7 2 Rn to this set, and call S. Let E [R]= 6 . 7 2 Rn be the vector of expected returns of the it 6 . 7 6 . 7 4 . 5 4 . 5 1 Rn n securities in the market. b 11 b 21 . . . b n1 b 12 b 22 . . . b n2 Suppose 2 S ? , and consider the portfolio induced by w1; w 2; :::; w n ). Sincew ? 1, we have w P ( P n =0, so there is no cost to entering into P . Sincew ? b , we have i =1 wi b =0 for 1 j J , so j ij Pbearsnoindex risk. Also, bysome diversication arguments, wecan assumethe residual riskis0. Thus, P P n has no expected return, so wi Ri =0, and w ? E [R]. i =1 P
n i =1 wi

Thus,S ? E [R] , and so E [R] 2 S ?? =span S and so E [R]= 1 + for some 1 ; :::; ; 2
J J X j =1

j bj :

2 R. Thus Ri = +

J X j =1

j bij

(3.2)
J

for 1 i n and some 1 ;2:::; J 2 R. Note that this holds for every ; ; i, with and 1 ; :::; ; 2 the same each time. It remains to nd out what they are. Note that for any P , portfolio RP = +
J X j =1

j bPj :

Considering a riskless portfolio, with=0 for each , we have = r. Hence b j Pj RP = r +

J X j =1

j bPj :

Choosing a portfolio with = 1 once only, andPj = 0 otherwise - this is the portfolio exposed only to b b Pj indexj , so we must think of it as being a position in index - we have j itself RI j = r + j and so j = R I j r. Thus Ri = r +
J X j =1

(R I j r)b : ij

(3.3)

3.2 Consistency of APT and CAPM

CAP-M says that R I j = r + I j (R M r): 22

Hence Ri = r + where

J X j =1

(R I j r)b = r + ij

J X j =1 J X

I j (R M r)b := r + i (R M r) ij

i =

Ib j

ij

j =1

23

Chapter 4

The Black-Litterman model

This model was introduced in (Black & Litterman 1992).

Let us start by revisiting expected returns of each stock under the Markowitz model and under CAP-M. Under the Markowitz model and in equilibrium, the ratio
@ @w i

4.1 Another look at Markowitz and CAPM

i i p := 0 w i w

(4.1)

is a constant. See (Litterman 2003, Chapter 2). More precisely, it should be a constant, for if 1 2 < 1 2 then by decreasing the holding of asset 1 and increasing the holding of asset 2, we can obtain an increase in the expected return of the portfolio without an increase in risk. To be precise, sell asset 1 and buy worth of 1 worth of asset 2. 2 The change in expected cash return 1 + 1 2 is 0. However, the change in risk is given (using 2 > 1 Taylor series) by 1 2 2 to rst order. Thus, we have an improved position, with the same risk + =0 but greater return. Now, @p 0 ww = @w i = and so = Kw 24 (4.2)
n X 1 p 2 wj ij 2 w0 w j =1

n X 1 wj ij w0 w j =1

for some constant, sometimes called the risk aversion coecient. Here as before vector of market K w is the capitalisation weights, and will be the (historical) covariance matrix. Thisapproachactuallyenablesustoreducetheestimationofexpectedreturnstothechoiceofrisk-aversion coecient. Having chosen our own personal risk aversioncoecient, we can nowderive the portfolio we must hold via (1.7).

The Black-Litterman model essentially allows us to have some views on some of the stocks in the portfolio. Thus, the investor is not asked to specify a vector of expected excess returns, one for each asset. Rather, the investor focuses on one or more views. If,forexample,wehaveoneview,namelythatoneofthestocksisgoingtoperformbetterthanthereturn foundintheprevioussection,theportfoliowithviewswillbebroadlysimilartothemarketportfolio,butwill be overweight in that share. This overweightedness is known as a tilt. It will in fact also have tilts towards shares that are strongly correlated to it. Moreover, the degree of tilt will be a function of the condence that we have in that view. Let the equilibrium return vector found above. The vector of expected returns including views is be given by 1 v = ( ) 1 + P 0 1P D () 1 + P 0 1V D (4.3) where v is the new (posterior) expected return vector, is a scalar, is the covariance matrix of returns, P is a k n matrix identifying the stocks in k dierent views, the D is a k k diagonal matrix of error terms in expressed views representing the level of condence in each view (diagonal since these error terms are residuals, and so assumed independent of each other), V is thek 1 view vector. We will focus on three dierent types of views (Idzorek 2002): 1. A will have an absolute return of 10% (Condence of View = 50%); 2. A will outperform B by 3% (Condence of View = 65%); A 3. market weighted portfolio of A and B will outperform a market weighted portfolio of C, D and E by 1.5% (Condence of View = 30%). and we can comment on these views as follows:

4.2 Expected returns under BlackLitterman

25

If1. has an equilibrium return of less than 10%, the BL portfolio will tilt towards A, if more than 10%, A the BL portfolio will tilt away; 2.If A has an equilibrium return of less than 3% more than B, the BL portfolio will tilt towards A and away from B, if more than 3% more than B, the BL portfolio will tilt away from A and towards B; As3. above, but we compare the weighted equilibrium return of A and B with the weighted equilibrium return of C, D and E. 3 2 0:10 0:50 1c 0 6 7 6 Now, in the above example, = 4 0:03 5, D = 4 V 0 0:65 1c 0:15 0 0 factor to be discussed shortly, and 2 1 0 0 0 0 6 P =4 1 1 0 0 0 wA wB wC w wE wC +wD +wE wC +wD +wE wC +wD +wE wA +wB wA +wB D talisation weights. 2 3 0 7 0 5 wherec is a calibration 1 0:30 c 3 7 w where the are market capi5

The scalar is more or less inversely proportional to the relative weight given to the implied equilibrium returns. The only issues outstanding are the parameter the calibration factor These problems are in fact and c. related, and the literature is not clear about resolving this problem. According to (He & Litterman 1999), (Idzorek 2002) we have c = = 10 P P 01k k trace( ) D 2kc
1 2

(4.4) (4.5)

Having determined the expected returns, we choose the portfolio with weights given by w = K 1( v r1n ) where again is the risk aversion parameter (and again be eliminated if desired). K K can If all views are relative views then only the stocks which are part of those views will have their weights aected. If there are absolute views then all the stocks will have change in returns and weights, since each individual return is linked to the other returns via the covariance matrix of returns. As usual as soon as there are constraints on short selling the model is much more dicult to use. One possibility is to use the vector of returns with views in a Markowitz model which has those short selling constraints included, as1.4.1. in (4.6)

26

In the following example we nd our optimal portfolio with the only extraneous input being the risk aversion factor . K w 36.66% 4.30% 23.37% 6.55% 15.05% 3.54% 10.53% 31.24% 37.80% 32.08% 38.25% 34.96% 30.32% 30.78% Correl AGL AMS BIL IMP RCH SAP SOL AGL AMS BIL IMP RCH SAP SOL 1 AGL AMS BIL IMP RCH SAP SOL =Kw 10.44% 8.92% 10.16% 8.00% 8.60% 6.99% 6.51% r 2.9% 1.4% 2.7% 0.5% 1.1% -0.5% -1.0% AGL AMS BIL IMP RCH SAP SOL 100% 56% 78% 44% 49% 55% 42% 56% 100% 49% 61% 38% 46% 35% 78% 49% 100% 39% 47% 50% 43% 44% 61% 39% 100% 38% 37% 32% 49% 38% 47% 38% 100% 41% 37% 55% 46% 50% 37% 41% 100% 50% 42% 35% 43% 32% 37% 50% 100% AGL 0.098 0.066 0.078 0.053 0.053 0.052 0.041 AMS 0.066 0.143 0.059 0.088 0.050 0.052 0.041 BIL 0.078 0.059 0.103 0.048 0.053 0.048 0.042 IMP 0.053 0.088 0.048 0.146 0.050 0.043 0.037 IMP -0.8 -5.6 -0.2 11.4 -1.5 -0.6 -0.7 rfr K RCH 0.053 0.050 0.053 0.050 0.122 0.043 0.039 RCH -2.3 -0.3 -2.1 -1.5 11.9 -1.5 -1.5 7.50% 1.4 SAP 0.052 0.052 0.048 0.043 0.043 0.092 0.046 SAP -4.2 -2.0 -1.2 -0.6 -1.5 18.5 -5.0 SOL 0.041 0.041 0.042 0.037 0.039 0.046 0.095 SOL -0.5 -0.4 -2.2 -0.7 -1.5 -5.0 15.3

AGL AMS BIL 31.7 -4.4 -17.9 -4.4 13.7 -0.6 -17.9 -0.6 26.3 -0.8 -5.6 -0.2 -2.3 -0.3 -2.1 -4.2 -2.0 -1.2 -0.5 -0.4 -2.2

1( r) w 0.39119772 498% 0.03180503 41% 0.1696619 216% -0.0580698 -74% 0.01717531 22% -0.2498161-318% -0.2234732-285% 0.07848087

27

Now we go on to the situation where in addition we have three views: 1. SOL will have a return of 12% (condence 60%). 2. AGL will outperfrom BIL by 2% (condence 50%). 3. AGL, AMS, BIL, IMP will outperform SAP, SOL by 5% (condence 70%). V 12% 2% 5% c Conf 60% 50% 70% 6.49% 85% D AGL 0 1 0.52 0.11 0 0 AGL 37.4 -5.2 -21.1 -0.9 -2.7 -5.0 -0.6 AGL 10.596 0.339 -5.869 0.516 0 -1.404 -4.178 AGL 0 0 0.06 0 0.13 0 AGL -5.2 16.1 -0.7 -6.6 -0.4 -2.3 -0.5 AMS 0.339 0.040 0.216 0.061 0 -0.165 -0.490 AGL 0 -1 0.33 0 0 0.09 AGL -21.1 -0.7 31.0 -0.3 -2.5 -1.5 -2.5 BIL -5.869 0.216 8.883 0.329 0 -0.895 -2.664 w 78% 6% 19% -6% 2% -40% -9% 50% AGL AGL -0.9 -2.7 -6.6 -0.4 -0.3 -2.5 13.5 -1.8 -1.8 14.0 -0.7 -1.8 -0.9 -1.8 IMP RCH 0.516 0 0.061 0 0.329 0 0.092 0 0 0 -0.251 0 -0.747 0 w 157% 12% 37% -12% 5% -81% -17% AGL -5.0 -2.3 -1.5 -0.7 -1.8 21.8 -5.9 SAP -1.404 -0.165 -0.895 -0.251 0 0.683 2.032 AGL -0.6 -0.5 -2.5 -0.9 -1.8 -5.9 18.0 SOL -4.178 -0.490 -2.664 -0.747 0 2.032 15.299 AGL AGL 0 0 0 0 0.09 0.00 AGL 0 0 -0.25 AGL 1 0 -0.75

() 1 AGL AMS BIL IMP RCH SAP SOL P 0 1P D AGL AMS BIL IMP RCH SAP SOL

AGL AMS BIL IMP RCH SAP SOL

() 1 P 0 1V D r 60% 43% 12.36% 4.9% 7% 3% 10.51% 3.0% 39% 2% 11.70% 4.2% 11% 5% 9.42% 1.9% 25% 0% 9.88% 2.4% 6% -14% 8.18% 0.7% 17% 71% 8.49% 1.0%

28

29