Professional Documents
Culture Documents
Eric Zivot
University of Washington
BlackRock Alternative Advisors
Outline
1. Introduction
Introduction
Factor models for asset returns are used to
Decompose risk and return into explanable and unexplainable components
Generate estimates of abnormal return
Describe the covariance structure of returns
Predict returns in specified stress scenarios
Provide a framework for portfolio risk analysis
(1)
Rit is the simple return (real or in excess of the risk-free rate) on asset i
(i = 1, . . . , N ) in time period t (t = 1, . . . , T ),
fkt is the kth common factor (k = 1, . . . , K),
ki is the factor loading or factor beta for asset i on the kth factor,
it is the asset specific factor.
Assumptions
1. The factor realizations, ft, are stationary with unconditional moments
E[ft] = f
cov(ft) = E[(ft f )(f t f )0] = f
2. Asset specific error terms, it, are uncorrelated with each of the common
factors, fkt,
cov(fkt, it) = 0, for all k, i and t.
3. Error terms it are serially uncorrelated and contemporaneously uncorrelated across assets
cov(it, js) = 2i for all i = j and t = s
= 0, otherwise
Notation
Vectors with a subscript t represent the cross-section of all assets
R1t
..
Rt =
, t = 1, . . . , T
(N 1)
RNt
Ri1
.
Ri =
. , i = 1, . . . , N
(T 1)
RiT
Matrix of all assets over all time periods (columns = assets, rows = time period)
R11 RN 1
.
..
...
R = .
(T N)
R1T RNT
Rt =
+ B
ft + t ,
(N 1) (N K)(K1) (N1)
(N 1)
01
11 1K
..
...
B
= .. = ..
(N K)
N1 NK
0N
E[t0t|ft] = D = diag( 21, . . . , 2N )
t = 1, . . . , T
(2)
Ri = 1T i + F
i + i , i = 1, . . . , N
(T K)(K1)
(T 1)
(T 1)
(T 1)(11)
f0
f11 fKt
..1
.
..
...
F
=
=
(T K)
fT0
f1T fKT
E[i0i] = 2i IT
Note: Time series homoskedasticity
(3)
Multivariate Regression
Collecting data from i = 1, . . . , N allows the model (3) to be expressed as the
multivariate regression
[R1, . . . , RN ] = 1T [1, . . . , N ] + F[1, . . . , N ] + [1, . . . , N ]
or
0 + F
B0 + E
(T
K)
(T N)
(1T
)
(KN)
(T 1)
= X0 + E
"
#
0
0
.
= [1T . F],
=
,
X
B0
(T (K+1))
((K+1)N )
R
(T N )
1T
R0
(N T )
10T +
(N 1)(1T )
F0
(N K)(KT )
0 + E0
= X
"
#
10T
0
X
=
= [ .. B],
0 ,
F
(N (K+1))
((K+1)T )
E0
(N T )
E[Rit] = i + 0iE[ft]
0iE[ft] = explained expected return due to systematic risk factors
i = E[Rit] 0iE[ft] = unexplained expected return (abnormal return)
Note: Equilibrium asset pricing models impose the restriction i = 0 (no
abnormal return) for all assets i = 1, . . . , N
Covariance Structure
Using the cross-section regression
Rt = + B
ft + t , t = 1, . . . , T
(N 1) (NK)(K1) (N 1)
(N1)
and the assumptions of the multifactor model, the (N N ) covariance matrix
of asset returns has the form
cov(Rt) = F M = Bf B0 + D
Note, (4) implies that
var(Rit) = 0if i + 2i
cov(Rit, Rjt) = 0if j
(4)
Portfolio Analysis
Let w = (w1, . . . , wn) be a vector of portfolio weights (wi = fraction of
wealth in asset i). If Rt is the (N 1) vector of simple returns then
Rp,t = w0Rt =
N
X
wiRit
i=1
Rt
Rp,t
p
var(Rp,t)
= + Bf t + t
Active portfolios have weights that change over time due to active asset
allocation decisions
Static portfolios have weights that are fixed over time (e.g. equally weighted
portfolio)
Factor models can be used to analyze the risk of both active and static
portfolios
Choice of factors
Estimate factor betas, i, and residual variances, 2i , using time series
regression techniques.
(5)
where RMt denotes the return or excess return (relative to the risk-free rate)
on a market index (typically a value weighted index like the S&P 500 index) in
time period t.
Risk-adusted expected return and abnormal return
E[Rit] = iE[RMt]
i = E[Rit] iE[RMt]
F M = 2M 0 + D
(6)
where
2M = var(RMt)
= ( 1, . . . , N )0
D = diag(21, . . . , 2N ),
2i = var(it)
Estimation
Because RMt is observable, the parameters i and 2i of the single factor
model (5) for each asset can be estimated using time series regression (i.e.,
ordinary least squares) giving
b +
b i, i = 1, . . . , N
b i1T + RM
Ri =
i
b = cov(R
d
d
iM /
2M
it, RMt)/var(R
Mt) =
i
iR
M
i
bi = R
1
b2
b0 bi
i =
T 2 i
The estimated single factor model covariance matrix is
0
bb
b
c
b2
FM =
M + D
Remarks
1. Computational eciency may be obtained by using multivariate regression.
The coecients i and i and the residual variances 2i may be computed
in one step in the multivariate regression model
R = X0 + E
The multivariate OLS estimator of 0 is
b 0 = (X0X)1X0R0.
EE
T 2
0
decomposition
var(Rit) = 2i var(RMt) + var(it) = 2i 2M + 2i
R2 can be estimated using
R2 =
2i
2M
d
var(R
it)
4. The single factor covariance matrix (6) is constant over time. This may
not be a good assumption. There are several ways to allow (6) to vary
over time. In general, i, 2i and 2M can be time varying. That is,
i = it, 2i = 2it, 2M = 2Mt.
To capture time varying betas, rolling regression or Kalman filter techniques
could be used. To capture conditional heteroskedasticity, GARCH models
may be used for 2it and 2Mt. One may also use exponential weights in
computing estimates of it, 2it and 2Mt. A time varying factor model
covariance matrix is
0
b b
b
c
b2
F M,t =
Mt t t + Dt,
Estimation
Because the factor realizations are observable, the parameter matrices B and
D of the model may be estimated using time series regression:
b +
b i = X
b i1T + F
Ri =
+ bi, i = 1, . . . , N
i
0
0
.
i) = (X0X)1X0Ri
X = [1T . F],
= (
i,
1
b2
b0ibi
i =
T K 1
The covariance matrix of the factor realizations may be estimated using the
time series sample covariance matrix
T
T
1 X
1 X
0
(ft f )(ft f ) , f =
ft
T 1 t=1
T t=1
The estimated multifactor model covariance matrix is then
b =
b
b b b0
c
F M = Bf B + D
(7)
Remarks
1. As with the single factor model, robust regression may be used to compute
i and 2i . A robust covariance matrix estimator may also be used to
compute and estimate of f .
2. F M can be made time varying by allowing i, f and 2i (i = 1, . . . , N )
to be time varying
BARRA Approach
In this approach, the observable asset specific fundamentals (or some transformation of them) are treated as the factor betas, i, which are time
invariant.
Fama-French Approach
This approach was introduced by Eugene Fama and Kenneth French (1992).
For a given observed asset specific characteristic, e.g. size, they determined
factor realizations using a two step process. First they sorted the crosssection of assets based on the values of the asset specific characteristic.
Then they formed a hedge portfolio which is long in the top quintile of the
sorted assets and short in the bottom quintile of the sorted assets. The
observed return on this hedge portfolio at time t is the observed factor
realization for the asset specific characteristic. This process is repeated for
each asset specific characteristic.
Given the observed factor realizations for t = 1, . . . , T, the factor betas
for each asset are estimated using N time series regressions.
Rt =
ft + t , t = 1, . . . , T
(N1)
(N 1)
(N 1)(11)
is an N 1 vector of observed values of an asset specific attribute (e.g.,
market capitalization, industry classification, style classification)
Estimation
For each time period t = 1, . . . T, the vector of factor betas, , is treated as
data and the factor realization ft, is the parameter to be estimated. Since the
error term t is heteroskedastic, ecient estimation of ft is done by weighted
least squares (WLS) (assuming the asset specific variances 2i are known)
ft,wls = (0D1)10D1Rt, t = 1, . . . , T
(8)
= diag( 21, . . . , 2N )
1)10D
1Rt, t = 1, . . . , T
ft,f wls = (0D
= diag(
21, . . . ,
2N )
D
Note 2: Other weights besides
2i could be used
h0 = (0D1)10D1
The estimated factor realization is then the portfolio return
ft,wls = h0Rt
When the portfolio h is normalized such that
factor mimicking portfolio.
PN
i hi = 1, it is referred to as a
cov(fit, fjt) = ij , i, j = 1, . . . , K
where
ik = 1 if asset i is in industry k (k = 1, . . . , K)
= 0, otherwise
It is assumed that there are Nk firms in the kth industry such
PK
k=1 Nk = N .
Rt = 1f1t + + K fKt + t,
= Bf t + t
0
E[tt] = D, cov(ft) = f
Since the industries are mutually exclusive it follows that
0j k = Nk for j = k, 0 otherwise
An unbiased but inecient estimate of the factor realizations ft can be obtained
by OLS:
1 PN1
1
N1 i=1 Rit
..
..
b
ft,OLS = (B0B)1B0Rt =
1 PNK K
fbKt,OLS
NK i=1 Rit
fb1t,OLS
OLS =
f OLS =
T
1 X
(bft,OLS f OLS)(bft,OLS f OLS)0,
T 1 t=1
T
1 X
b
ft,OLS
T t=1
i,OLS =
T
1 X
(bit,OLS i,OLS)2, i = 1, . . . , N
T 1 t=1
T
1 X
i,OLS =
bit,OLS
T t=1
c
b2
where D
OLS is a diagonal matrix with
i,OLS along the diagonal.
=
(bft,GLS f GLS)(bft,GLS f GLS)0
GLS
T 1 t=1
T
1 X
b2
=
(bit,GLS i,GLS)2, i = 1, . . . , N
i,GLS
T 1 t=1
0
b
bF
c
N = RR
T
(N N )
where R is the (N T ) matrix of observed returns.
RR0.
T =
N
(T T )
Factor Analysis
Traditional factor analysis assumes a time invariant orthogonal factor structure
Rt =
+ B
ft + t
(N 1)
(N 1) (NK)(K1) (N 1)
cov(ft, s)
E[ft]
var(ft)
var(t)
=
=
=
=
(9)
0, for all t, s
E[t] = 0
IK
D
where D is a diagonal matrix with 2i along the diagonal. Then, the return
covariance matrix, , may be decomposed as
= BB0+D
Hence, the K common factors ft account for all of the cross covariances of
asset returns.
Variance Decomposition
For a given asset i, the return variance variance may be expressed as
var(Rit) =
K
X
2ij + 2i
j=1
PK
2
j=1 ij , is called the commu-
Rt = + BHH0ft + t
= + Bft + t
where B= BH, ft= H0ft and var(ft) = IK .
Because the factors and factor loadings are only identified up to an orthogonal
transformation (rotation of coordinates), the interpretation of the factors may
not be apparent until suitable rotation is chosen.
Estimation
Estimation using factor analysis consists of three steps:
(10)
b is the sample mean vector of Rt. The error terms in (10) are hetwhere
eroskedastic so that OLS estimation is inecient.
ft
Estimation of Factor Realizations
Using (10), the factor realizations in a given time period t, ft, can be estimated
using the cross-sectional feasible weighted least squares (FWLS) regression
b
b 0D
c1B
b )1B
b 0D
c1(R
b)
ft,f wls = (B
t
(11)
Performing this regression for t = 1, . . . , T times gives the time series of factor
realizations (bf1, . . . , bfT ).
1
2
b | ln |B
bB
b0 + D
c| .
LR(K) = (T 1 (2N + 5) K) ln |
6
3
Remarks:
Traditional factor analysis is only appropriate if it is cross-sectionally uncorrelated, serially uncorrelated, and serially homoskedastic.
Principal Components
Principal component analysis (PCA) is a dimension reduction technique
used to explain the majority of the information in the sample covariance
matrix of returns.
With N assets there are N principal components, and these principal
components are just linear combinations of the returns.
The principal components are constructed and ordered so that the first
principal component explains the largest portion of the sample covariance
matrix of returns, the second principal component explains the next largest
portion, and so on. The principal components are constructed to be orthogonal to each other and to be normalized to have unit length.
In terms of a multifactor model, the K most important principal components are the factor realizations. The factor loadings on these observed
factors can then be estimated using regression techniques.
p1 solves
max p01p1 s.t. p01p1 = 1.
p1
= PP0
P
= [p1 .. p2 .. .. pN ], P0 = P1
(N N )
Variance Decomposition
N
X
i=1
var(Rit) =
N
X
var(fit) =
i=1
N
X
i=1
N = P
P
0
0 = P
1
= [p
1 .. p
2 .. .. p
N ], P
P
(N N )
= diag(
1, . . . ,
N ),
1 >
2 > >
N
The estimated factor realizations are simply the first K sample principal components
fbkt = p
0
k Rt, k = 1, . . . , K.
ft = (f1t, . . . , fKt)0
(12)
The factor loadings for each asset, i, and the residual variances, var(it) =
2i can be estimated via OLS from the time series regression
ft + it, t = 1, . . . , T
Rit = i + 0i
(13)
b and
b2
giving
i
i for i = 1, . . . , N . The factor model covariance matrix of
returns is then
where
and
b
b b b0
c
F M = Bf B + D
0
b = ..
c=
, D
B
0
b
N
b =
(14)
b2
0
0
1
.
.
. 0
0
,
2
bN
0
T
T
1 X
1 X
0
b
b
b
(ft f )(ft f ) , f =
ft
T 1 t=1
T t=1
!
1
w
k =
p
k , k = 1, . . . , K
(15)
0
1N p
k
where 1 is a (N 1) vector of ones, and the factor mimicking portfolio returns
are
k0 Rt
Rk,t = w
If traditional factor analysis is used, then there is a likelihood ratio test for
the number of factors. However, this test will not work if N > T .
i,f t
T
N X
X
(Rit i 0ift)2
i=1 t=1
Bai and Ngs model selection or information criteria are of the form
b 2(K) + K g(N, T )
IC(K) =
b 2(K) =
N
1 X
b2
N i=1 i
The preferred model is the one which minimizes the information criteria IC(K)
over all values of K < Kmax. Bai and Ng consider several penalty functions
and the preferred criteria are
N +T
NT
ln
,
N
T
N
+
T
N +T
2
2
2
b (K) + K
b (Kmax)
ln CNT
,
P Cp2(K) =
N
T
CNT = min( N , T )
b 2(K) + K
b 2(Kmax)
P Cp1(K) =
Algorithm
First, select a number Kmax indicating the maximum number of factors to be
considered. Then for each value of K < Kmax, do the following:
1. Extract realized factors bft using the method of APCA.
2. For each asset i, estimate the factor model
Rit = i + 0ibftK + it,
where the superscript K indicates that the regression has K factors, using
time series regression and compute the residual variances
b2
i (K) =
T
X
1
b2it.
T K 1 t=1
N
1 X
b2
(K)
N i=1 i