Factor Models

Fernando Chague

2019
Beta Pricing Models

Definition: A single-factor beta pricing model exists if there is a factor f˜ and

constants Rz and λ such that for each R̃ we have
 
h i cov f˜, R̃
E R̃ = Rz + λ  
var f˜

Remarks:
cov (f˜,R̃ )
βf = var (f˜)
is the beta of R̃ (factor loading, “quantity” of risk)
λ is the factor risk premium (the price of a unit of beta-risk)
 
Rz is the return on a cov f˜, R˜z = 0 asset; if ∃ Rf , then Rz = Rf
Beta Pricing Models

Remarks:
cov (R̃ ∗ ,R̃ ∗ )
h i
1. If the factor is also a return R̃ ∗ = f˜, then E R̃ ∗ = Rz + λ var R̃ ∗
( )
 
h i cov R̃ ∗ , R̃  h i 
E R̃ = Rz +   E R̃ ∗ − Rz
var R̃ ∗

2. One can write a beta pricing model in terms of covariances:

h i  
E R̃ = Rz + ψcov f˜, R̃
h i
where ψ = λ /var f˜
Beta Pricing Models

Remarks:
3. Assume a single-factor beta pricing exists and let Rz 6= 0. Then, the following
random variable is a Stochastic Discount Factor (SDF):

1 ψ ˜ h i
m̃ = − f − E f˜
Rz Rz

To see this, note that for any R̃:

 h i 
h i 1 ψ ˜
E m̃R̃ = E − f − E f˜ R̃
Rz Rz
1  h i h h i i
= E R̃ − ψE f˜ − E f˜ R̃
Rz
1  h i  
= E R̃ − ψcov f˜, R̃
Rz
= 1
Capital Asset Pricing Model (CAPM)
Theorem
Suppose either:
i) Returns are jointly normal
ii) all investors have quadratic utility
iii) two-fund separation holds
Then in equilibrium the market portfolio (“aggregate wealth portfolio”) is on the
MV efficient frontier and the previous theorem imply that for any portfolio q

E [Rq ] = Rf + βqm (E [Rm ] − Rf )

with E [Rm ] > Rf

Intuition:
I Under i), ii), or iii), MV analysis is valid
I Under MV analysis, each investor hold a portfolio on the MV frontier
I Each of these portfolios is a linear combination of two frontier portfolios
I The market portfolio is a convex combination of all investor’s portfolios and,
by the convexity of the efficient set, is also a frontier portfolio
Capital Market Line

Theorem
Suppose m is on the MV efficient frontier and q is another MV efficient frontier
portfolio, then:
E [Rm ] − Rf
E [Rq ] = Rf + σq
σm
which defines the Capital Market Line (CML).

Remarks:
E [Rq ]−Rf E [Rm ]−Rf
I Since q is MV efficient, σq = σm
E [Rm ]−Rf
I
σm is also known as the Sharpe Ratio (SR)
I The inclination of the CML is the SR (the maximum SR among all possible
portfolios)
I The market portfolio is the tangency portfolio with only risky assets
I The Security Market Line (SML) has (µ, β ) as coordinates
CAPM Derivations

Suppose there are i = 1, 2, ..., I investors and j = 1, 2, ..., N assets

Let πij denote the ith investor’s proportional investment in asset j
The terminal wealth of investor i is:
" #
N  
w̃i,1 = wi,0 1 + Rf + ∑ πi,j R̃j − Rf
j =1

The terminal value of the market portfolio is:

I  
∑ w̃i,1 = w̃m,1 = wm,0 1 + R̃m
i =1

where
I N
wm,0 = ∑ wi,0 and R̃m = ∑ πm,j R̃j
i =1 j =1
CAPM Derivations

Market clearing implies:

I I
wi,0
∑ πi,j wi,0 = πm,j wm,0 ∀j and πm,j = ∑ πi,j wm,0
i =1 i =1

Finally, the investor’s FOC are:

h i h   i
Cov ui0 (w̃i,1 ) , R̃j = −E u 0 (w̃i,1 ) E R̃j − Rf

CAPM Derivations
Method 1: Assume joint normality
I Using Stein’s lemma on the LHS:

E ui00 (w1i ) Cov w1i , Rj = −E ui0 (w1i ) E Rj − Rf

     


or
E u 0 (w1i ) 
 
Cov w1i , Rj = −  00i
 

 E Rj − Rf
E ui (w1i )
 00   0 
I Define θi = −E ui (w1i ) /E ui (w1i ) and aggregate over investors:
!
l 
 l    
∑ 1/θi E Rj − Rf = ∑ Cov w1i , Rj = Cov w1m , Rj
i=1 i=1
   
= Cov w0m (1 + Rm ) , Rj = w0m Cov Rm , Rj

I For j = m, we have that:

!
l
∑ 1/θi [E (Rm ) − Rf ] = w0m Var [Rm ]
i=1

I Dividing the last two equations yields the CAPM

CAPM Derivations

Method 2: Assume quadratic utility

With quadratic utility ui0 wi,1 = ai − bi wi,1 . Substituting into the FOCs:


I

 
 
   
ai − bi E wi,1 E Rj − Rf = −Cov ai − bi wi,1 , Rj = bi Cov wi,1 , Rj
I Divide by bi and aggregate over investors
!
I
ai    
 
∑ bi − E wm,1 E Rj − Rf = −wm,0 Cov Rm , Rj
i=1

I For j = m, we have:
!
I
ai  
∑ − E wm,1 (E [Rm ] − Rf ) = −wm,0 Var [Rm ]
i=1 bi

I Dividing the last two equations yields the CAPM

multi-factor models
Beta Pricing Models

Definition: A multi-factor beta pricing model exists if there are factors

F = (f1 , ..., fk )0 , a constant Rz and a k × 1 vector λ such that for each R we have

E [R] = Rz + λ 0 Σ−1
F cov (F , R)

Remarks:
I β = Σ−1F cov (F , R) is vector of betas (“factor loadings”, quantity of risky)
I λ is the vector of factor risk premia (the price of a unit of beta-risk)
I Rz is the return on a Σ−1F cov (F , R) = 0 asset; if ∃ Rf , then Rz = Rf
Beta Pricing Models

Remarks:
1. One can write a beta pricing model in terms of covariances:

E [R] = Rz + ψ 0 cov (F , R)

where ψ 0 = Σ−1
F λ
2. We can transform factors to have zero means, unit variances and to be
mutually uncorrelated:
E [R] = Rz + ψ 0 cov (G, R)
where ψ 0 = L−1 λ and G is the Gram-Schmidt orthogonalization of F (that is
G = L−1 (F − E [F ]), where L is the Cholesky decomposition of ΣF )
3. The number of factors is not uniquely determined; e.g. we can combine factors
in one single factor λ 0 Σ−1
F F
APT

(Different but) Factor-Related Concepts:

I Stochastic Discount Factor
I Factors in Beta Pricing Model
I Factor Model (or factor structure of returns)

Arbitrage Pricing Theory (Ross, 1977)

I Derived under the assumption that assets have a factor structure (potentially a
multiple factor structure)
APT

Project Ri on k random variables F = (f1 , ..., fk )0 and a constant

Ri = E [Ri ] + Cov (F , Ri )0 Σ−1 ˜

F (F − E [F ]) + εi

where E [εi ] = 0 and cov (fj , εi ) = 0 for j = 1, ..., k.

Remarks:
I Return variation is decomposed into a common and an asset-specific
component
I Two potential source of risk:
F Systematic risk (comes from common component)
F Idiosyncratic risk (is asset specific)
I Intuition: If idiosyncratic shocks can be diversified away, by arbitrage, they
should not be priced
APT

Definition: Returns have a factor structure with f1 , ..., fk as factors if

cov (εi , ε` ) = 0

for i, ` = 1, ..., n and i 6= `.

Portfolio diversification is an investment strategy that makes idiosyncratic
risk disappear
Let Rp = n1 ∑ni=1 Ri
" #
1 n 1 n 1
Var [εp ] = Var ∑ εi = 2 ∑ var (εi ) ≤ 2 × max var (εi )
n i =1 n i =1 n i =1,...,n

which is arbitrarily close to zero as n → ∞ if var (εi ) is finite for all i

Thus, there is (at least an approximate) beta pricing model with f1 , ..., fk as
the factors (i.e. the systematic factor are the beta-pricing factors)
Multifactor Models
Data suggest asset returns variation can be summarized by few factors.

Principal Component Analysis

The first principal component is f1t = x1∗0 Rt , where

 
x1∗ = argmax x1 Σ̂x1 s.t. x10 x1 = 1
x1

and Σ̂ is a consistent estimator of the sample covariance matrix.

The second principal component is f2t = x2∗0 Rt
 
x2∗ = argmax x2 Σ̂x2 s.t. x20 x2 = 1 and x20 x1∗ = 0
x2

And so on...
Evidence shows that the first 3, 4 principal components are sufficient to
capture most of the variation (Roll and Ross, 1980)
Multifactor Models

With statistically chosen factors:

I Connor and Korajczyk (1988) use 6 principal components

With macroeconomic chosen factors:

I Chen, Roll, Ross (1986): maturity premium (long-short treasury yields
spread), expected inflation, unexpected inflation, industrial productivity
growth, default premium (BAA-AAA spread)

With empirically chosen factors

I Fama-French (1996): MKT-RF, SMB and HML (3-factor)
I Carhart (1997): + momentum (4-factor)
I Pastor-Stambaugh (2001): + liquidity (5-factor) (some theory)