Fin econometricslecture

Methods Lectures: Financial Econometrics
Linear Factor Models and Event Studies

Michael W. Brandt, Duke University and NBER

NBER Summer Institute 2010

Michael W. Brandt Methods Lectures: Financial Econometrics

Outline

1 Linear Factor Models
Motivation
Time-series approach
Cross-sectional approach
Comparing approaches
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles


Linear Factor Models Motivation

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Linear factor pricing models
Many asset pricing models can be expressed in linear factor form

E[rt+1 − r0 ] = β λ

where β are regression coefﬁcients and λ are factor risk premia.

The most prominent examples are the CAPM

E[ri,t+1 − r0 ] = βi,m E[rm,t+1 − r0 ]

and the consumption-based CCAPM

E[ri,t+1 − r0 ] βi,∆ct+1 γVar[∆ct+1 ]
λ
Note that in the CAPM the factor is an excess return that itself
must be priced by the model ⇒ an extra testable restriction.


Econometric approaches

Given the popularity of linear factor models, there is a large
literature on estimating and testing these models.

These techniques can be roughly categorized as
Time-series regression based intercept tests for models in
which the factors are excess returns.
Cross-sectional regression based residual tests for models in
which the factors are not excess returns or for when the
alternative being considered includes stock characteristics.

The aim of this ﬁrst part of the lecture is to survey and put into
perspective these econometric approaches.


Linear Factor Models Time-series approach

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Black, Jensen, and Scholes (1972)

Consider a single factor (everything generalizes to k factors).

When this factor is an excess return that itself must be priced by
the model, the factor risk premium is identiﬁed through
ˆ ˆ
λ = ET [ft+1 ].

With the factor risk premium ﬁxed, the model implies that the
intercepts of the following time-series regressions must be zero
e
ri,t+1 − r0 ≡ ri,t+1 = αi + βi ft+1 + i,t+1

e
so that E[ri,t+1 ] = βi λ.

This restriction can be tested asset-by-asset with standard t-tests.



Joint intercepts test
The model implies, of course, that all intercepts are jointly zero,
which we test with standard Wald test

α (Var[ˆ ])−1 α ∼ χ2 .
ˆ α ˆ N

To evaluate Var[ˆ ] requires further distributions assumptions
α

With iid residuals t+1 ,
standard OLS results apply
 
−1
α
ˆ 1 1 Eˆ T [ft+1 ]
Var ˆ =  ˆ ˆ 2 ⊗Σ 
β T ET [ft+1 ] ET [ft+1 ]2

which implies  
2
1 ˆ
ET [ft+1 ]
Var[ˆ ] =
α 1+ ˆ
Σ .
T σT [ft+1 ]
ˆ



Gibbons, Ross, and Shanken (1989)
Assuming further that the residuals are iid multivariate normal, the
Wald test has a ﬁnite sample F distribution
2 −1
 
(T − N − 1)  ˆ T [ft+1 ]
E
1+  α Σ−1 α ∼ FN,T −N−1 .
ˆ ˆ ˆ
N σT [ft+1 ]
ˆ

In the single-factor case, this test can be further rewritten as
2 2
 
ˆ e
ET [rq,t+1 ] ˆ e
ET [rm,t+1 ]
e −
ˆ e
 
(T − N − 1)  σT [rq,t+1 ]
 ˆ σT [rm,t+1 ] 
,

N
 2
 ˆ e
ET [rm,t+1 ] 
1+
 
e
σT [rm,t+1 ]
ˆ

where q is the ex-post mean-variance portfolio and m is the
ex-ante mean variance portfolio (i.e., the market portfolio).


MacKinlay and Richardson (1991)
When the residuals are heteroskedastic and autocorrelated, we
can obtain Var[ˆ ] by reformulating the set of N regressions as a
α
GMM problem with moments

ˆ rt+1 − α − βft+1 ˆ t+1
gT (θ) = ET = ET =0
(rt+1 − α − βft+1 )ft+1 t+1 ft+1

Standard GMM results apply
α
ˆ 1 −1
Var ˆ = DS −1 D
β T
with
∂gT (θ) 1 ˆ
ET [ft+1 ]
D= =− ˆ T [ft+1 ] ˆ 2 ⊗ IN
∂θ E ET [ft+1 ]
∞
t t−j
S= E .
t ft t−j ft−j
j=−∞


Linear Factor Models Cross-sectional approach

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Cross-sectional regressions
When the factor is not an excess return that itself must be priced
by the model, the model does not imply zero time-series intercepts.

Instead one can test cross-sectionally that expected returns are
proportional to the time-series betas in two steps
1. Estimate asset-by-asset βi through time-series regressions
e
rt+1 = a + βi ft+1 + t+1 .

2. Estimate the factor risk premium λ with a cross-sectional
regression of average excess returns on these betas
ˆ ˆ
¯ie = ET [rie ] = λβi + αi .
r

Ignore for now the "generated regressor" problem.



OLS estimates
The cross-sectional OLS estimates are

λ = (β β)−1 (β¯ie )
ˆ ˆˆ ˆr
ˆˆ
α = ¯e − λβ.
ˆ r i

A Wald test for the residuals is then

αVar[α]−1 α ∼ χ2
ˆ ˆ ˆ N−1
with 1
Var[ˆ ] = I − β(β β)−1 β
α ˆ ˆˆ Σ I − β(β β)−1 β
ˆ ˆˆ
T
Two important notes
Var[ˆ ] is singular, so Var[ˆ ]−1 is a generalized inverse and
α α
the χ 2 distribution has only N − 1 degrees of freedom.

Testing the size of the residuals is bizarre, but here we have
additional information from the time-series regression (in red).


GLS estimates

Since the residuals of the second-stage cross-sectional regression
are likely correlated, it makes sense to instead use GLS.

The relevant GLS expressions are

λGLS = (β Σ−1 β)−1 (βΣ−1¯ie )
ˆ ˆ ˆ ˆ r

and
1
Var[ˆ GLS ] =
α Σ − β(β Σ−1 β)−1 β
ˆ ˆ ˆ .
T



Shanken (1992)
The generated regressor problem, the fact that β is estimated in
the ﬁrst stage time-series regression, cannot be ignored.
Corrected estimates are given by

1 λ2
Var[ˆ OLS ] = I − β(β β)−1 β
α ˆ ˆˆ Σ I − β(β β)−1 β
ˆ ˆˆ × 1+
T σf2
1 λ2
Var[ˆ GLS ] =
α Σ − β(β Σ−1 β)−1 β
ˆ ˆ ˆ × 1+
T σf2

In the case of the CAPM, for example, the Shanken correction for
annual data is roughly
2 2
ˆ
ET [rmkt,t+1 ] 0.06
1+ =1+ = 1.16
σT [rmkt,t+1 ]
ˆ 0.15



Cross-sectional regressions in GMM
Formulating cross-sectional regressions as GMM delivers an
automatic "Shanken correction" and allows for non-iid residuals.
The GMM moments are
e
 
rt+1 − a − βft+1
e
gT (θ) = ET [rt+1  (rt+1 − a − βft+1 )ft+1  = 0.
rt+1 − βλ

The first two rows are the first-stage time-series regression.
The third row over-identifies λ
If those moments are weighted by β we get the first order
conditions of the cross-sectional OLS regression.
If those moments are instead weighted by β Σ−1 we get the
first order conditions of the cross-sectional GLS regression.
Everything else is standard GMM.


Linear factor models in SDF form
Linear factor models can be rexpressed in SDF form
e
E[mt+1 rt+1 ] = 0 with mt+1 = 1 − bft+1 .

To see that a linear SDF implies a linear factor model, substitute in
e e e
0 = E[(1 − bft+1 )rt+1 ] = E[rt+1 ] − b E[rt+1 ft+1 ]
e e
Cov[rt+1 , ft+1 ] + E[rt+1 ]E[ft+1 ]
and solve for

e b e
E[rt+1 ] = Cov[rt+1 , ft+1 ]
1 − bE[ft+1 ]
e
Cov[rt+1 , ft+1 ] bVar[ft+1 ]
= .
Var[ft+1 ] 1 − bE[ft+1 ]
β λ



GMM on SDF pricing errors
This SDF view suggests the following GMM approach
ˆ e ˆ e ˆ e
gT (b) = ET (1 + bft+1 )rt+1 = ET [rt+1 ] − bET [rt+1 ft+1 ] = 0.
pricing errors = π

Standard GMM techniques can be used to estimate b and test
whether the pricing errors are jointly zero in population.

GMM estimation of b is equivalent to
Cross-sectional OLS regression of average returns on the
second moments of returns with factors (instead of betas)
when the GMM weighting matrix is an identity.
Cross-sectional GLS regression of average returns on the
second moments of returns with factors when the GMM
weighting matrix is a residual covariance matrix.



Fama and MacBeth (1973)
The Fama-MacBeth procedure is one of the original variants of
cross-sectional regressions consisting of three steps
1. Estimate βi from stock or portfolio level rolling or full sample
time-series regressions.
2. Run a cross-sectional regression
e ˆ
rt+1 = λt+1 β + αi,t+1

for every time period t resulting in a time-series
ˆ ˆ
{λt , αi,t }T .
t=1

3. Perform statistical tests on the time-series averages

1 T ˆ 1 T ˆ
αi =
ˆ t=1 αi,t
ˆ and λi = T t=1 λi,t .
T


Linear Factor Models Comparing approaches

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Time-series regressions

Source: Cochrane (2001)
Avg excess returns versus betas on CRSP size portfolios, 1926-1998.



Cross-sectional regressions

Avg excess returns versus betas on CRSP size portfolios, 1926-1998.



SDF regressions

Avg excess returns versus predicted value on CRSP size portfolios, 1926-1998.



Estimates and tests

CRSP size portfolios, 1926-1998.


Linear Factor Models Odds and ends

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Firm characteristics
So far

H0 : αi = 0 i = 1, 2, ..., n
HA : αi = 0 for some i

We might, typically on the basis of preliminary portfolio sorts, a
more specific alternative in mind

H0 : αi = 0 i = 1, 2, ..., n
HA : αi = γ xi

for some firm level characteristics xi .
We can test against this more specific alternative by including the
characteristics in the cross-sectional regression
ˆ
¯ie == λβi + γ xi + αi .
r



Portfolios
Stock level time-series regressions are problematic
Way too noisy.
Betas may be time-varying.

Fama-MacBeth solve this problem by forming portfolios
1. Each portfolio formation period (say annually), obtain a noisy
estimate of ﬁrm-level betas through stock-by-stock time series
regressions over a backward looking estimating period.
2. Sort stocks into "beta portfolios" on the basis of their noisy
but unbiased "pre-formation beta".

The resulting portfolios should be fairly homogeneous in their
beta, should have relatively constant portfolio betas through time,
and should have a nice beta spread in the cross-section.



Characteristic sorts

Somewhere along the way sorting on pre-formation factor
exposures got replaced by sorting on ﬁrm characteristics xi .

Sorting on characteristics is like sorting on ex-ante alphas (the
residuals) insead of betas (the regressor).

It may have unintended consequence
1. Characteristic sorted portfolios may not have constant betas.
2. Characteristic sorted portfolios could have no cross-sectional
variation in betas and hence no power to identify λ.



Characteristic sorts (cont)

Source: Ang and Liui (2004)



Conditional information
In a conditional asset pricing model expectations, factor
exposures, and factor risk premia all have a time-t subscripts

Et [rt+1 − r0 ] = βt λt .

This makes the model fundamentally untestable.
Nevertheless, there are at least two ways to proceed
1. Time-series modeling of the moments, such as

βt = β0 − κ(βt−1 − β0 ) + ηt .

2. Model the moments as (linear) functions of observables

βt = γβ Zt or λt = γλ Zt .

Approach #2 is by far the more popular.


Conditional SDF approach
In the context of SDF regressions
e
Et [mt+1 rt+1 ] = 0 with mt+1 = 1 − bt ft+1 .

Assume bt = θ0 + θ1 Zt and substitute in
e
Et (1 − (θ0 + θ1 Zt )ft+1 ) rt+1 = 0.

Condition down
e e
E Et [(1 − (θ0 + θ1 Zt )ft+1 )rt+1 ] = E (1 − (θ0 + θ1 Zt )ft+1 )rt+1 = 0.

Finally rearrange as unconditional multifactor model
e
E (1 − θ0 f1,t+1 − θ1 f2,t+1 )rt+1 = 0,

where f1,t+1 = ft+1 and f2,t+1 = zt ft+1 .

Event Studies Motivation

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



What is an event study?

An event study measures the immediate and short-horizon
delayed impact of a speciﬁc event on the value of a security.

Event studies traditionally answer the question
Does this event matter? (e.g., index additions)
but are increasingly used to also answer the question
Is the relevant information impounded into prices immediately
or with delay? (e.g., post earnings announcement drift)

Event studies are popular not only in accounting and ﬁnance, but
also in economics and law to examine the role of policy and
regulation or determine damages in legal liability cases.



Short- versus long-horizon event studies
Short-horizon event studies examine a short time window, ranging
from hours to weeks, surrounding the event in question.

Since even weekly expected returns are small in magnitude, this
allows us to focus on the information being released and (largely)
abstract from modeling discount rates and changes therein.

There is relatively little controversy about the methodology and
statistical properties of short-horizon event studies.

However event study methodology is increasingly being applied to
longer-horizon questions. (e.g., do IPOs under-perform?)

Long-horizon event studies are problematic because the results
are sensitive to the modeling assumption for expected returns.

The following discussion focuses on short-horizon event studies.


Event Studies Basic methodology

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Setup

Basic event study methodology is essentially unchanged since
Ball and Brown (1968) and Fama et al. (1969).

Anatomy of an event study
1. Event Deﬁnition and security selection.
2. Speciﬁcation and estimation of the reference model
characterizing "normal" returns (e.g., market model).
3. Computation and aggregation of "abnormal" returns.
4. Hypothesis testing and interpretation.

#1 and the interpretation of the results in #4 are the real economic
meat of an event study – the rest is fairly mechanical.



Reference models

Common choices of models to characterize "normal" returns
Constant expected returns
ri,t+1 = µi + i,t+1
Market model
ri,t+1 = αi + βi rm,t+1 + i,t+1

Linear factor models

ri,t+1 = αi + βi ft+1 + i,t+1

(Notice the intercepts.)

Short-horizon event study results tend to be relatively insensitive
to the choice of reference models, so keeping it simple is ﬁne.



Time-line and estimation

A typical event study time-line is

Source: MacKinlay (1997)

The parameters of the reference model are usually estimated over
the estimation window and held ﬁxed over the event window.



Abnormal returns
Using the market model as reference model, deﬁne
ˆ ˆ
AR i,τ = ri,τ − αi − βi rm,τ τ = T1 + 1, ..., T 2.
ˆ

Assuming iid returns
2
ˆ 1 (rm,τ − µm )
ˆ
Var AR i,τ = σ 2i + 1+ .
T1 − T0 σm
ˆ

due to estimation error

Estimation error also introduces persistence and cross-correlation
in the measured abnormal returns, even when true returns are iid,
which is an issue particularly for short estimation windows.

Assume for what follows that estimation error can be ignored.



Aggregating abnormal returns

Since the abnormal return on a single stock is extremely noisy,
returns are typically aggregated in two dimensions.
1. Average abnormal returns across all ﬁrms undergoing the
same event lined up in event time

¯ 1 N ˆ
AR τ = AR i,τ .
N i=1
2. Cumulate abnormal returns over the event window
ˆ T2 ˆ
CAR i (T1 , T2 ) = τ =T1 +1 AR i,τ

or
¯ T2 ¯
CAR(T1 , T2 ) = τ =T1 +1 AR τ .



Hypothesis testing
Basic hypothesis testing requires expressions for the variance of
¯ ˆ ¯
AR τ , CAR i (T1 , T2 ), and CAR(T1 , T2 ).

If stock level residuals i,τ are iid through time

ˆ
Var CAR i (T1 , T2 ) = (T2 − T1 )σ 2i .

The other two expressions are more problematic because they
depend on the cross-sectional correlation of event returns.

If, in addition, event windows are non-overlapping across stocks

¯ 1 N 2
Var AR τ = 2 i=1 σ i
N
¯ T2 ¯
Var CAR(T1 , T2 ) = τ =T1 +1 Var AR τ .



Example
Earnings announcements

Source: MacKinlay (1997)


Event Studies Bells and whistles

Outline

Motivation
Odds and ends

2 Event Studies
Motivation
Basic methodology
Bells and whistles



Dependence

If event windows overlap across stocks, the abnormal returns are
correlated contemporaneously or at lags.

One solution is Brown and Warner’s (1980) "crude adjustment"
Form a portfolio of ﬁrms experiencing an event in a given
month or quarter (still lined up in event time, of course).
Compute the average abnormal return of this portfolio.
Normalize this abnormal return by the standard deviation of
the portfolio’s abnormal returns over the estimation period.



Heteroskedasticity

Two forms of heteroskedasticity to deal with in event studies
1. Cross-sectional differences in σ i .
¯
• Standardize ARi,τ by σ i before aggregating to AR τ .
• Jaffe (1974).

2. Event driven changes in σ i ,t
.
• Cross-sectional estimate of Var [ARi,τ ] over the event window.
• Boehmer et al. (1991).



Regression based approach
An event study can alternatively be run as multivariate regression
L
r1,t+1 = α1 + β1 rm,t+1 + γ1,τ D1,t+1,τ + u1,t+1
τ =0
L
r2,t+1 = α2 + β2 rm,t+1 + γ2,τ D2,t+1,τ + u2,t+1
τ =0
...
L
rn,t+1 = αn + βn rm,t+1 + γn,τ Dn,t+1,τ + un,t+1
τ =0

where Di,t,τ = 1 if ﬁrm i had an event at date t − τ (= 0 otherwise).

In this case, γi,τ takes the place of the abnormal return ARi,τ .


Fin econometricslecture

More Related Content

What's hot (17)

Viewers also liked (7)

Similar to Fin econometricslecture (20)

More from NBER (20)

Fin econometricslecture