Topic2 Reg Handout

Applied Financial Econometrics
Topic 2: Linear Regression Models

Hurn, Martin, Phillips & Yu (2020) Chapter 3
Semester 2, 2023/24
Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 1 / 61

Linear Regression
Simple Linear Regression

Linear Regression
Introduction
Linear regression is one of the most important tools in empirical finance.

The linear regression model provides a way to explain and predict the
movement of one financial variable based on the movement of other variables.
Does a regression always carry causal interpretations?

Linear Regression
CAPM introduction
One of the most important applications of linear regression in financial

econometrics is the capital asset pricing model (CAPM).
It is a way of relating the risk of an asset to market risk.
We can extend the simple CAPM by extending the set of explanatory
variables to contain size and growth factors.

Linear Regression
CAPM
The CAPM describes the risk characteristics of an asset in terms of β-risk,

given by the ratio
cov(rit − rft , rmt − rft )
β= .
var(rmt − rft )
rit − rft is the return on asset i relative to risk-free rate rft . We call this excess
return.
rmt − rft is market return relative to the risk-free rate.
We can use, e.g., treasury bill rate as rft and S&P 500 returns as rmt .
Does this expression look familiar?
What is the support of β? β ∈ R.

Linear Regression
Characterising β-risk
Individual stocks, or even the portfolios of stocks, are classified as follows in

terms of their degree of β-risk:
Aggressive: β>1
Benchmark: β=1
Conservative: 0<β<1
Uncorrelated: β=0
Imperfect Hedge: −1 < β < 0
Perfect Hedge: β = −1.
With which type do we earn the market return? Benchmark. Which type gives
us the risk-free return? Uncorrelated. (Independent of the market.)

Linear Regression
The model
How to estimate β?
We use the linear regression model
rit − rft = α + β(rmt − rft ) + ut .
The disturbance term ut captures additional movements in the dependent

variable not predicted by CAPM.
We assume zero conditional mean, that is E[ut |rmt − rft ] = 0.

Linear Regression
Model parameters
The regression model contains two unknown parameters.

The intercept parameter α captures the average abnormal return to the asset
given the relative risks.
The slope parameter β corresponds to the asset’s β-risk.

Linear Regression
Decomposing risk
Risk → 2nd moment (variance)

CAPM facilitates the decomposition:
E[(rit − rft )2 ] = E[(α + β(rmt − rft ))2 ] + E(ut2 ),

| {z } | {z } | {z }
Total risk Systematic risk Idiosyncratic risk
Which assumption did we rely on? Zero conditional mean so that the
correlation between rmt − rft and ut is zero.
Systematic risk = risk inherent to the entire market → nondiversifiable
Idiosyncratic risk → diversifiable R2

Linear Regression
In general ...
The simple linear regression model is
yt = α + βxt + ut .
The disturbance term, ut , represents movements in yt that are not explained

by the model and it has the properties
E(ut ) = 0,
E(ut2 ) = σu2 ,
E(ut ut−j ) = 0,
E(xt ut ) = 0.
If a series satisfies the first three conditions, we call it... white noise.

Linear Regression
The disturbance term
E(ut ) = 0 implies that the additional movements in yt that are in excess of the
movements predicted by CAPM balance out to be zero on average.
How does it differ from α? Note the subscript t.
E(ut2 ) = σu2 means that idiosyncratic risk as given by the variance of ut is
taken to be constant over time. What is the technical term we have for it?
Homoskedasticity.

Linear Regression
E(ut ut−j ) = 0, j ̸= 0 movements in ut are uncorrelated with previous (or

future) movements.
If this is not the case, the idiosyncratic component ut exhibits...
autocorrelation.
Autocorrelation implies arbitrage opportunities (potential of making a profit).
This is because this information may be utilised to improve the predictions of
future movements in yt .
Remember the EMH?

Linear Regression
E(xt ut ) = 0, the fourth and final property, ensures that movements in the
market (xt ) are independent of other factors (manifesting via ut ) caused by
non-market movements.
Is it a very realistic assumption?

Linear Regression
OLS in scalar form I
With simple linear regression
yt = α + βxt + ut ,
we have that
T −1 Tt=1 (yt − y )(xt − x)

P
βb =
T −1 Tt=1 (xt − x)2
P
b = y − βb x.
α

Linear Regression
OLS in scalar form II
and that
v
u PT
u bu2
σ t=1 xt2
se(b
α) = t PT ,
T t=1 (xt − x)2
s
bu2
σ
se(β)
b = PT ,
t=1 (xt − x)2
in which
T
1 X
bu2 =
σ (yt − α b t )2 .
b − βx
T
t=1
is the estimate of the residual variance.

These were derived in EofE.

Linear Regression
CAPM for nondurables
As an example, consider data on a United States nondurable industry portfolio

over the period January 1927 to December 2013.
Monthly returns to the market: value-weighted return of all CRSP (Center for
Research in Security Prices) firms incorporated in the US and listed on NYSE,
AMEX, and NASDAQ.
Risk free rate of interest: 1-month US Treasury Bill rate.
T −1 Tt=1 (xt − x)(yt − y )

P
23273.5059/1044
βb = 2 = = 0.7578,
−1
PT 30712.9719/1044
T t=1 xt − x
b = y − βx
α b = 0.6942 − 0.757 × 0.6449 = 0.2055,
T T
1 X b t )2 = 1
X 5026.2404
bu2 =
σ (yt − α
b − βx bt2 =
u = 4.8236.
T −2 T −2 1044 − 2
t=1 t=1
How would you categorise the β-risk? What about abnormal returns? Why
T − 2?

Linear Regression
CAPM for nondurables
The standard errors of αb and βb are, respectively,

v
bu2 Tt=1 xt2 √
u P r
u σ 4.8236 × 31147.2117
se(b
α) = t PT = = 0.0047 = 0.0684,
T t=1 (xt − x) 2 1044 × 30712.9719
s
√
r
b2
σ 4.8236
se(β) = PT
b = = 0.0002 = 0.0125.
t=1 (xt − x)
2 30712.9719
With more than one x, what is the analytical expression?

Linear Regression
Linear model in matrix form
Define the following column vectors

 
βb0
     

y1 β0 u1 u
b1
 y2  β1   u2  βb 
 1 ub2 
y =  . , β =  . , u =  . , βb = 
 ..  , u
b=
 . ,
       
 ..   ..   ..  .  .. 
yT βk uT βbk u
bT
and the matrix  

1 x11 ... x1K
1 x21 ... x2K 
X = . .. ..  .
 
 .. . ... . 
1 xT 1 ... xTK
The population multiple regression model is written in observation form as
y = X β + u.

Linear Regression
The least squares problem I
We have the objective function
T
X
S = ut2 = u ′ u = (y − X β)′ (y − X β)
t=1
= y y − y ′ X β − β ′ X ′ y + β ′ (X ′ X )β
′
= y ′ y − 2y ′ X β + β ′ (X ′ X )β .

Linear Regression
The least squares problem II
To find the β that minimises S, we take the derivative wrt β:
∂S
= 0 − 2X ′ y + 2(X ′ X )β.
∂β
Some matrix differentiation rules...

Setting the derivative to zero yields
−2X ′ y + 2(X ′ X )βb = 0 ⇒ (X ′ X )βb = X ′ y .

Linear Regression
The least squares problem III
Pre-multipling both sides of the equation with (X ′ X )−1 , we have the least
squares solution
βb = (X ′ X )−1 X ′ y .
−1
This follows because by definition, (X ′ X ) (X ′ X ) = I, and I is a k × k identity
matrix so I β̂ = β̂.
It can be shown that
b = σ 2 (X ′ X )−1 .
Var(β)

Linear Regression
The least squares problem IV
For (X ′ X )−1 to exist, we relied on the no perfect collinearity assumption.

Recall that y and X are nothing more than matrices of data so OLS is a
simple series of matrix operations (multiplication and inversion).
To ensure β̂ is a local minimum, we take the second derivative of S with
respect to β to find 2X ′ X , which is a positive definite matrix provided X has
full rank.

Linear Regression
Nondurables portfolio
The estimates of the CAPM model for the Nondurables portfolio are now computed
using the matrix formulation. The y and X matrices respectively are
−0.92 1 −0.11
   
 3.16  1 4.11 
1 −0.15
   
 2.46 
   
 3.12  1 0.52 
   
 7.88  1 5.40 
   
−2.03 1 −2.04
y =  ...  , X =  ... ..  .
   
  
   . 

 2.95  1 5.65 
   
−4.00 1 −2.69
   
 1.94  1 3.76 
   
 4.75  1 4.17 
   
 1.29  1 3.12 
2.65 1 2.81
The following matrices are needed

1044.0000 673.3100 724.7300
X ′X = , X ′y = .
673.3100 31147.2117 23740.9082
Linear Regression
Nondurables portfolio
The least squares estimates are then deduced as follows
βb = (X ′ X )−1 X ′ y
−1
1044.0000 673.3100 724.7300
=
673.3100 31147.2117 23740.9082

0.000971 −0.000021 724.7300
=
−0.000021 0.000033 23740.9082

0.2055
= ,
0.7578
and the residual variance is

b′ u
u 5026.2404
bu2 =
b
σ = = 4.8236.
T −2 1044 − 2
These estimates agree with the estimates reported previously.

Linear Regression
In Stata
. gen e_ind1 = ind1-rf
. regress e_ind1 mktrf // nondurables
Source | SS df MS Number of obs = 1,050

-------------+---------------------------------- F(1, 1048) = 3670.25
Model | 17653.4218 1 17653.4218 Prob > F = 0.0000
Residual | 5040.7421 1,048 4.80986841 R-squared = 0.7779
-------------+---------------------------------- Adj R-squared = 0.7777
Total | 22694.1639 1,049 21.6340933 Root MSE = 2.1931
------------------------------------------------------------------------------
e_ind1 | Coefficient Std. err. t P>|t| [95% conf. interval]
-------------+----------------------------------------------------------------
mktrf | .7577539 .0125078 60.58 0.000 .7332107 .782297
_cons | .207834 .0681665 3.05 0.002 .0740756 .3415923
------------------------------------------------------------------------------

Linear Regression
Extensions of CAPM
The CAPM model has been extensively studied.

There are useful extensions to the standard CAPM model.
We will focus on one extension today, the Fama-French 3 factor model.
Fama and French (1993) have proposed to augment the CAPM model by
including two additional risk factors to explain investment returns.
Colloquially these factors are referred to as size and value.

Linear Regression
Additional factor: SMB
Size, or “Small minus big” (SMB) accounts for the spread in returns between
small- and large-sized firms.
Size is determined by market capitalisation = share price × number of shares
outstanding.
Incorporating SMB into the CAPM shows whether management/investors are
relying on the small firm effect (investing in stocks with low market
capitalisation) to earn an abnormal return.

Linear Regression
Additional factor: HML
Value, or High-Minus-Low, (HML), is an additional factor that captures the

performance of “value” stocks relative to growth stocks.
Value stocks: companies with high book-to-market ratios. The market is
undervaluing the company.
Growth stocks: companies with low book-to-market ratios. The market is
placing a high premium on expected future earnings or growth potential.
Book-to-market ratio = book value of firm (historical cost or accounting value) /
market value of firm (market capitalisation).
HML = historic excess returns of (value stocks - growth stocks).

Linear Regression
Fama-French 3 factor model
Taken together with the ”market factor”, SMB and HML form the multi-factor,
or three-factor, CAPM:
rit − rft = α + β1 (rmt − rft ) + β2 SMBt + β3 HMLt + ϵt .

Fama and French propose the following interpretations:
For β2 , an estimated value greater than .5 signifies a portfolio composed mainly of
small cap stocks.
For β3 , an estimated value greater than .3 signifies a portfolio composed mainly of
value stocks.

Linear Regression
Other factors
We can add other factors, such as

RMW (Robust/high Minus Weak/low profitability),
CMA (Conservative Minus Aggressive)
momentum (size intersecting prior return)
These are left for you to explore. For data and definitions, see https:
//mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html

Diagnostics
Diagnostics

Diagnostics
Background
We have assumed that the CAPM is correctly specified.

But we might have omitted factors and misspecified the model.
To test the adequacy of the model, a number of diagnostic procedures are
available. These may be classified into three categories involving diagnostics
on
1 the dependent variable yt ,
2 the explanatory variables {x1t , x2t , · · · , xKt }, and
3 the disturbance term ut .

Diagnostics
Coefficient of determination, R 2
A natural measure of the success of an estimated model is given by the proportion

of the variation in the dependent variable explained by the model. This measure is
called the coefficient of determination and is defined as follows.
In terms of total variation, explained variation and unexplained variation:
Explained SS Unexplained SS
R2 = =1−
Total SS Total SS
Note: variation = sum of squares (SS)
In the context of CAPM, CAPM risk
systematic risk idiosyncratic risk

R2 = =1−
total risk total risk

Diagnostics
Adjusted coefficient of determination, R̄ 2
We use R̄ 2 for comparing models with the same dependent variable, but
different number of explanatory variables.
Adding explanatory variables will always increase R 2 .
Penalty adjustment in R̄ 2 gets larger with each additional model parameter:
T −1
R̄ 2 = 1 − (1 − R 2 )
T −K −1
Adding explanatory variables may increase or decrease R̄ 2 .

Diagnostics
Ordinary least squares estimates of the CAPM using 10 industry portfolios using
monthly data for the United States beginning January 1927 and ending December
2013.
2
Industry α
b βb R2 R RSS
Nondurables 0.205 0.758 0.778 0.7780 5026.2
(0.068 ) (0.013)
Durables 0.003 1.244 0.747 0.747 16110.2
(0.123) (0.022 )
Manufacturing 0.008 1.128 0.924 0.923 3234.9
(0.055) (0.010)
Energy 0.231 0.856 0.595 0.595 15289.8
(0.119 ) (0.022)
Tech 0.009 1.236 0.825 0.825 9939.8
(0.096) (0.018)
Telecom 0.152 0.657 0.591 0.591 9176.5
(0.092) (0.017)
Retail 0.107 0.969 0.789 0.788 7734.7
(0.085) (0.016)
Health 0.255 0.841 0.650 0.650 11696.9
(0.104) (0.019)
Utilities 0.089 0.782 0.576 0.576 13805.6
(0.113) (0.021)
Other -0.103 1.126 0.876 0.876 5524.7
(0.072) (0.013)

Diagnostics
Single parameter test: t-test
Consider
yt = α + β1 x1t + β2 x2t + . . . + βK xKt + ut
H0 : β = c
H1 : β ̸= c
β̂ − c
Test statistic and distribution: t= ∼ tT −K −1
se(β̂)
Recall that in large samples the distribution is nearly Normal, and so Normal
distribution critical values may be used:
right tail 10% 5% 2.5% 1% 0.5%
1.282 1.645 1.960 2.326 2.576

Diagnostics
t-test for parameter significance
This is a special case of the t-test:
H0 : β = 0
H1 : β ̸= 0
β̂
Test statistic and distribution: t= ∼ tT −K −1
se(β̂)

Diagnostics
Joint parameter test
H0 : list of restrictions on the model parameters

H1 : at least one of the listed restrictions does not hold
Builds on the concepts of “restricted” and “unrestricted” models, and compares

these two via their respective RSS (residual sum of squares) values.
RSS0 −RSS1
Test statistic and distribution: J= RSS1 /(T −K −1)
∼ χ2R
(RSS0 −RSS1 )/R
Test statistic and distribution: F = RSS1 /(T −K −1)
∼ FR,T −K −1
(Small samples)
where K = number of explanatory variables in the unrestricted model, and R = the

number of restrictions listed in H0 . Which is from the restricted model, RSS0 or
RSS1 ? RSS0 .

Diagnostics
Test for overall model significance
Special case:
H0 : β1 = β2 = . . . = βK = 0 i.e., all slope parameters are zero

H1 : at least one of the slopes is nonzero
Other forms of linear restrictions:

H0 : β1 = 1 and β2 + βK = 0
H1 : at least one of the restrictions listed above does not hold

Diagnostics
Disturbance diagnostics
The third set of diagnostic tests concerns the properties of the disturbance
term ut .
If CAPM is correctly specified, there should be no information left in the ut .
The adoption of tests concerning ut is especially important for those situations
where the coefficient of determination is found to be extremely small.
But, a low R 2 ↛ misspecification. Data could be noisy (high frequency).
An R 2 lower than 5% is common. According to EMH, there is no effective
predictor of future returns.

Diagnostics
Disturbance diagnostics
If the model is correctly specified, there should be no information regarding the

systematic part (the model) remaining in the disturbance term.
H0 : ut is random (model is correctly specified)
H1 : ut is not random (model is misspecified)
Recall two of the crucial regression model assumptions on the disturbance term
E(ut ut−j ) = 0, j ̸= 0 (no autocorrelation)
E(ut2 ) = σ 2 (constant variance)
Diagnostics on the disturbance term are concerned with its mean (autocorrelation),
variance (heteroskedasticity), and distributional shape (normality).

Diagnostics
Autocorrelation
Breusch-Godfrey test:
H0 : No autocorrelation in the disturbances
H1 : Autocorrelation up to lag p in the disturbances
Test statistic and distribution: AR(p) ∼ χ2p
where p = the autocorrelation lag length tested.
What is the effect of autocorrelated disturbances on the OLS estimator? OLS is no

longer efficient.

Diagnostics
Autocorrelation
Details:
Estimate
and get the residuals ût .
We want to test if the residuals have autocorrelation up to lag p. Estimate
ût = γ0 + γ1 x1t + γ2 x2t + . . . + γK xKt + ρ1 ût−1 + ρ2 ût−2 + . . . + ρp ût−p + vt
and test the joint restriction ρ1 = ρ2 = . . . = ρp = 0.
Test statistic and distribution: AR(p) = TR 2 ∼ χ2p

where R 2 is the coefficient of determination from the second regression, and T is
the sample size from the second regression.

Diagnostics
Heteroskedasticity
White’s test:
H0 : Homoskedastic disturbances (σu2 is constant)
H1 : Heteroskedastic disturbances (σu2 is time-varying)
Test statistic and distribution: HSK ∼ χ2k

where k is the number of explanatory regressors in the auxiliary regression.
What is the effect of heteroskecastic disturbances on the OLS estimator? Also

efficiency loss.

Diagnostics
Heteroskedasticity
Details:
Estimate
We want to test if the variance of the residuals (use squared residuals as a proxy)
is influenced by the explanatory variables. Estimate
ût2 = γ0 + γ1 x1t + γ2 x2t + . . . + γK xKt

+γK +1 x1t2 + γK +2 x2t2 + . . . + γK +K xKt
2
!
all bivariate combinations
+γ2K +1 x1 x2 + + vt
of regressors
and test the joint restriction γ1 = γ2 = . . . = γm = 0 (without γ0 ).

Test statistic and distribution: HSK = TR 2 ∼ χ2m where R 2 is the coefficient of
determination from the second (auxiliary) regression, and m is the number of
explanatory regressors in the auxiliary regression. What is m?
K

2
+ 2K = K × (K − 1)/2 + 2K

Diagnostics
ARCH
AutoRegressive Conditional Heteroskedasticity – volatility (variance) clustering.

Low variance is followed by low variance, and high variance is followed by high
variance (memory in variance).
ARCH test:
H0 : No ARCH in disturbances
H1 : ARCH of order p in disturbances
Test statistic and distribution: ARCH(p) ∼ χ2p
where p = the order of ARCH tested.

Diagnostics
ARCH
Details:
Estimate
We want to test if the variance of the residuals is influenced by the past variance of
the residuals. Estimate
ût2 = γ0 + γ1 ût−1
2 2
+ γ2 ût−2 2
+ . . . + γp ût−p + vt
and test the joint restriction γ1 = γ2 = . . . = γp = 0.
Test statistic and distribution: ARCH(p) = TR 2 ∼ χ2p

where R 2 is the coefficient of determination from the second regression, T is the
sample size from the second regression, and p = the number of lags in ARCH
tested.

Diagnostics
Normality
Normality test (Jarque–Bera test):

H0 : Disturbances are normally distributed
H1 : Disturbances are not normally distributed
Test statistic and distribution: JB ∼ χ22
We first calculate the measures of skewness and kurtosis for the disturbances:
T 3 T 4
1 X ût 1 X ût
SK = KT =
T σ̂u T σ̂u
t=1 t=1
If the disturbances come from a Normal distribution, then SK = 0 and KT = 3.
SK 2 (KT − 3)2

Test statistic and distribution: JB = T + ∼ χ22
6 24

Event Analysis
Event Analysis

Event Analysis
Introduction
Event analysis is used in empirical finance to model the effects of qualitative

changes on financial variables arising from a discrete event.
Typical events that are relevant in finance arise from announcements such as
the change in a company’s CEO, a monetary policy announcement, or
dramatic news events.
The overall event is decomposed into three sub-events: the part that is
anticipated by the market, the part that occurs at the time of the event, and the
part that happens after the event has occurred.
A typical event study involves specifying a regression equation based on
some given model that represents normal market returns, and then modifying
this equation to account for designated events through the inclusion of
indicator variables.

Event Analysis
Dummy variable
To model qualitative effects in an event analysis, we use an indicator variable,

It , that takes the value 1 if the event occurs at a point in time t and 0 for a
non-occurrence of the event.
Formally a dummy variable is defined as

1 : [event]
It =
0 : [non-event].
The parameter on a particular dummy measures the abnormal return

associated with that event, representing the return over and above the normal
return.

Event Analysis
Lee Raymond retires
In December of 2005, Lee Raymond retired as the CEO of Exxon receiving a

package of around US $400 mil.
How did the market react?
We model returns on Exxon (rt ) using market returns (rmt ) as
rt = β0 + β1 rmt
| {z }
Normal return
+ δ−2 IOct,t + δ−1 INov ,t + δ0 IDec,t + δ1 IJan,t + δ2 IFeb,t +ut ,

| {z }
Abnormal return
where, for example, (

1 : Oct. 2005
IOct,t =
0 : Otherwise
Which indicator(s) measure the reactions before, during, and after the event
takes place?

Event Analysis
Results
Model is estimated using data from January 1970 to February 2006.

Estimates are
rt = 0.009 + 0.651 rmt − 0.121 IOct,t + 0.007 INov ,t

(0.002) (0.044) (0.041) (0.041)
− 0.041 IDec,t + 0.086 IJan,t − 0.059 IFeb,t + u

bt .
(0.041) (0.041) (0.041)
Market anticipated the event with the October dummy being statistically
significant.
The market corrects itself, as the coefficients have alternating signs.
Overall, market viewed the event unfavourably. The net effect of the retirement
package on the market is negative with the total abnormal return equalling
Total = −0.121 + 0.007 − 0.041 + 0.086 − 0.059 = −0.128 with p-value = 0.16.

Event Analysis
Statistical significance
We next test the hypothesis that the parameters on the 5 event indicator
variables are jointly zero.
The null and alternative hypotheses are
H0 : δ−2 = δ−1 = δ0 = δ1 = δ2 = 0 [Normal returns]

H1 : at least one restriction is not valid [Abnormal returns].
Under the null hypothesis that δ−2 = δ−1 = δ0 = δ1 = δ2 = 0, the regression

model reduces to the simple CAPM.
Using a chi-square test the statistic is 16.399. As there are 5 restrictions being
tested, χ25 = 11.07, and this statistic has a p value of 0.0058.
H0 : δ−2 = δ−1 = 0, p-value = 0.01; H0 : δ1 = δ2 = 0, p-value = 0.04.
What do we conclude? The retirement package has a significant dampening
effect on Exxon returns.

Portfolio Performance

Portfolio performance measures
The performance of a portfolio is commonly measured in terms of its expected

return relative to its risk, in excess of a risk-free return.
Three well-known measures of a portfolio’s performance are
1 Sharpe ratio
µp − rf
S=
σp
2 Treynor index
µp − rf
T =
β
3 Jensen’s alpha
α = µp − rf − β(µm − rf )
where µp and µm are respectively the expected returns on the portfolio and
the market, rf is the risk-free rate, and the risk measures are portfolio risk σp
and β-risk from the CAPM.

Sharpe ratio
µp − rf
S=
σp
Since William Sharpe’s creation of the Sharpe ratio in 1966, it has been one
of the most referenced risk/return measures used in finance, and much of this
popularity is attributed to its simplicity.
The advantage is that the Sharpe ratio considers both systematic and
idiosyncratic risks.
The Sharpe ratio assumes that returns follow a normal distribution, which is
often not true.
It does not quantify the value added compared to the market return.

Treynor index
µp − r f
T =
β
The Treynor index also does not quantify the value added.
The Treynor index considers only systematic risk, since β is in the
denominator. Portfolios with identical systematic risk, but different total risk,
will be rated the same.

Jensen’s alpha
α = µp − rf − β(µm − rf )
Jensen’s alpha quantifies the added return as the excess return above that
predicted by the CAPM.
Rankings based on Jensen’s alpha take account of systematic risk alone, so it
will rank portfolios in a similar way to the Treynor index.
Jensen’s alpha is perhaps the most widely used among the three, as a
positive α is a necessary condition for good performance.

Industry portfolios
Industry Mean Std. Dev. Sharpe Treynor Jensen’s Rank Rank Rank
µ
bp σ
bp Ratio Index Alpha Sharpe Treynor Jensen
Nondur. 0.981 4.661 0.149 0.916 0.205 1 2 3
Durables 1.093 7.794 0.103 0.648 0.003 9 9 9
Manuf. 1.023 6.355 0.116 0.652 0.008 6 8 7
Energy 1.070 6.009 0.130 0.915 0.231 3 3 2
Tech. 1.094 7.371 0.109 0.653 0.009 7 7 7
Telecom. 0.863 4.639 0.124 0.876 0.152 4 4 4
Retail 1.019 5.914 0.124 0.755 0.107 5 6 5
Health 1.085 5.658 0.141 0.948 0.255 2 1 1
Utilities 0.881 5.591 0.106 0.759 0.089 8 5 6
Other 0.910 6.523 0.096 0.553 -0.103 10 10 10
What is the ranking criterion? With all three measures, the larger the value, the better the
performance.
Health has the highest Treynor index and Jensen’s alpha.
Nondurable sector is also highly ranked.
The worst performing portfolio is Other.

References I
Fama, E. F. & K. R. French (Feb. 1993). “Common risk factors in the returns on stocks and bonds”. J. financ.
econ. 33.1, pp. 3–56. ISSN: 0304-405X. DOI: 10.1016/0304-405X(93)90023-5.

Topic2 Reg Handout

Uploaded by

Copyright:

Available Formats

Topic2 Reg Handout

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Topic2 Reg Handout

Uploaded by

Copyright:

Available Formats

Applied Financial Econometrics

Topic 2: Linear Regression Models

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 1 / 61

Simple Linear Regression

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 2 / 61

Linear regression is one of the most important tools in empirical finance.

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 3 / 61

One of the most important applications of linear regression in financial

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 4 / 61

The CAPM describes the risk characteristics of an asset in terms of β-risk,

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 5 / 61

Individual stocks, or even the portfolios of stocks, are classified as follows in

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 6 / 61

rit − rft = α + β(rmt − rft ) + ut .

The disturbance term ut captures additional movements in the dependent

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 7 / 61

The regression model contains two unknown parameters.

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 8 / 61

Risk → 2nd moment (variance)

E[(rit − rft )2 ] = E[(α + β(rmt − rft ))2 ] + E(ut2 ),

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 9 / 61

The simple linear regression model is

The disturbance term, ut , represents movements in yt that are not explained

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 10 / 61

The disturbance term

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 11 / 61

The disturbance term

E(ut ut−j ) = 0, j ̸= 0 movements in ut are uncorrelated with previous (or

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 12 / 61

The disturbance term

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 13 / 61

OLS in scalar form I

With simple linear regression

T −1 Tt=1 (yt − y )(xt − x)

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 14 / 61

OLS in scalar form II

is the estimate of the residual variance.

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 15 / 61

CAPM for nondurables

As an example, consider data on a United States nondurable industry portfolio

T −1 Tt=1 (xt − x)(yt − y )

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 16 / 61

CAPM for nondurables

The standard errors of αb and βb are, respectively,

With more than one x, what is the analytical expression?

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 17 / 61

Linear model in matrix form

Define the following column vectors

and the matrix  

The population multiple regression model is written in observation form as

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 18 / 61

The least squares problem I

We have the objective function

Applied Financial Econometrics Topic 2. Linear Regression Models Semester 2, 2023/24 19 / 61

The least squares problem II

To find the β that minimises S, we take the derivative wrt β:

Some matrix differentiation rules...

−2X ′ y + 2(X ′ X )βb = 0 ⇒ (X ′ X )βb = X ′ y .