INSTITUTIONAL FINANCE

Lecture 05: Portfolio Choice, CAPM, Black-Litterman

OVERVIEW

1. Portfolio Theory in a Mean-Variance world

2. Capital Asset Pricing Model (CAPM)
3. Estimating Mean and CoVariance matrix
4. Black-Litterman Model
 Takinga view
 Bayesian Updating
EXPECTED RETURNS & VARIANCE

 Expected returns (linear)

j
¹ p := E [r p ] = w j ¹ j , w h ere each w j = Ph
hj

Variance
j


µ ¶ µ ¶
¾ 12 ¾ 12 w1
¾ p2 := V a r [r p ] = w 0 V w = (w 1 w 2 )
¾ 21 ¾ 22 w2
µ ¶
w1
= (w 1 ¾ 12 + w 2 ¾ 2 1 w 1 ¾ 1 2 + w 2 ¾ 22 )
w2
= w 12 ¾ 12 + w 22 ¾ 22 + 2 w 1 w 2 ¾ 1 2 ¸ 0
s in ce ¾ 1 2 · ¡ ¾ 1 ¾ 2 : recall that correlation
coefficient 2 [-1,1]
ILLUSTRATION OF 2 ASSET CASE

 For certain weights: w1 and (1-w1)

mp = w1 E[r1]+ (1-w1) E[r2]
s2p = w12 s12 + (1-w1)2 s22 + 2 w1(1-w1)s1 s2 r1,2
(Specify s2p and one gets weights and mp’s)
 Special cases [w1 to obtain certain sR]
 r1,2 = 1 ) w1 = (+/-sp – s2) / (s1 – s2)
 r1,2 = -1 ) w1 = (+/- sp + s2) / (s1 + s2)
2 ASSETS ½ = 1
¾p = jw 1 ¾ 1 + (1 ¡ w 1 )¾ 2 j
Hence, w 1 = §¾p ¡¾2
¾1 ¡ ¾2
¹p = w 1 ¹ 1 + (1 ¡ w 1 )¹ 2

¹p = ¹1 + ¹¾22¡¹
¡¾1 (§ ¾p ¡ ¾1 )
1

E[r2]
mp

E[r1]

s1 sp s2

Lower part with … is irrelevant

¹p = E[r1] + E[r¾2]¡E[r
¡¾
1] (¡¾ ¡ ¾ )
R 1
2 1
The Efficient Frontier: Two Perfectly Correlated Risky Assets
2 ASSETS ½ = -1
For r1,2 = -1: ¾p = jw 1 ¾ 1 ¡ (1 ¡ w 1 )¾ 2 j Hence, w 1 = §¾p +¾2
¾1 + ¾2
¹p = w 1 ¹ 1 + (1 ¡ w 1 )¹ 2

¹2¡¹1
¹p = ¾ ¾+¾
2 ¹ +
1
¾1
¾1+¾2 § ¾1+¾2 ¾p
1 2

E[r2]
¹2 ¡¹1
slop e: ¾1 + ¾2
¾p
¾2 ¾1
¹
¾ 1 +¾ 2 1
+ ¹
¾ 1 +¾ 2 2 ¹2 ¡¹1
slop e: ¡ ¾p
E[r1] ¾1 + ¾2

s1 s2

Efficient Frontier: Two Perfectly Negative Correlated Risky Assets

2 ASSETS -1 <½ < 1

E[r2]

E[r1]

s1 s2

Efficient Frontier: Two Imperfectly Correlated Risky Assets

2 ASSETS ¾1 = 0

E[r2]
mp

E[r1]

s1 sp s2

The Efficient Frontier: One Risky and One Risk Free Asset
EFFICIENT FRONTIER WITH N RISKY ASSETS

 A frontier portfolio is one which displays minimum variance

among all feasible portfolios with the same expected portfolio
return.
E[r]

1 T
min w Vw
w 2

s
  s.t. w e  E  ~r )  E 
T N
 w E ( 
 i 1 
i i

 N

  w T1  1 
 i 1
w i  1 

 @L = V w ¡ ¸e ¡ °1 = 0
@w
@L = E ¡ wT e = 0
@¸
@L = 1 ¡ wT 1 = 0
@°
The first FOC can be written as:

V wp = ¸e + °1 or
¡1
wp = ¸V e + °V 1 ¡1
T T ¡1 T ¡1
e wp = ¸(e V e) + °(e V 1)
Noting that eT wp = wTpe, using the first foc, the second foc
can be written as

rp] = eT wp = ¸ |(eT V{z¡1e)} +° |(eT V{z¡11)}

E[~
:=B =:A
pre-multiplying first foc with 1 (instead of eT) yields

1T wp = wpT 1 = ¸(1T V ¡1e) + °(1T V ¡11) = 1

1 = ¸ |(1T V{z¡1 e)} +° |(1T V{z¡11)}
=:A =:C
Solving both equations for  and 

CE¡A
¸ = D and ° = D B¡AE

where D = BC ¡ A2.
Hence, wp =  V-1e +  V-11 becomes

CE  A 1 B  AE 1
wp  V e V 1
D (vector) D (vector)
 scalar)  scalar)


1
D
 1
D
 
BV 11  AV 1e   CV 1e   AV 11 E 
(6.15)
wp  g  h E
linear in expected return E!
(vector) (vector) (scalar)

If E = 0, wp = g
Hence, g and g+h are portfolios
If E = 1, wp = g + h
on the frontier.
EFFICIENT FRONTIER WITH RISK-FREE ASSET

The Efficient Frontier: One Risk Free and n Risky Assets

EFFICIENT FRONTIER WITH RISK-FREE ASSET

minw 12 w TV w

s.t. wT e + (1 ¡ wT 1)rf = E[rp]

FOC: wp = ¸V ¡1(e ¡ rf 1)
E[rp]¡rf
Multiplying by (e–rf 1)T and solving for  yields ¸ =
(e¡rf 1)T V ¡1(e¡rf 1)

¡1 E[rp ]¡rf
wp = V (e ¡ rf 1) H2
| {z }
n£1
q
where H = B ¡ 2Arf + Crf2 is a number
EFFICIENT FRONTIER WITH RISK-FREE ASSET
 Result 1: Excess return in frontier excess return
Cov[rq ; rp ] = wqT V wp
E[rp ] ¡ rf
= wqT (e ¡ rf 1)
| {z } H2
E[rq ]¡rf

(E[rq ] ¡ rf )([E[rp ] ¡ rf )
=
H2
(E[rp ] ¡ rf )2
V ar[rp ; rp ] =
H2
Cov[rq ; rp]
E[rq ] ¡ rf = (E[rp] ¡ rf )
V ar[rp]
| {z }
:=¯q;p
Holds for any frontier portfolio p, in particular the market portfolio!
EFFICIENT FRONTIER WITH RISK-FREE ASSET

 Result 2: Frontier is linear in (E[r], ¾)-space

(E[rp ] ¡ rf )2
V ar[rp ; rp ] =
H2
E[rp ] = rf + H¾p

E[rp ]¡rf
H= ¾p
where H is the Sharpe ratio
TWO FUND SEPARATION

 Doing it in two steps:

 First
solve frontier for n risky asset
 Then solve tangency point

 Advantage:
 Same portfolio of n risky asset for different agents
with different risk aversion
 Useful for applying equilibrium argument (later)
TWO FUND SEPARATION

Price of Risk =
= highest
Sharpe ratio

Optimal Portfolios of Two Investors with Different Risk Aversion

MEAN-VARIANCE PREFERENCES
 U(mp, sp) with @U
@ ¹p
> 0, @U
2
@ ¾p
< 0

 Example: E[W ] ¡ °
2V ar[W ]
 Also in expected utility framework
 quadratic utility function (with portfolio return R)
U(R) = a + b R + c R2
vNM: E[U(R)] = a + b E[R] + c E[R2]
= a + b mp + c mp2 + c sp2
= g(mp, sp)

 asset returns normally distributed ) R=j wj rj normal

 if U(.) is CARA ) certainty equivalent = mp - rA/2s2p
(Use moment generating function)
OVERVIEW

1. Portfolio Theory in a Mean-Variance world

2. Capital Asset Pricing Model (CAPM)
3. Estimating Mean and CoVariance matrix
4. Black-Litterman Model
 Takinga view
 Bayesian Updating
2. EQUILIBRIUM LEADS TO CAPM

 Portfolio theory: only analysis of demand

 price/returns are taken as given
 composition of risky portfolio is same for all investors

 Equilibrium Demand = Supply (market portfolio)

 CAPM allows to derive

 equilibrium prices/ returns.
 risk-premium
THE CAPM WITH A RISK-FREE BOND

 The market portfolio is efficient since it is on

the efficient frontier.
 All individual optimal portfolios are located on
the half-line originating at point (0,r f).
E[ RM ]  R f
 The slope of Capital Market Line (CML): sM .

E[ RM ]  R f
E[ R p ]  R f  sp
sM
CAPITAL MARKET LINE

CML

M
rM

rf
j

sM sp
SECURITY MARKET LINE
E(r)

SML
E(ri)

E(rM)

rf
slope SML = (E(ri)-rf) /b i

b M 1 bi b

Cov[rj ; rM ]
E[rj ] = ¹j = rf + (E[rM ] ¡ rf )
V ar[rM ]
| {z }
¯j
OVERVIEW

1. Portfolio Theory in a Mean-Variance world

2. Capital Asset Pricing Model (CAPM)
3. Estimating Mean and CoVariance matrix
4. Black-Litterman Model
 Takinga view
 Bayesian Updating
3. ESTIMATING MEAN AND CO-VARIANCE
 Consider returns as stochastic process (e.g. GBM)
 Mean return (drift)
 For any partition of [0,T] with N points (∆t=T/N),
N*E[r] =Ni=1 ri∆t= pT -p0 (in log prices)
 Knowing first p0 and last price pT is sufficient
 Estimation is very imprecise!
 Variance
 Var[r]=1/N Ni=1 (ri∆t-E[r])2 → σ2 as N→∞
 Theory: Intermediate points help to estimate co-
variance
 Real world:
 time-varying
 Market microstructure noise
3. 1000 ASSETS

 Invert a 1000x1000 matrix

 Estimate 1000 expected returns

 Estimate 1000 variances

 Estimate 1000*1001/2 – 1000 co-variances

Reduce to fewer factors

… so far we used past data

(and assumed future will behave the same)
OVERVIEW

1. Portfolio Theory in a Mean-Variance world

2. Capital Asset Pricing Model (CAPM)
3. Estimating Mean and CoVariance matrix
4. Black-Litterman Model
 Takinga view
 Bayesian Updating
4. BLACK-LITTERMAN MODEL

 So far we estimated expected returns using

historical data.
 We ignored statistical priors:
 A sector with an unusually high (or low) past return
was assumed to earn (on average) the same high (or
low) return going forward.
 We should have attributed some of this past return to
luck, and only some to the sector being unusual
relative to the population.
EXPECTED RETURNS

 We also ignored economic priors:

 A sector with a negative past return should not be
expected to have negative expected returns going
forward.
 A sector that is highly correlated with another sector
should probably have similar expected returns.
 A “good deal” in the past (i.e. good realized return
relative to risk) should not persist if everyone is
applying mean-variance optimization.
 What is a good starting point from which to
update based on our analysis?
BAYESIAN UPDATING
 Bayes’ Rule allows one to update distribution after
observing some signal/data
 from prior to posterior distribution
 Recall if all variables are normally distributed with can use
the projection theorem
 E.g. prior: µ = N (¹,¿ 2); signal/view x = µ + ², where ² = N(0, ¾ 2)
 Weights depend on relative precision/confidence of prior vs.
signal/view (on portfolio)

.  s2   t2 

E ( | x)   2  m   2
2 
x
2 
t s  t s 
BLACK LITTERMAN PRIOR

 All expected returns are in proportion to their

risk.
 Expected returns are distributed around
bi (E[Rm] – Rf)
PROPERTIES OF A CAPM PRIOR
 All expected returns are in proportion to their
risk.
 Expected returns are distributed around
bi (E[Rm] – Rf)

 Is this a good starting point?

 We can still use optimization
 We don’t throw out data (e.g. still can estimate
covariance structure accurately)
 It is internally consistent – if we don’t have an
edge, the prior will lead us to holding the market
BLACK-LITTERMAN
 The Black-Litterman model simply takes the
starting point that there are no good deals…
 And then adjusts returns according to any
“views” that the investor has from:
 Seeing abnormal returns in the past that
expected to persist (or reverse)
 Fundamental analysis
 Alphas of active trading strategies
 “views” concern portfolios and not necessarily
individual assets
BLACK LITTERMAN PRIORS – MORE SPECIFIC
See He and Litterman

 Suppose returns of N-assets (in vector/matrix notation)

r » N(¹; §)
 Equilibrium risk premium,
eq
¦ = °§w
where ° risk aversion, weq market portfolio weights
 Bayesian prior (with imprecision)
¹ = ¦ + "0 , where "0 » N(0; ¿§)
VIEWS

 View on a single asset affects many weights

 “Portfolios views”
 views on K portfolios
 P: K x N-matrix with portfolio weights

 Q: K-vector of expected returns on these portfolios

 Investor’s views
P ¹ = Q + "v , where "v » N(0; -)
  is a off-diagonal values are all zero
 "v and "0are all orthogonal
BAYESIAN POSTERIOR - REWRITTEN

 s2   t2 
E ( | x)   2  m   2
2 
x
2 
t s  t s 
 1t2   1s2 
  2  m   2
2 
x
2 
1 t 1 s  1 t 1 s 

1
1 t 1 s
2 2
1 
t 2
 m  1 s 2
x 
BAYESIAN UPDATING
 Black Litterman updates returns to reflect views using
Bayes’ Rule.
 The updating formula is just the multi-variate (matrix)
version of

E ( | x) 
1
1 t 1 s
2 2
2

1 t  m 1 s  x
2



E[ R | Q]  t  P  P
1 T 1
 t
1 1
  PT 1Q 
BAYESIAN UPDATING

E ( | x) 
1
1 t 1 s
2 2
2 2

1 t  m 1 s  x 


E[ R | Q]  t  P  P
1 T 1
 t
1 1
P  Q
T 1

Scaling term – Total precision
BAYESIAN UPDATING

E ( | x) 
1
1 t 1 s
2 2

1 t  m 1 s  x
2 2



E[ R | Q]  t  P  P
1 T 1
 t
1 1
P  Q T 1

CAPM Prior
expected returns
BAYESIAN UPDATING IN BLACK
LITTERMAN
E ( | x) 
1
1 t 1 s
2 2

1 t  m 1 s  x
2 2



E[ R | Q]  t  P  P
1 T 1
 t
1 1
P  Q
T 1

Weighted by
precision of
CAPM Prior
BAYESIAN UPDATING

E ( | x) 
1
1 t 1 s
2 2

1 t  m 1 s  x
2 2



E[ R | Q]  t  P  P
1 T 1
 t
1 1
P  QT 1

Vector of expected
return views
BAYESIAN UPDATING

E ( | x) 
1
1 t 1 s
2 2

1 t  m 1 s  x
2 2



E[ R | Q]  t  P  P
1 T 1
 t
1 1
P  QT 1

Weighted by
precision of views
ADVANTAGES OF BLACK-LITTERMAN

 Returns are only adjusted partially towards the investor’s

views using Bayesian updating
 Recognizes that views may be due to estimation error
 Only highly precise/confident views are weighted heavily
 Returns are modified in a way that is consistent with
economic priors
 highly correlated sectors have returns modified in the same way
 Returns can be modified to reflect absolute or relative views
 The resulting weights are reasonable and do not load up on
estimation error
OVERVIEW

1. Portfolio Theory in a Mean-Variance world

2. Capital Asset Pricing Model (CAPM)
3. Estimating Mean and CoVariance matrix
4. Black-Litterman Model
 Takinga view
 Bayesian Updating