Lecture 2

Andrea Buraschi

Winter 2016
Road Map

1. The Problems of the Traditional Optimal Asset Allocation Portfolio

2. A simple static solution: the 1/N Portfolio:
3. Time-Series Predictability in Asset Returns
4. Using Predictability in Optimal Portfolio Choice: A Bayesian Approach
5. The Black Litterman (i.e. Goldman Sachs model)
Review
Asset Allocation with One Asset

The optimal portfolio of one risky asset and one risk free asset solves the
following problem:
h α 2i
max U (rP ) = max E (rP ) σ
w w 2 P
where
E (rP ) = rf + w (ri rf ) and σ2P = w 2 σ2i
and α is the coe¢ cient of risk aversion.
Review
I 1 E (r i ) r f
... the solution is given by w = α σ2i
If you have N risky assets, you write the solution as
2 3
1 1 6 7 1 1
w = Σ 4E (r ) rf 1 5 = Σ µ
α | {z } α
µ

where Σ is a N N matrix, and µ = [E (r) rf 1] is a N 1 vector of expected

excess returns.
Summary

I Here are the 2 key formulas to remember:

I Tangency portfolio:
Σ 1µ
wTG =
(N 1 ) [10 Σ 1 µ ]
where Σ is a N N Covariance matrix of returns, µ is a N 1 vector
of excess returns and 1 is a N 1 vector of ones.

I Global Minimum Variance portfolio:

Σ 11
wGMV =
[10 Σ 1 1]
I The denominator, i.e. [10 Σ 1 µ] and [10 Σ 1 1], is a simple scaling factor
that just enforces the requirement that w 0 1 = 1.
Implementing Optimal Portfolios
In practice

I In practice, to obtain the portfolio weights w one needs estimates of µ

and Σ, which are unknown. Typically we follow a two-step procedure.
I Suppose an investor has T periods of observed returns data
ΦT = fR1 , R2 , ..., RT g and would like to form a portfolio for period
T + 1.
1. Step 1: Suppose that the mean and covariance matrix of the asset
returns are estimated based on historical data.
2. Step 2: Assume that sample estimates are simply plugged into the
condition to calculate optimal weights (wi = α1i Σµ). (i.e. sample
estimates are treated as if they were true parameters)
I Such a portfolion rule is called “plug-in rule”.
Estimation Risk in Optimal Allocation
Understanding Estimation Risk
Experiment: Consider Emerging Markets

The following experiment is based on DeMiguel & Nogales (in Operation

Resarch, 2009).
I Consider an economy in which (true) returns are generated by:

G = 99% N (µ, Σ) + 1% D(µd , Σ)

I The distribution D is a Deviation distribution, since 99% of the times

returns follow the Normal distribution N (µ, Σ).

0.01 0.0021 0
µ = , Σ=
0.01 0 0.0021
0.01
µd = , Σd = Σ
0.50

I We consider two EM bond assets (#1 is good; #2 less good). Only the
mean in di¤erent.
I We allow the second EM bond asset to have a large negative return with
a small 1% probability.
Understanding Estimation Risk
Experiment: Consider Emerging Markets

I Generate 240 returns Rt N (µ,Σ) and D(µd ,Σd ), i = 1, 2, ..., 240.

I Rolling windows estimation: for di¤erent values of t calculate the sample
moments:
120 +t 1
1
^
µ =
120 ∑ Ri
i =t
120 +t 1
1
Σ̂ =
120 1 ∑ (Ri ^ µ)0
µ ) ( Ri ^
i =t

I µ and Σ̂ using R1 , ..., R120 ;

Rebalancing date t = 1: Calculate ^
I µ and Σ̂ using R2 , ..., R121 ;
Rebalancing date t = 2: Calculate ^
I ...
I ... until t = 20.
I Then, assume α = 1 and compute rolling optimal portfolios.
Understanding Estimation Risk
Portfolio Instability

I In the sample there are 2 dates in which the EM bond asset #2

experience a large negative return (at t = 169 and t = 207)
Understanding Estimation Risk
Portfolio Instability: Tangency portfolio

I Notice the instability of the MVP:

I Optimal MV portfolio weight wTG should be constant:
w1,TG = 143% and w2,TG = 43%.
I Instead, MV portfolio weights w1,TG range from +200 and +450%
Understanding Estimation Risk
Portfolio Instability: Global minimum variance

I Also the GMV portfolio is unstable: Notice the increase in w1,TG

immediately after t = 169 (i.e. 49 days after t = 120)
Alternative Solutions

I Portfolio constraints
I ω TG with short-selling constraints
I Moment restrictions
I Instead of ω TG , use ω GMV
I Value-weighted portfolio
I Bayesian approach
I Bayesian di¤use prior
I Bayes-Stein estimator
I Optimal combination of portfolios
I Black Litterman
I Three-fund (Kan and Zhou): a combination of ω TG and ω GMV .
The 1/N Portfolio
A Naive Portfolio

I DeMiguel, Garlappi, and Uppal (2007) have proposed a naive portfolio

that they claim dominates (out of sample) the MV portfolio.
I The naive (“ew ” or “1/N”) strategy involves holding a portfolio weight
ω ew
t = 1 /N in each of the N risky assets. This strategy does not involve
any optimization or estimation and completely ignores the data!
The 1/N Portfolio
A Comparison with several alternatives
The Performance of the 1/N Portfolio
Predictability and Optimal Portfolio Choice
Road Map

I Return predictability: Time Series

I Can we forecast equity returns?
I Using Predictability in the Context of Optimal Portfolios
I A Bayesian Approach
I Consider the optimal portofolio choice we obtained in a simple MV
setting:
1 E t (R t + τ R f j Ωt )
ωt =
γ σ2t (Rt +τ jΩt )
I If information contained in Ωt (price/earning ratios, the slope of the term
structure, etc..) can be used to forecast future returns, then this would
a¤ect optimal portfolios
I If we can forecast ahead of time, then Et (Rt +τ Rf jΩt ) is time-varying
as a function of the value of our signals.
Return Predictability for the Market

To see if returns are predictable by a variable xt , we run regressions of

tomorrow’s return, Rt +1 , on xt .

Rt +1 = a + bxt + εt +1

If b is statistically signi…cant and R 2 is large, then there is predictability:

I I If b > 0, then buy when xt is high and positive (or sell otherwise)
I If b < 0, then buy when xt is large and negative (or sell otherwise)
Return Predictability
Market e¢ ciency

I If markets are e¢ cient, in the regressions of tomorrow’s return, Rt +1 , on

a(ny) variable we can see today, xt .

Rt +1 = a + bxt + εt +1

we expect that b = 0 and R 2 = 0, which implies:

E t (R t +1 ) = a

I Unpredictable returns means that expected return is constant over time.

Prices follow a random walk if and only if returns are unpredictable.
I Why should we believe in market e¢ ciency? If everyone sees a high xt ,
they also try to buy, driving prices up until today’s price is the same as the
price we expect tomorrow. Competition in stock markets should drive out
any predictable movement in stock prices.
I Competition means that the information in the signal xt will quickly
be impounded in today’s price
The Facts
Case 1: Do lagged returns predict future returns?

I As a …rst test, let’s test whether past returns Rt predict future returns
R t +1 :

I a positive b coe¢ cient means "momentum" (i.e. past good returns mean
higher future returns)
I a negative b means "overreaction" or "mean-reversion"
The Facts
Stocks

I If stock returns go up 100% this year, you expect a rise of just 4% next
year (b = 0.04).
I This is a trivial amount of “momentum.”
I The R 2 , measuring the proportion of return variance that can be forecast
one year ahead, is very small (R 2 = 0.002).
The Facts
Bonds

I If bond returns go up 100% this year, you expect a rise of 91% next year
I The R 2 , measuring the proportion of return variance that can be forecast
one year ahead, is very high (R 2 = 0.83).
The Facts
Stock vs. bonds
The Facts
Stock vs. bonds

I If you know stock returns will be high next year, you can borrow money
and invest in the market.
I If you know interest rates are high (T-bill returns will be high next year)
you’d have to borrow at the same high rate to invest.
I For this reason, The right way to run predictability tests is to focus on
EXCESS returns. The return you achieve in borrowing a dollar and
investing: this takes no money from agents’pocket, and thus reveals
willingness to bear risk , separating out intertemporal substitution.
The Facts
Excess returns

Notice that E (R ) = 7.25 but σ (E (R )) = 0.91. Expected excess returns are

almost constant.
Fama and French (1988): Permanent and Temporary
Components of Stock Prices

I The implied predictability of daily or weekly returns is economically insigni…cant.

I Fama and French ran the following long-horizon forecasting regression:

R t,t +T = b 0,T + b 1,T R t T ,t + εt,t +T

I Some statistical they excplicitly address (explained later):

I Finite sample bias in AR(1) regressions - coe¢ cients negatively biased.

I OLS standard errors are wrong because of overlapping observations.
I None-the-less robust standard errors are computed and Monte Carlo
simulations are used to correct unbiased coe¢ cents.
Fama and French (1988): results

I Bias adjusted slopes reach minimum for 2 - 5 years: U-shaped pattern.

I The U-shaped pattern is consistent with the hypothesis that stock prices contain
both random walk (short term) and slowly decaying stationary components (5
years mean reversion).
I Predictable variation amounts to 40% for small …rms while it drops to 25% for
large …rms (3 5 year returns).
Fama and French (1988): Permanent and Temporary
Components of Stock Prices

I There is evidence of long-run reversals at longer (3-5 years) horizons

I Two stories for long horizon predictability:

I Models of irrational markets in which stock prices take long temporary

swings away from fundamental values.
I Time-varying equilibrium expected returns generated by rational pricing in
an e¢ cient market.
I Both stories imply observationally equivalent price behaviour.
I If expected returns are stationary and vary over time stock prices should exhibit
mean-version.
Intuition : holding expected future dividends constant, only way to give investors
higher expected future returns is for prices to contemporaneously fall.
Predictability using accounting and macro variables
Case 2: Do dividend yields predict future returns?
Predictability using accounting and macro variables
Case 3: Do Cons/Income and Detrended short rate predict future returns?
Returns and the Dividend/Price ratio
Returns and Dividends
Using Information from Predictability: The
Black-Litterman Approach
Using Information from Predictability

I Suppose you want to use both historical information ^

µ about returns and
contemporaneous information that comes from a dynamic model (signal
υt )
I Both sources of information are assumed to be uncertain and are
expressed in terms of probability distributions.
A Simple Bayesian Approach
5 steps

1. Compute Σ̂ based on historical returns

2. Use the CAPM-implied expected excess returns ^
µ, as a
prior : E (R Rf ) = β[E (Rm Rf )]
3. Express views, expressed via a signal vector υt . This could be a valuation
model based on P /D ratios or other subjective views
4. Merge optimally the prior ^
µ and the signal υt to form a posterior using
Bayes Theorem.
5. Use the posterior beliefs to form optimal portfolio weights.
A Simple Bayesian Approach
Step 2

I The …rst two steps give us an "initial estimate" or a set of prior beliefs
about the distribution of returns.
I Suppose our con…dence in the CAPM model is not perfect, and it is
parametrized by the coe¢ cient τ .

Re µ, τ Σ̂)
N (^

I So that in the special case of 1 asset case, we have

p
Rje = ^ µj + τσj εj εj N (0, 1)
E (Rje ) = ^
µj

Higher velues of τ express lower con…dence in CAPM.

I A commonly assumed value for τ is the neutral value of 1. But if you are
even less con…dent, then τ > 1.
A Simple Bayesian Approach
Step 3

I µ = 40%; The naive

Intuition: Suppose that, based on the last 5 years, ^
approach implies that E (Rje ) = 40%
I However, you also observe that P /E = 90. Would you revise your
expectations? How much?
A Simple Bayesian Approach
Step 3

I We need to introduce the P /E signal into the model. How? This is

Black-Litterman.
I The model forecast can be expressed as a (correlated) signal υj about the
expected return of asset j

υj = Rje + σεj εjυ

where εjυ N (0, 1), with εj and εjυ uncorrelated.

I Stack into a vector:
p
Rje µj τσj εj
= +
υj Rje σεj εjυ

where ε N (0, 1).

I Notice that since E (υj ) = E (Rje ) the evolution of the signal and of Rje are
not independent on average.
Remember Bayesian Theorem?
Step 4

Theorem
If a multidimensional variable (X, Y ) is jointly normally distributed N (µ, Σ)
where
µX ΣXX ΣXY
µ= Σ=
µY ΣYX ΣYY
then

(XjY = y ) N [µX + ΣXY ΣYY1 (y µY ) ; ΣXX ΣXY ΣYY1 ΣXY ]

Apply Bayes Theorem in Our Context (Kalman Bucy)
Step 4

Corollary
If the prior is based on the implied returns Re µ, τ Σ̂) and investor’s views
N (^
are given by a signal vector υ N (^ µ, τ Σ̂ + Ω), then the posterior is given by

Re j υ µPOST , Σ̂POST )
N (^

where
0 1

1 1 B 1 C
1 1 B 1 C
^
µPOST = Ω + Σ̂ B Ω υ + Σ̂ 1 ^
µ C
τ @ | {z } τ
| {z } A
Tactical Part
Strategic Part

Ω 1
= µ+
^ (υ ^µ)
Ω + τ1 Σ̂
1 1 | {z }
Distance View from Priors
1
1
Σ̂POST = Ω 1 + Σ̂ 1
.
τ

I Recall that Ω is the diagonal matrix with the variance σ2εj as elements.
... Now the Optimal Portfolio
Step 5

I The new portfolio is therefore:

1
w = Σ̂ 1 ^
µ =
γ POST POST
A Simple Bayesian Approach
Example - 1 asset

I Suppose you are a mean variance investor and have £ 100 million to
allocate between a U.S. Treasury risk-free bond and Turkish bonds.
I The risk free bond yields a return of rf = 3.5%.
I Historical data on Turkish bond returns are normally distributed,
with standard deviation σG = 20% .
I The CAPM implied expected return of Turkish bonds is
E (rG ) = 10%.
I Your experience in portfolio management makes you (dis)trust CAPM,
with a factor τ = 1.2. A larger τ proxy for more distrust.
I Your coe¢ cient of risk-aversion is 2.
A Simple Bayesian Approach
Example - 1 asset

Which fraction of your wealth do you place in the U.S. Treasuries and in
Turkish bonds?
I The optimal fraction of the wealth placed in the Turkish bonds satis…es

E (rG ) rf
w =
γ τσ2G
0.10 0.035
w = = 0.677
2 (1.2 0.04)
A Simple Bayesian Approach
Example - 1 asset

Suppose that your understanding of recent market developments tells you that
probably the expected return of Turkish bonds will be 2% lower than the
market expects. While you are usually right, the standard deviation of your
predictions is 25%. Would you adjust your portfolio? How?
I Your signal can be represented as

υ = rG + 0.1ε
where ε N (0, 1).
I In this case your posterior beliefs are normally distributed with:
(σ2v ) 1
E (rG j υ ) = E (rG ) + [υ E (rG )]
(σ2v ) 1 + (τσ2g ) 1

1/(0.252 )
= 0.10 + (0.08 0.10) = 9.13%
1/(0.252 ) + 1/(1.2 0.202 )
and
1
1 1
σ2G jυ = Ω 1
+ Σ̂ 1
= = 2.715%
τ 1/(0.252 ) + 1/(1.2 0.202 )
I It compars to τσ2G = 1.2 0.202 = 0.048. The signal is reducing both the
(posterior) Expected and Variance of the Returns if you believe your views
A Simple Bayesian Approach
Example - 1 asset

I The fraction of wealth invested in Turkish bonds is now

9.13 3.5
100 100
w = 2.715
= 1.0368
2 100

I You are investing more because even if the signal for Turkish Treasuries is
less than before (9.13% instead than 10%), the signal plus the CAPM
increase your con…dence in a large positive return.
 Appendix:
Extensions not covered in class
 What about
Predictability on Relative Returns?
Now the Most General Case
Mix Absolute and Relative weights

Alternativley, views can be expressed as follows

υ = P T Re + Ωευ

where
I Re is the vector of excess returns for the N assets
I P is N k matrix of view weights
I υ is a vector of k views
I ευ is a vector of k error terms associated with the views (εjυ N (0, 1))
I Ω represents a diagonal k k matrix with elements σεj (i.e. the views are
independent).
Now the Most General Case
Mix Absolute and relative weights

Example - 6 assets. Consider the following views

1. The investor is sure that the expected return of share 1 will amount to 20
%.
2. With a probability of 70 %, the investor believes that the di¤erence of the
expected returns between an equally weighted portfolio of share 1 and
share 3 and an equally weighted portfolio of share 4, 5 and 6 will amount
to 6 %.
Now the Most General Case
Mix Absolute and relative weights

Example - 6 assets. The views can be represented as

0 1
R1e
B R2e C
B C
0.2 1 0 0 0 0 0 B R3e C ε1υ
= B C+Ω
0.06 0.5 0 0.5 0.33 0.33 0.33 B R4e C ε2υ
B C
@ R5e A
R6e

where
0 0
Ω= 5
0 9.317 10
Now the Most General Case
Mix Absolute and relative weights

Example - 6 assets.

I View 1 is speci…ed in absolute terms: the …rst entry (for the …rst stock) of
the …rst view row in matrix P T contains the value 1.

I View 2 is a relative view. Since the relative comparison is related to

portfolios between stocks 1, 3 and stocks 4, 5 and 6, the respective
column contains 1/2 for stocks 1 and 3 and 1/3 for stocks 4, 5 and 6.

I The sum of the weights in the absolute view has to be 1.

I The sum of the weights in the relative view has to be 0.
I Note: the investor does not need to have views on all assets.
Now the Most General Case
Mix Absolute and relative weights

When views are expressed as

υ = P T Re + Ωευ

then
υ µ, τ P T Σ̂P + Ω)
N (P T ^
Now the Most General Case
Mix Absolute and relative weights

Corollary
If the prior is based on the implied returns Re µ, τ Σ̂) and investor’s views
N (^
are given by a signal vector υ N (P T ^ µ, τ P T Σ̂P + Ω), then the posterior is
given by
Re jυ N (^ µPOST , Σ̂POST )
where
1
1 1 1
^
µPOST = PΩ PT + Σ̂ 1
PΩ 1
υ+ Σ̂ 1
^
µ
τ τ
1
1 1
Σ̂POST = PΩ PT + Σ̂ 1
.
τ

Note: If P = IN (P is the identity matrix), then we obtain the same expression

as in the previous Corrolary.
Now the Most General Case
Mix Absolute and relative weights

I If the uncertainty about own beliefs converges to in…nity, the investor only
trusts in the equilibrium expected returns

lim E (Re jυ) = ^

µ
Ω!∞

I If the uncertainty about expected returns converges to zero, the investor

only trusts in the equilibrium expected returns

lim E (Re jυ) = ^

µ
τ !0
Now the Most General Case
Mix Absolute and relative weights

The variance of the views is inversely related to the investors con…dence in the
views. It is up to the investor to compute the variance of the views Ω.
I Proportional to the variance of the prior
I Use a con…dence interval
I Use the variance of residuals in a factor model
The Bayesian Approach
Specifying the views

The variance of the views is proportional to the variance of the prior

Ω = diag P (τΣ) P T

I This seems to be the most common method used in the

literature/practice.
I This speci…cation of the variance, or uncertainty, of the views essentially
equally weights the investor’s views and the market equilibrium weights.
I By including τ in the expression, the …nal solution becomes independent
of τ as well.
The Bayesian Approach
Specifying the views

De…ne a con…dence interval around the estimated mean return

I Asset 2 has an estimated 3% mean return with the expectation it is 68%
likely to be within the interval (2.0%, 4.0%).
I Knowing that 68% of the normal distribution falls within 1 standard
deviation of the mean, allows us to translate this into a
variance for the view of (1%)2 .
The Bayesian Approach
Specifying the views

Use the variance of residuals from a factor model

n
E (R e ) = ∑ β i fi + ε
i =1

where βi is the factor loading for factor i , fi is the return due to factor and ε is
an independent normally distributed residual.

The general expression for the variance of the return from a factor model is

B V (F ) B T + V ( ε )

where B is the factor loading matrix and F is the vector of returns due to the
various factors.

While the regression might yield a full covariance matrix, the BL model will be
more robust if only the diagonal elements are used.
 Appendix 2:
Kalman-Bucy for N Assets
N-Assets Extension
Example - 2 assets

Suppose you are a mean variance investor and have £ 100 million to allocate
between two risky assets, A and B .
I σA = 15% and σB = 25% and ρAB = 0.2.
I the CAPM implied returns are E (rA ) = 20% and E (rB ) = 40%.
I the coe¢ cient of risk-aversion is 1.5.
N-Assets Extension
Example - 2 assets

Suppose your views are di¤erent than the market, and you think that the return
of asset A is going to be 24% while the return of asset B is 36%. You are not
entirely sure about this, such that the standard deviation of your predictions is
20% for both A and B .

υA = rA + 0.2εA
υB = rB + 0.2εB
N-Assets Extension
Example - 2 assets

At this step we need to update our initial believes, conditional on the signal we
have. The posterior opinions are given by
1
E (rA j υ ) 1 1 1 1 υA 1 1 E (rA )
= Ω + Σ Ω + Σ
E (rB j υ ) τ υB τ E (rB )

and
σ2A jυ 1 1 1
1
= Ω + Σ
σ2B jυ ) τ

where
υA 0.24
=
υB 0.36

σ2A ρσA σB 0.152 0.2 0.15 0.25

Σ = =
ρσA σB σ2B 0.2 0.15 0.25 0.252
0.22 0
Ω =
0 0.22
N-Assets Extension
Example

Calculate the new weights using the posterior mean and variance covariance
matrix
Dealing with Corner Solutions in Mean-Variance Portfolios
Suppose our estimates of the expected returns are given by
Mean Variance Portfolios

Then...

I Emergence of corner solution

I Too low the weight to emerging markets
 Appendix 3:
Using Information from Market Clearing
Solution 1: Impose Equilibrium Information
Markets Must Clear!

In equilibrium, markets clear such that

Z
wi0 = wS
i

where wS is a vector of outstanding value of shares for the risky assets. This
implies: Z
1
Σ 1 µ = wS
i αi
CAPM Equilibrium Returns

Since in equilibrium excess returns must satisfy

1
µ= Σw S
γ
R
where γ = α1i is the aggregate (average) inverse-risk aversion of market
participants. The higher the average risk-aversion, the higher the expected
return
CAPM Equilibrium Returns

e
In aggregate we should hold the market Rm,t +1 . Thus, with Mean-Variance
investors the CAPM relation must be satis…ed:

E (R i ) R f = βi [E (R m ) Rf ]

This implies that given (a) βi and (b) the market risk premium [E (Rm ) R f ],
we can recover equilibrium E (Ri ) Rr = µ
Then, we can plug-in into the usual solution

1 1
w = Σ µ
α
Proof of the CAPM from …rst principles..

Notice that:

e wS0 Ret+1
Rm,t +1 =
wS0 1N
which has an expected return

wS0 γΣwS
µm = .
wS0 1N

Then

µ = βµm
cov (µ, µm )
where β =
var (µm )
How does it look like in practice?
The optimal portfolio is
How does it look like in practice?
The optimal portfolio is