Avramov Doron Financial Econometrics

Lecture Notes
Financial Econometrics
Professor Doron E. Avramov
Hebrew University of
Jerusalem
'
Prof. Doron Avramov
1
Syllabus: Motivation
Why do we need a course in financial
econometrics?
The past few decades have been characterized
by an extraordinary growth in the use of
quantitative methods in financial markets in
analyzing various asset classes; be it equities,
fixed income instruments, commodities, or
derivative securities.
'
Prof. Doron Avramov
2
Syllabus: Motivation
Financial market participants, both academics
and practitioners, have been routinely using
advanced econometric techniques in a host of
applications including portfolio management,
risk management, modeling volatility,
understanding pivotal issues in corporate
finance, asset pricing, interest rate modeling, as
well as regulatory purposes.
'
Prof. Doron Avramov
3
Syllabus: Objectives
This course attempts to provide a fairly deep
understanding of topical issues in asset pricing
and deliver econometric methods in which to
develop research agenda in financial economics.
The course targets advanced master level and
PhD level students in finance and economics.
'
Prof. Doron Avramov
4
Syllabus: Prerequisite
I will assume some prior exposure to matrix
algebra, distribution theory, Ordinary Least
Squares, and Maximum Likelihood
Estimation.
I will also assume you have some skills in
computer programing beyond Excel.
Suggested packages include MATLAB,
STATA, SAS, R, etc.
'
Prof. Doron Avramov
5
Syllabus: Topics to be Covered
Overview:
Matrix algebra
Regression analysis
Law of iterated expectations
Variance decomposition
Taylor approximation
Distribution theory
Hypothesis testing
OLS
MLE
'
Prof. Doron Avramov
6
Testing asset pricing models including the CAPM
and multi factor models: Time series analysis.
The econometrics of the mean variance frontier
Estimating expected returns
Estimating the covariance matrix of stock returns.
Forming mean variance efficient portfolio, the
Global Minimum Volatility Portfolio, and the
minimum Tracking Error Volatility Portfolio.
'
Prof. Doron Avramov
7
The Sharpe ratio: estimation and distribution
The Delta method
The Black-Litterman approach for estimating expected
returns.
Principal component analysis.
'
Prof. Doron Avramov
8
Risk management and the downside risk measures:
value at risk, shortfall probability, expected
shortfall, target semi-variance, downside beta, and
drawdown.
Option pricing: testing the validity of the B&S
formula
Model verification based on failure rates
'
Prof. Doron Avramov
9
Syllabus: Topics to be
Covered
Variance ratio tests
Predicting asset returns using time series regressions
The econometrics of back-testing
Understanding time varying volatility models
including ARCH, GARCH, EGARCH, stochastic
volatility, implied volatility, and realized volatility
'
Prof. Doron Avramov
10
Syllabus: Grade Components
Assignments (35%): there will be three problem sets
during the term. You can form study groups to prepare
the assignments with up to four students per group.
Class Participation (15%) - Attending AT LEAST 80%
of the sessions is mandatory.
Take-home final exam (50%): based on class material,
handouts, assignments, and readings.
'
Prof. Doron Avramov
11
Let us Start
This session is mostly an overview. Major contents:
Why do we need a course in financial econometrics?
Normal, Bivariate normal, and multivariate normal
densities
The Chi-squared, F, and Student t distributions
Regression analysis
Basic rules and operations applied to matrices
Iterated expectations and variance decomposition
'
Prof. Doron Avramov
12
In previous courses in finance and economics you
mastered the concept of the efficient frontier.
A portfolio lying on the frontier is the highest
expected return portfolio for a given volatility target.
Or it is the lowest volatility portfolio for a given
expected return target.
'
Prof. Doron Avramov
13
Plotting the Efficient Frontier
'
Prof. Doron Avramov
14
However, how could you Practically Form
an Efficient Portfolio?
Problem: there are TOO many parameters to estimate.
For instance, investing in ten assets requires:
which is about ten estimates for expected return, ten for
volatility, and 45 for covariances/correlations.
Overall, 65 estimates are required. That is a lot!!!
(
(
(
(
(
(
10
2
1
.
.
(
(
(
(
(
2
10 10 , 1
2
2 2 , 1
10 , 1
2
,
1

K K K K K
K K K K K K K K
K K K K K
K K K K K
'
Prof. Doron Avramov
15
More generally, if there are N
investable assets, you need:
N estimates for the means,
N estimates for the volatilities,
0.5N(N-1) estimates for correlations.
Overall: 2N+0.5N(N-1) estimates are required!
Mean, volatility, and correlation estimates are noisy
as reflected through their standard errors.
'
Prof. Doron Avramov
16
The volatility estimate is less noisy than the mean.
I will later show that the standard error of the volatility
estimate is lower than that of mean return.
We have T asset return observations:
The sample mean and volatility are given by
T
R R R R K K K
3 2 1
, ,
( )
1
2
1
=

=
T
R R
T
t
t
Less noisy
'
Prof. Doron Avramov
17
T
R
R
T
t
t
=
=
1
Sample Mean and Volatility
Estimation Methods
One of the ideas here is to introduce robust methods
in which to estimate the comprehensive set of
parameters.
We will discuss asset pricing models and the Black
Litterman approach for estimating expected returns.
We will further introduce several methods for
estimating the large-scale covariance matrix of asset
returns.
'
Prof. Doron Avramov
18
Mean-Variance vs. Down
Side Risk
We will comprehensively cover topics in mean
variance analysis.
We will also depart from the mean variance
paradigm and consider down side risk measures
to form as well as evaluate investment strategies.
Why do we need to resort to down side risk?
'
Prof. Doron Avramov
19
Down Side Risk
For one, investors often care more about the
down side of investment payoffs than the upside
potential.
The practice of risk management as well as
regulations of financial institutions are typically
about downside risk such as targeting VaR,
shortfall probability, and expected shortfall.
Moreover, there is a major weakness embedded
in the mean variance paradigm.
'
Prof. Doron Avramov
20
Drawback in the Mean-Variance Setup
To illustrate, consider two stocks, A and B, with
correlation coefficient smaller than unity.
There are five states of nature in the economy.
Returns in the various states are described on the
next page.
'
Prof. Doron Avramov
21
Stock A dominates stock B in every state of nature.
Nevertheless, a mean-variance investor may consider
stock B because it can reduce the portfolios volatility.
Investment criteria based on down size risk measures
could get around this weakness.
S5 S4 S3 S2 S1
20.05% 10.50% 15.00% 8.25% 5.00% A
3.01% 3.10% 2.95% 2.90% 3.05% B
'
Prof. Doron Avramov
22
Drawback in Mean-Variance
The Normal Distribution
In various applications in finance and economics, a
common assumption is that quantities of interests, such
as asset returns, economic growth, dividend growth,
interest rates, etc., are normally (or log-normally)
distributed.
The normality assumption is primarily done for
analytical tractability.
The normal distribution is symmetric.
It is characterized by the mean and the variance.
'
Prof. Doron Avramov
23
Confidence Level
Normality suggests that deviation of 2 SD away from
the mean creates an 80% range (from -32% to 48%) for
the realized return with 95% confidence level.
'
Prof. Doron Avramov
24
Dispersion around the Mean
Assuming that x is a zero mean random variable.
As the distribution takes the form:
When is small, the distribution is concentrated and
vice versa.
( )
2
, 0 ~ N x
'
Prof. Doron Avramov
25
Probability Distribution Function
The Probability Density Function (pdf) of the normal
distribution is:
If then:
The Cumulative Density Function (CDF), which is the
integral of the pdf, is 50% at that point.
( )
(

=
2
2
2
2
1
exp
2
1
) (
r
r pdf
(
=
=
r
1
2
1
) ( = r pdf
'
Prof. Doron Avramov
26
Normally Distributed Return
Assume that the US excess rate of return on the
market portfolio is normally distributed with annual
mean (equity premium) and volatility given by 8% and
20%, respectively.
That is to say that with a nontrivial probability the
actual excess annual market return can be negative.
See figures on the next page.
'
Prof. Doron Avramov
27
Confidence Intervals for Annual Excess
Return on the Market
The probability that the realization is negative
% 99 ) 2 . 0 3 08 . 0 2 . 0 3 08 . 0 ( Pr
% 95 ) 2 . 0 2 08 . 0 2 . 0 2 08 . 0 ( Pr
% 68 ) 2 . 0 08 . 0 2 . 0 08 . 0 ( Pr
= + < <
= + < <
= + < <
R ob
R ob
R ob
'
Prof. Doron Avramov
28
% 4 . 34 ) 4 . 0 ( Pr
2 . 0
08 . 0 0
2 . 0
08 . 0
Pr ) 0 ( Pr
= < =
|
\
|

<
= <
z ob
R
ob R ob
Higher Moments
Skewness the third moment - is zero.
Kurtosis the fourth moment is three.
Odd moments are all zero.
Even moments are (often complex) functions of the
mean and the variance.
In the next slides, skewness and kurtosis are presented
for other distribution functions.
'
Prof. Doron Avramov
29
Skewness
The skewness can be negative (left tail) or
positive (right tail).
'
Prof. Doron Avramov
30
Kurtosis
'
Prof. Doron Avramov
31
Mesokurtic - A term used in a statistical context where the
kurtosis of a distribution is similar, or identical, to the
kurtosis of a normally distributed data set.
Positive skewness means non-zero probability for large payoffs.

Kurtosis is a measure for how tick the distributions tails are.
When is the skewness zero? In symmetric distributions.
'
Prof. Doron Avramov
32
The Essence of Skewness and
Kurtosis
Bivariate Normal Distribution
Bivariate normal:
The marginal densities of x and y are
What is the distribution of y if x is known?
(
(
|
|
\
|
|
|
\
|
(
2 2
2
,
,
, ~
y xy
xy x
y
x
N
y
x

( ) | |
( ) | |
2
2
, ~
, ~
y y
x x
N y
N x

'
Prof. Doron Avramov
33
Conditional Distribution
Conditional distribution:
always non-negative
When the correlation between x and y is positive
and - then the conditional expectation is
higher than the unconditional one.
( )
(
(
+
2
2
2
2
, ~
x
xy
y x
x
xy
y
x N x y
x
x >
'
Prof. Doron Avramov
34
Conditional Moments
If we have no information about x, or if x and y are
uncorrelated, then the conditional and unconditional
expected return of y are identical.
Otherwise, the expected return of y is .
If , then:
Here, the realization of x does not say anything about y.
0 =
xy
x y
y
x
xy
y x y

= =
2
2
2
'
Prof. Doron Avramov
35
Conditional Standard Deviation
Developing the conditional standard deviation further:
When goodness of fit is higher the conditional standard
deviation is lower.
( )
2
2 2
2
2
2
2
2
1 1 R
y
y x
xy
y
x
xy
y x y
=
|
|
\
|
= =

'
Prof. Doron Avramov
36
Define Z=AX+B.
( )
m m m
m
m
N
x
x
x
X

(
(
(
(
(
(
= , ~
.
.
1
2
1
1

'
Prof. Doron Avramov
37
Multivariate Normal
Let us make some transformations to end up with
N (0,I):
( )
( ) ( ) ) , 0 ( ~
) , 0 ( ~
) , ( ~
1 1
2
1
'
'
1 1
'
1 1 1 1
I N B A Z A A
A A N B A Z
A A B A N B X A Z
n m m n n m m m m n
n m m m m n n m m n
n m m m m n n m m n n m m n

+
+ + =

'
Prof. Doron Avramov
38
Multivariate Normal
Multivariate Normal
Consider an N-vector or returns which is normally
distributed:
Then:
What does mean? A few rules:
( )
N N N N
N R

, ~
1 1

( )
( ) ( ) I N R
N R
, 0 ~
, 0 ~
2
1
2
1
I =
=
=

2
1
2
1
2
1
2
1
1
2
1
2
1
'
Prof. Doron Avramov
39
If X1~N(0,1) then
If X1~N(0,1), X2~N (0,1) & then:
( ) 2 ~
2 2
2
2
1
X X +
2 1
X X
'
Prof. Doron Avramov
40
The Chi-Squared Distribution
( ) 1 ~
2 2
1
X
Moreover if
then
( )
( )
2 1
2
2
2
1
~
~
X X
n X
m X
( ) n m X X + +
2
2 1
~
'
Prof. Doron Avramov
41
More about the Chi-Squared
Gibbons, Ross & Shanken (GRS) designated a finite
sample asset pricing test that has the F - Distribution.
If
Then
) , ( ~
2
1
n m F
n
X
m
X
A =
( )
( )
2 1
2
2
2
1
~
~
X X
n X
m X
'
Prof. Doron Avramov
42
The F Distribution
The Students t Distribution
'
Prof. Doron Avramov
43
The pdf of student-t is given by:
Where is beta function and is a number of
degrees of freedom
The t Distribution
'
Prof. Doron Avramov
44
( )
( )
2
1
2
1
2 1 , 2
1
,
+
|
|
\
|
+
=
v
v
x
v B v
v x f
( ) b a B ,
1 v
As noted earlier, is a sample of length T of
stock returns which are normally distributed.
The sample mean and variance are
The resulting t-value which has the t distribution with T-1
d.o.f is
The t Distribution
'
Prof. Doron Avramov
45
T
r r r , , ,
2 1
K
T
r r
r
T
K +
=
1
and
( )
=
T
t
t
r r
T
s
1
2
2
1
1
1
=
T
s
r
t

The t-distribution is the sampling distribution of the t-value
when the sample consist of independent and identically
distributed observations from a normally distributed population.
It is obtained by dividing a normally distributed random variable
by a square root of a Chi-squared distributed random variable
when both random variables are independent.
Indeed, later we will show that when returns are normally
distributed the sample mean and variance are independent.
The t Distribution
'
Prof. Doron Avramov
46
Regression Analysis
Various applications in corporate finance and asset
pricing require the implementation of a regression
analysis.
We will estimate regressions using matrix notation.
For instance, consider the time series regression
t t t t
Z Z R + + + =
2 2 1 1
T , 2 , 1 , K K =
'
Prof. Doron Avramov
47
Rewriting the system in a matrix form:
yields
We will derive regression coefficients and their standard
errors using OLS, MLE, and Method of Moments.
(
(
(
(
(
(
2 1
22 21
12 11
3
, , 1
.
.
, , 1
, , 1
T T
T
Z Z
Z Z
Z Z
X
(
(
(
2
1 1 3
1 1 3 3 1
+ =
T T T
X R
(
(
(
(
(
(
T
T
.
.
2
1
1
'
Prof. Doron Avramov
48
(
(
(
(
(
(
T
T
R
R
R
R
.
.
2
1
1
Vectors and Matrices: some Rules
If A is a row vector:
Then the transpose of A is a column vector:
The identity matrix satisfies
| |
t t
a a a A
1 12 , 11 1
, K =
(
(
(
(
(
(
1
21
11
1
.
. '
t
t
a
a
a
A
B I B B I = =
'
Prof. Doron Avramov
49
Multiplication of Matrices:
,
The inverse of the matrix A is A
-1
which satisfies:
(
22 , 21
12 , 11
2 2
a a
a a
A
(
22 , 12
21 , 11
2 2
'
a a
a a
A
(
22 , 21
12 , 11
2 2
b b
b b
B
(
+ +
+ +
=

22 22 12 21 , 21 22 11 21
22 12 12 11 , 21 12 11 11
2 2 2 2
b a b a b a b a
b a b a b a b a
B A
(
= =

1 , 0
0 , 1
1
2 2 2 2
I A A
1 1 1
) (
' ' )' (

=
=
A B AB
A B AB
'
Prof. Doron Avramov
50
Solving Large Scale Linear Equations:
Of course, A has to be a square invertible matrix.
b A X
b X A
m m m m
1
1 1

=
=
'
Prof. Doron Avramov
51
Linear Independence and Norm
The vectors are linearly independent if there
does not exists scalars such that
The norm of a vector V is
'
Prof. Doron Avramov
52
N
V V , ,
1
K
N
c c , ,
1
K
unless
V c V c V c
N N
0
2 2 1 1
= + + + K
0
2 1
= = =
N
c c c K
V V V ' =
A Positive Definite Matrix
An matrix is called positive definite if
for any nonzero vector V.
A non positive definite matrix cannot be inverted and
its determinant is zero.
'
Prof. Doron Avramov
53
N N

0 '
1 1
>

N
N N
N
V V
The Trace of a Matrix
Let
Then
Trace is the sum of diagonal elements.
'
Prof. Doron Avramov
54
( ) ( )
| |
1 1
1
2
, 1 ,
2
2 2 , 1
, 1
2
, ,
,
1

=
(
(
(
(
(
=
N
N
N
N N
N
N N

K
K K K K K
K K K K K K K K
K K K K K
K K K K K
( )
2 2
2
2
1 N
tr + + + = K
Matrix Vectorization
This is the vectorization operator
it works for both square as well
as non square matrices.
'
Prof. Doron Avramov
55
( )
( )
( )
( )
(
(
(
(
(
N
N
Vec
M
2
1
1
2
The VECH Operator
Similar to but takes only the distinct elements
of .
'
Prof. Doron Avramov
56
( )
( )
( )
(
(
(
(
(
(
(
(
+
2
2
2
2
1
1
1
2
1
N
N
N N
Vech
M
M
( ) Vec
Partitioned Matrices
A, a square nonsingular matrix, is partitioned as
Then the inverse of A is given by
Check that
'
Prof. Doron Avramov
57
(
(
(
=

2 2 1 2
2 1 1 1
22 21
12 11
,
,
m m m m
m m m m
A A
A A
A
( ) ( )
( )
(
(
12
1
11 21 22
1
21 22 12 11 21
1
22
22
1
12
1
21 22 12 11
1
21 22 12 11
1
,
,
A A A A A A A A A A
A A A A A A A A A A
A
( )
2 1
1
m m
I A A
+
=
Matrix Differentiation
Y is an M-vector. X is an N-vector. Then
Let
'
Prof. Doron Avramov
58
(
(
(
(
(
(
N
M M M
N
N M
X
Y
X
Y
X
Y
X
Y
X
Y
X
Y
X
Y
, , ,
, , ,
2 1
1
2
1
1
1
K
M
K
A
X
Y
=
1 1
=
N N M M
X A Y
Let
'
Prof. Doron Avramov
59
' '
'
A X
Y
Z
A Y
X
Z
=
1 1
'

=
N N M M
X A Y Z
Let
If C is symmetric, then
'
Prof. Doron Avramov
60
( ) ' ' C C X
X
+ =
1 1
'

=
N N N N
X C X
C X
X
' 2 =
It is given by
g p m n mg np
B A C

=
(
(
(
B a B a
B a B a
C
nm n
m
mg np
,
,
1
1 11
K K K K K
K K K K K K K K K
K K K K K
'
Prof. Doron Avramov
61
Kronecker Product:
For A,B square matrices, it follows that
'
Prof. Doron Avramov
62
( )
( )( ) ( ) BD AC D C B A
B A B A
B A B A
N M
N N M M
=
=
=

1 1
1
Kronecker Product:
Operations on Matrices
'
Prof. Doron Avramov
63
| |
| | | |
)' )( ( ), (
) ( ' ) ' ( ) ' (
) ' ( ) ' (
) ( )' ' ( ) (
) ( ) ( ) (
) ( ) ( ) (
) ( ) ( ) (

= =
+ = = =
=
=
+ = +
+ = +
=
X X E X E
where
A tr A A XX E tr A XX tr E
AX X tr E AX X E
A vec B vec AB tr
B vech A vech B A vech
B vec A vec B A vec
B tr A tr B A tr
The expectation and the variance of a portfolios rate of
return in the presence of three stocks are formulated as

=
+ +
+ + + =
= + + =
'
2 2 2
'
23 3 2 3 2 13 3 1 3 1 12 2 1 2 1
2
3
2
3
2
2
2
2
2
1
2
1
2
3 3 2 2 1 1
p
p
'
Prof. Doron Avramov
64
Using Matrix Notation in a Portfolio
Choice Context
where
(
(
(
3
2
1
1 3
(
(
(
2
3 32 31
23
2
2 21
13 12
2
1
3 3
, ,
, ,
, ,

(
(
(
3
2
1
1 3
'
Prof. Doron Avramov
65
Using Matrix Notation in a Portfolio
Choice Context
The Expectation, Variance, and
Covariance of Sum of Random
Variables
) ( ) ( ) ( ) ( z cE y bE x aE cz by ax E + + = + +
'
Prof. Doron Avramov
66
) , ( 2 ) , ( 2 ) , ( 2
) ( ) ( ) ( ) (
2 2 2
z y bcCov z x acCov y x abCov
z Var c y Var b x Var a cz by ax Var
+ + +
+ + = + +
) , ( ) , (
) , ( ) , ( ) , (
w y bdCov z y bcCov
w x adCov z x acCov dw cz by ax Cov
+ +
+ = + +
Law of Iterated Expectations (LIE)
The LIE relates the unconditional expectation of a
random variable to its conditional expectation via the
formulation
Paul Samuelson shows the relation between the LIE and
the notion of market efficiency which loosely speaking
asserts that the change in asset prices cannot be predicted
using current information.
Prof. Doron Avramov
| | | | { } X Y E E Y E
x
| =
'
67
LIE and Market Efficiency
Under rational expectations, the time t security
price can be written as the rational expectation of
some fundamental value, conditional on information
available at time t:
Similarly,
The conditional expectation of the price change is
The quantity does not depend on the information.
| | ] [ |
* *
P E I P E P
t t t
= =
Prof. Doron Avramov
| | ] [ |
*
1 1
*
1
P E I P E P
t t t + + +
= =
| | 0 ] [ ] [ ] | [
* *
1 1
= =
+ +
P E P E E I P P E
t t t t t t
'
68
Variance Decomposition (VD)
Let us decompose :
Shiller (1981) documents excess volatility in the
equity market. We can use VD to prove it:
The theoretical stock price:
based on actual future dividends
The actual stock price:
| | y Var
| | ( ) | | ( ) | | x y Var E x y E Var y Var
x x
| | + =
| |
t t t
I P E P
*
=
( )
L L
2
2 1
*
1
1
r
D
r
D
P
t t
t
+
+
+
=
+ +
'
Prof. Doron Avramov
69
Variance Decomposition
What do we observe in the data? The opposite!!
| | ( ) | | ( ) | |
t t t t t
I P Var E I P E Var P Var
* * *
+ =
| | | | positive P Var P Var
t t
+ =
*
positive
| | | |
t t
P Var P Var >
*
'
Prof. Doron Avramov
70
Lecture Notes in Financial
Econometrics
Taylor Approximations in
Financial Economics
'
Prof. Doron Avramov
71
Several major applications in finance require the use of
Taylor series approximation.
See three applications in the following table.
We will also derive a test statistic for the Sharpe ratio
the price of risk which is based on the delta method,
which in turn is using the first order Taylor
approximation.
TA in Finance
'
Prof. Doron Avramov
72
TA in Finance: Major
Applications
'
Prof. Doron Avramov
73
Expected Utility Bond Pricing Option Pricing
First Oder Mean Duration Delta
Second Order Volatility Convexity Gamma
Third Order Skewness
Fourth Order Kurtosis
Taylor Approximation
Taylor series is a representation of a function as
an infinite sum of terms that are calculated from
the values of the function's derivatives at a single
point.
Taylor approximation is written as:
It can also be written with notation:
.... ) )( (
! 3
1
) )( (
! 2
1
) )( (
! 1
1
) ( ) (
3
0 0
' ' ' 2
0 0
' '
0 0
'
0
+ + + + = x x x f x x x f x x x f x f x f
( )
=
=
0
0
0
!
) (
) (
n
n
n
x x
n
x f
x f

'
Prof. Doron Avramov
74
Maximizing Expected Utility of
Terminal Wealth
). 1 )......( 1 )( 1 ( 1
) 1 (
2 1 K T T T
T K T
R R R R
where
R W W
+ + +
+
+ + + = +
+ =
The invested wealth is
The investment horizon is K periods.
The terminal wealth, the wealth in the end of the investment
horizon, is
T
W
'
Prof. Doron Avramov
75
Transforming to Log Returns
) 1 ln(
.....
) 1 ln(
) 1 ln(
2 2
1 1
K T K T
T T
T T
R r
R r
R r
+ +
+ +
+ +
+ =
+ =
+ =
'
Prof. Doron Avramov
76
Assume that the investor has the power utility function
of the from
The cumulative log return (CLR) over the investment
horizon is:
The terminal wealth as a function of CLR is
'
Prof. Doron Avramov
77
K T
K T
W W u
+
+
=

1
) (
K T T T
r r r r
+ + +
+ + + = ...
2 1
) exp( ) 1 ( r W R W W
T T K T
= + =
+
Power utility as a Function of
Log Return
Then the utility function of terminal wealth can be
expressed as a function of CLR
We can use the TA to express the utility as a function
of the moments of CLR.
That is, we will approximate around
'
Prof. Doron Avramov
78
) exp(
1 1
) ( r W W W u
T k T K T

= =
+ +
) (
K T
W u
+
) (r E =
Power utility as a Function of
Log Return
'
Prof. Doron Avramov
79
] ) (
24
1
) (
6
1
) (
2
1
1 )[ exp( )] ( [
.... ) )( exp(
24
1
) )( exp(
6
1
) )( exp(
2
1
) )( exp( ) exp( ) (
4 3
2
4 4
3 3 2 2

r Kur r Ske
r Var
W
W u E
r
r r
r
W W
W u
T
K T
T T
K T
+
+ +
+ +
+ +
+ =
+
+
The Utility Function
Approximation
Let us assume that log return is normally distributed:
The exact solution under normality is
| |
|
|
\
|
+ +
= =
4
4
2
2
4 2
8 2
1 ) exp( :
3 , 0 ) , ( ~
T
W
E Then
Kur Ske N r
'
Prof. Doron Avramov
80
| |
|
|
\
|
+ =
2
2
2
exp
T
W
E
Expected Utility
Taylor Approximation Bond Pricing
Taylor approximation is also used in bond pricing.
The bond price is:
where y0 is the yield to maturity.
Assume that y0 changes to y1
The delta (the change) of the yield to maturity is written
as:
( )
=
+
=
n
i
i
i
y
CF
y P
1
0
0
1
) (
0 1
y y y =
'
Prof. Doron Avramov
81
Changes in Yields and Bond
Pricing
Using Taylor approximation we get
2
0 1 0
' '
0 1 0
'
0 1
) )( (
! 2
1
) )( (
! 1
1
) ( ) ( y y y P y y y P y P y P + +
2
0 1 0
' '
0 1 0
'
0 1
) )( (
! 2
1
) )( (
! 1
1
) ( ) ( y y y P y y y P y P y P +
'
Prof. Doron Avramov
82
Duration and Convexity
Dividing by P(y
0
) yields
Instead of: we can write -MD
(Modified Duration).
Instead of: we can write Con
(Convexity).
) ( 2
) (
) (
) (
0
2
0
' '
0
0
'
y P
y y P
y P
y y P
P
P
+

) (
) (
0
0
'
y P
y P
) (
) (
0
0
' '
y P
y P
'
Prof. Doron Avramov
83
The Approximated Bond Price
Change
It is given by
What if the yield to maturity falls?
2
2
y Con
y MD
P
P
+
'
Prof. Doron Avramov
84
The change of the bond price is:
According to duration - bond price should increase.
According to convexity - bond price should also
increase.
Indeed, the bond price rises.
2
2
y Con
y MD
P
P
+ =
'
Prof. Doron Avramov
85
The Bond Price Change when
Yields Fall
And what if the yield to maturity increases?
Again, the change of the bond price is:
According to duration the bond price should
decrease.
2
2
y Con
y MD
P
P
+ =
'
Prof. Doron Avramov
86
Yields Increase
According to convexity - bond price should increase.
So what is the overall effect?
Notice: the influence of duration is always stronger
than that of convexity as the duration is a first order
effect while the convexity is second order.
'
Prof. Doron Avramov
87
Yields Increase
A call price is a function of the underlying asset price
What is the change in the call price when the
underlying asset pricing changes?
Focusing on the first order term this establishes the
delta neutral trading strategy.
( ) ( )
( ) ( )
2
2
2
2
2
1
) (
2
1
) ( ) (
t s t s t
t s t s t s
P P P P P C
P P
P
C
P P
P
C
P C P C
+ +
+
'
Prof. Doron Avramov
88
Option Pricing
Suppose that the underlying asset volatility increases.
Suppose further the implied volatility lags behind.
The call option is then underpriced buy the call.
However, you take the risk of fluctuations in the price
of the underlying asset.
To hedge that risk you sell Delta units of the
underlying asset.
'
Prof. Doron Avramov
89
Delta Neutral Strategy
Econometrics
OLS, MLE, MOM
'
Prof. Doron Avramov
90
Ordinary Least Squares (OLS)
The goal is to estimate the regression parameters.
Using optimization can help us reach the goal.
Consider the regression:
, ,
=
=
=
+ =
T
T
t
t t t
t t t
f
x y
x y
1
2
) (

(
(
(
=
T
y
y
Y .
1
(
(
(
=
KT T
K
x x
x x
X
, , , 1
, , , 1
1
1 11
K
K K K K K
K
(
(
(
=
T
E
.
1
'
Prof. Doron Avramov
91
First Order Conditions
Let us derive the function with respect to beta and
make the derivative equal to zero
Recall: X and Y are observations based on real data.
Could we choose other to obtain smaller
Y X X X
Y X X X
f
' 1 '
' '
) (
) (
0 2 ) ( 2
=
= =

=
=
T
T
t
E E
1
2 '
'
Prof. Doron Avramov
92
No! Because the optimization minimizes the quantity
.
We know:
Is the estimator unbiased for ?
, So it is indeed unbiased.
E X X X
E X X X X X X X
E X X X X E X Y
' 1 '
' 1 ' ' 1 '
' 1 '
) (
) ( ) (
) ( ) (
+ =
+ =
+ = + =

| |
| |
| | 0 ) (
) (
' 1 '
' 1 '
= =
=

E E X X X
E X X X E
E
E E
'
'
Prof. Doron Avramov
93
The OLS Estimator
What about the Standard Error of ?
Reminder:
( ) ( )
(
=
'

)
( E VAR
( )
1 ' '
'
' 1 '
) (
) (
=
=
X X X E
E X X X

'
Prof. Doron Avramov
94
The Standard Errors of the OLS Estimates
Continuing:
where K is the number of explanatory variables in
the regression specification.
| |
| |
E E
K T
X X
X X X X X EE E X X X
X X X EE X X X E

1
1
) (
) ( ) ( ) (
) ( ) (
' 2
2 1 '
2 1 ' 1 ' ' ' 1 '
1 ' ' ' 1 '

=
=
=
=

'
Prof. Doron Avramov
95
The Standard Errors of the OLS Estimates
MLE
We next turn to resorting to Maximum Likelihood as a
powerful tool for both estimating regression parameters
as well as testing models.
The MLE is an asymptotic procedure and it a
parametric approach in that the distribution of the
regression residual must be specified explicitly.
'
Prof. Doron Avramov
96
Implementing MLE
Assume that
Let us estimate and using MLE; then derive
the joint distribution of these estimates.
Under normality, the probability distribution
function (pdf) of the rate of return takes the form
( )
2
, ~ N r
iid
t
( )
(

=
2
2
2
2
1
exp
2
1
) (
t
t
r
r pdf
'
Prof. Doron Avramov
97
The Joint Likelihood
Since stock returns are assumed iid - it follows that
Now take the natural log of the joint likelihood:
( ) ( ) ( )
( )
(
(
\
|

=
=
T
T
t
T
T
r
L
r pdf r pdf r pdf L
1
2
2
2
2 1
2
1
exp 2
K
( ) ( )

=
|
\
|

=
T
t
t
r T T
L
1
2
2
2
1
ln
2
2 ln
2
ln

( ) ( ) ( ) ( )
T T T T
r pdf r r r pdf r r r pdf r r r pdf L = = K K K K
3 2 2 1 2 1
, , ,
'
Prof. Doron Avramov
98
MLE: Sample Estimates
Derive the first order conditions
Since - the variance estimator is not
unbiased.
( )
( )

= =
= =
= =
+ =
= =
T
t
T
t
t
t
T
t
t
T
t
t
r
T
r T L
r
T
r L
1 1
2
2
4
2
2 2
1 1
2
0
2
ln
1
0
ln

| |
2 2
E
'
Prof. Doron Avramov
99
MLE: The Information Matrix
Take second derivatives:
( )
( )
( )
=
=
=
=
|
|
\
|
=
|
|
\
|

=
|
|
\
|
= =
T
t
t
T
t
t
T
t
T L
E
r T L
L
E
r L
T L
E
T L
1
4 2
2
2
4
2
4 2
2
2
1
2
2
4 2
2
2
1
2
2
2 2 2
2
2
ln
2
ln
0
ln ln
ln 1 ln

'
Prof. Doron Avramov
100
MLE: The Covariance Matrix
Set the information matrix
Then the asy. distribution of the parameters is
In our context, the information matrix is derived as
|
|
\
|

=
'
ln
) (
2

L
E I
(
(
(
(
(
(
(
(
(
(
(
(

T
T
T
T
T
T
INVERSE MULTIPLY
4
2
4
2
" 1 "
4
2
2
, 0
0 ,
2
, 0
0 ,
2
, 0
0 ,
( )
1
) ( , 0 ~

\
|
I N T
'
Prof. Doron Avramov
101
To Summarize
The asy. distribution of the parameters is
In our context, the joint distribution of the sample
mean and variance assuming IID normal is
( )
1
) ( , 0 ~

\
|
I N T
|
|
\
|
(
(
(
(
(
(
= =
=
4
2
2
1 1
2
1
2 , 0
0 ,
,
0
0
~
)
1
(
1
1
N
r
T
r
T
r
T
T
T
t
t
T
t
t
T
t
t
'
Prof. Doron Avramov
102
The Sample Mean and Variance
Notice that when returns are normally distributed
the sample mean and variance are independent.
That would not be the case otherwise.
Departing from normality, the covariance between the
sample mean and variance is the skewness (see below).
Notice also that the ratio obtained by dividing the
variance of the variance by the variance of the mean is
smaller than one as long as volatility is below 70%.
Clearly, the mean return estimate is more noisy.
'
Prof. Doron Avramov
103
Departing from Normality:
Setting Moment Conditions
We know:
If:
Then:
That is, we set two momentum conditions.
( )
( )
2
2

=
=
t
t
r E
r E
( )
( )
(
(
=
2
2

t
t
t
r
r
g
( )
(
=
0
0
t
g E
'
Prof. Doron Avramov
104
Method of Moments (MOM)
There are two parameters:
Stage 1: Moment Conditions
Stage 2: estimation
2
,
( )
(
(
=
2
2

t
t
t
r
r
g
0
1
1
^
=
=
T
t
t
g
T
'
Prof. Doron Avramov
105
Method of Moments
Continue estimation:
( )
( ) | |
( )
=
=
=
=
=
=
=
T
t
t
T
t
t
T
t
t
t
r
T
r
T
r
T
r
T
1
2
2
1
2
2
1
0

1
1
'
Prof. Doron Avramov
106
T t , , 1 K =
MOM: Stage 3
( )
'
t t
g g E S =
(
(
4
4 3
3
2
,
,

( ) ( ) ( )
( ) ( ) ( ) ( )
(
(
+

=
4
2
2
4
2
3
2
3 2
2 ,
,

t t t t
t t t
r r r r
r r r
E S
'
Prof. Doron Avramov
107
MOM: stage 4
Memo:
Stage 4: derive wrt and take the expected
value
( )
( )
(
(
=
2
2

t
t
t
r
r
g
( )
t
g

( )
( )
(
=
(
=
(
=
1 , 0
0 , 1
1 , 2
0 , 1

t
t
r
E
g
E D
'
Prof. Doron Avramov
108
MOM: The Covariance Matrix
Stage 5:
( )
S
D S D
=
=

1
1 '
'
Prof. Doron Avramov
109
The Covariance Matrices
(
(
=
4
2
1
2 , 0
0 ,
(
(
=
4
4 3
3
2
2
,
,

Sample estimates of the Skewness and Kurtosis are

3
3
= = sk SK
( )
=
=
T
t
t
r r
T
1
3
3
1
( )
=
=
T
t
t
r r
T
1
4
4
1
sk is the skewness of the standardized

return
kr is the kurtosis of the standardized
return
4
4
= = kr KR
110
Under Normality the MOM Covariance
Matrix Boils Down to
1
4 4
4 3
3
2
2
2 , 0
0 ,

=
(
(
= =
=
=
'
Prof. Doron Avramov
111
MOM: Estimating Regression Parameters
Let us run the time series regression
There are two moment conditions:
t mt t
r r + + =
( )
( ) 0
0
=
=
t mt
t
r E
E
( )
( ) | |
| |
| |' ,
' , 1
'

=
=
=
(

=
mt t
t t t
mt mt t
mt t
t
r x
x r x
r r r
r r
g
'
Prof. Doron Avramov
112
MOM: Estimating Regression Parameters
Estimation:
( )
(
(
(
=
=
|
\
|
=
=
=
= =

mT
m
T
T
t
t t
T
t
t t
T
t
t t
T
t
t t
r
r
X
R X X X r x x x
x x
T
r x
T
, 1
, 1
0
1 1
1
2
'
1
'
1
1
1
'
^
1
^
'
1
K K
(
(
(
=
T
r
r
R K
1
'
Prof. Doron Avramov
113
MOM: Estimating Standard Errors
Estimation of the covariance matrix
T
X X
x x
T
g
T
D
x x
T
g g
T
S
t
T
t
t
T
t
t
T
t t
T
t
t t
T
t
t T
'
) ' (
1 ) ( 1
) ' (
1
)' ( ) (
1
1 1
^
^
2
1
^
1
^
= =
=
= =

= =
= =

'
Prof. Doron Avramov
114
MOM: Estimating Standard Errors
The covariance matrix estimate is thus
Then, asymptotically we get
1
1
2 1
1 1
) ' (
) ' ( ) ' (
) ' (
=
=
X X x x X X T
D S D
T
t
t t t
T T T
'
Prof. Doron Avramov
115
) , 0 ( ~ ) (
^
N T
Econometrics
Hypothesis Testing
'
Prof. Doron Avramov
116
Overview
A short brief of the major contents for todays class:
Hypothesis testing
TESTS: Skewness, Kurtosis, Bera-Jarque
A first step to testing asset pricing models
Deriving test statistic for the Sharpe ratio
'
Prof. Doron Avramov
117
Hypothesis Testing
Let us assume that a mutual fund invests in value
stocks (e.g., stocks with high ratios of book-to-market).
Performance evaluation is mostly about running the
regression of excess fund returns on the market premium
The hypothesis testing for examining performance is
H0: means no performance
H1: Otherwise
it
e
mt i i
e
it
R R + + =
0 =
i
'
Prof. Doron Avramov
118
Hypothesis Testing
Errors emerge if we reject H0 while it is true, or when
we do not reject H0 when it is wrong:
Reject H0 Dont
reject H0
Type 1
error
Good
decision
H0
Good
decision
Type 2 error
H1
True state
of world
'
Prof. Doron Avramov
119
Hypothesis Testing - Errors
is the first type error (size), while is the second
type error (related to power).
The power of the test is equal to 1- .
We would prefer both and to be as small as
possible, but there is always a trade-off.
When decreases increases and vice versa.
The implementation of hypothesis testing requires the
knowledge of distribution theory.

'
Prof. Doron Avramov
120
Skewness, Kurtosis & Bera
Jarque Test Statistics
We aim to test normality of stock returns.
We use three distinct tests to examine normality.
'
Prof. Doron Avramov
121
TEST 1 - Skewness (third moment)
The setup for testing normality of stock return:
H0:
H1: otherwise
Sample Skewness is
( )
2
, ~ N R
t
=
|
\
|
|
\
|

=
T
t
H
t
T
N
R
T
S
1
3
6
, 0 ~
1
0

'
Prof. Doron Avramov
122
Test I Skewness
Multiplying S by , we get
If the statistic value is higher (absolute value) than the
critical value e.g., the statistic is equal to -2.31, then
reject H0, otherwise do not reject the null of nomrality.
( ) 1 , 0 ~
6
N S
T
6
T
'
Prof. Doron Avramov
123
Test I Skewness
Kurtosis estimate is:
After transformation:
=
|
\
|
|
\
|

=
T
t
H
t
T
N
R
T
K
1
4
24
, 3 ~
1
0

( ) ( ) 1 , 0 ~ 3
24
0
N K
T
H
'
Prof. Doron Avramov
124
TEST 2 - Kurtosis
The statistic is:
Why ?
( ) ( ) 2 ~ 3
24 6
2
2
2
+ = K
T
S
T
BJ
) 2 (
2
'
Prof. Doron Avramov
125
TEST 3 - Bera-Jarqua Test
If X1~N(0,1) , X2~N (0,1) & then:
and are both standard normal and
they are independent random varaibles.
( ) 2 ~
2 2
2
2
1
X X +
2 1
X X
S
T
6
( ) 3
24
K
T
'
Prof. Doron Avramov
126
TEST 3 - Bera-Jarqua Test
Chi Squared Test
In financial economics, the Chi square test is
implemented quite frequently in hypothesis testing.
Let us derive it.
Suppose that:
Then:
( ) ( )
( ) ( ) ( ) N R R y y
I N R y
2 1
'
'
2
1
~
, 0 ~

=
=
'
Prof. Doron Avramov
127
Testing the CAPM
For one, the chi squared is used to test the CAPM
There are a few tests based on time series and cross
section specifications. Let us start with the time series.
The CAPM says that:
Thus, we first run the time-series market regressions:
) ( ) (
e
m i
e
i
R E R E =
N
e
m N N
e
N
e
m
e
R R
R R

+ + =
+ + =
K K K K K K K K
1 1 1 1
'
Prof. Doron Avramov
128
The expected excess return for asset i is given by:
According to the CAPM, the intercept i is restricted
to be zero for every test asset.
So, the joint hypothesis test is:
While doing the estimation, we will never get .
Rather, we examine whether the estimate is equal to
zero statistically. For example, the sample estimate could
be -- nevertheless this estimate could be
insignificant, or it is statistically equal to zero.
) ( ) (
e
m i i
e
i
R E R E + =
) (
otherwise H
H
N
:
0 :
1
2 1 0
= = = K
005 . 0
=
'
Prof. Doron Avramov
129
Testing the CAPM
0
=
The Test Statistic
Estimate :
If:
and
We get:
Later we will develop the exact analytics of this test.
(
(
(
=
N
( ) ( )
( )

, 0 ~
, ~
, ~
0
N
N N R
H
( ) N
H
2 1 '
0
~

'
Prof. Doron Avramov
130
Joint Hypothesis Test
You run the time series regression:
,
2 2 1 1 0 t t t t
x x y + + + = T t , , 2 , 1 K =
'
Prof. Doron Avramov
131
We have three orthogonal conditions:
Using matrix notation:
( )
( )
( ) 0
0
0
2
1
=
=
=
t t
t t
t
x E
x E
E
(
(
(
(
T
T
y
y
Y
.
.
1
1
(
(
(
(
T T
T
x x
x x
x x
X
2 1
22 12
21 11
3
, , 1
, , 1
, , 1
K K K K
| |
'
3 2 1 1 3
, , =
x
1 1 3 3 1
+ =
T T T
E X Y
(
(
(
(
T
T
E
.
.
1
1
'
Prof. Doron Avramov
132
Let us assume that: , are iid
Joint hypothesis testing:
( )
2
, 0 ~

N
t
T
K
3 2 1
, ,
( ) ( )
( ) Y X X X
X X N
'
1
'
2
1
'
, ~
=

otherwise H
H
:
0 , 1 :
1
2 0 0
= =
'
Prof. Doron Avramov
133
Define: ,
The joint test becomes:
otherwise H
q R H
H
:
0
1
1 , 0 , 0
0 , 0 , 1
:
1
2
0
2
1
0
0
0
(
=
(
=
(
(
(
=
1 , 0 , 0
0 , 0 , 1
R
(
=
0
1
q
'
Prof. Doron Avramov
134
q
R
Returning to the testing:
Under H0:
Chi squared:
test statistic
otherwise H
q R H
:
0 :
1
0
=
( )
'
, ~
R R q R N q R

( )
'
, 0 ~
0
R R N q R
H

( ) ( ) ( ) ( ) 2 ~

2
1
'
'

q R R R q R

'
Prof. Doron Avramov
135
Yet, another example:
, ,
(
(
(
T
T
y
y
Y .
1
1
(
(
(
(
T T
T
x x
x x
X
5 1
51 11
6
, , , 1
.
.
, , , 1
K
K
| |' , ,
6 1 0 1 6
K =
t t t t t t t
x x x x x y + + + + + + =
5 5 4 4 3 3 2 2 1 1 0
(
(
(
T
T
.
1
1
T t , , 3 , 2 , 1 K =
E X Y + =
'
Prof. Doron Avramov
136
Joint hypothesis test:
Here is a receipt:1.
2.
3.
4.
otherwise H
H
:
7 ,
4
1
,
3
1
,
2
1
, 1 :
1
5 3 2 1 0 0
= = = = =
( )
( ) ( )

=
=
=
2
1
'
' 2
'
1
'
, ~

6
1
X X N
E E
T
X Y E
Y X X X
'
Prof. Doron Avramov
137
5.
6 .
7 .
8 .
9 .
(
(
(
(
(
(
(
(
(
(
=
(
(
(
(
(
(
=
7
4
1
3
1
2
1
1
,
1 , 0 , 0 , 0 , 0 , 0
0 , 0 , 1 , 0 , 0 , 0
0 , 0 , 0 , 1 , 0 , 0
0 , 0 , 0 , 0 , 1 , 0
0 , 0 , 0 , 0 , 0 , 1
q R
( )
( )
( )( ) ( ) ( ) 5 ~

, 0 ~
, ~
0 :
2
1
'
'
'
'
0
0
0

H
H
q R R R q R
R R N q R
R R q R N q R
q R H

=
'
Prof. Doron Avramov
138
There are five degrees of freedom implied by the five
restrictions on
The chi-squared distribution is always positive.
5 3 2 1 0
, , , ,
'
Prof. Doron Avramov
139
Shrinkage Methods
One of the problems with the tests we have just
displayed is that the decision is binary: Reject or do
not reject (Classical econometrics) the null.
The possibility of partly rejecting/accepting the null,
for example, does not exist.
'
Prof. Doron Avramov
140
Shrinkage Methods and the CAPM
One can advocate a Bayesian approach in which
the proposed model (CAPM) is recognized to be non-
perfect, but at the same time it is worth something.
For example, let us assign a 50% weight to the
CAPM and 50% to data.
'
Prof. Doron Avramov
141
Shrinkage Methods and the CAPM
50% CAPM :
50% Data:
Data based estimate is the sample mean. (Check it!)
Let us consider three Mutual funds, with each of the
fund managers has one of the following beliefs:
M1 The CAPM is true
M2 The CAPM is wrong
M3 - Mix (equally weights) between CAPM and the Data
) ( ) (
e
m i
e
i
R E R E =
( ) ( )
e
m
e e
m
e
R E R E R R
1 1 1 1 1 1 1
+ = + + =
0 =
0
'
Prof. Doron Avramov
142
Shrinkage and Performance
Over the last several decades, the third Mutual fund
would have had the best performance.
Using the mixing method improves the estimation of
expected return.
The Black Litterman method (coming up later) is also
a type of a shrinkage approach - it is more rigorous than
the one presented here.
The weights are on the model versus views on
expected returns, either absolute or relative.
'
Prof. Doron Avramov
143
Estimating the Sample Sharpe Ratio
You observe time series of returns on a stock, or a
bond, or any investment vehicle (e.g., a mutual fund or
a hedge fund):
You attempt to estimate the mean and the variance of
those returns, derive their distribution, and test whether
the Sharpe Ratio of that investment is equal to zero.
Let us denote the set of parameters by
The Sharpe ratio is equal to
f
r
SR

= ) (
]' , [
2
=
( )
T
r r r , , ,
2 1
K
'
Prof. Doron Avramov
144
MLE vs. MOM
To develop a test statistic for the SR, we can
implement the MLE or MOM, depending upon our
assumption about the return distribution.
Let us denote the sample estimates by
MLE MOM
( )
2
, ~ N r
iid
t
( ) ( )
1
, 0 ~

N T
a
( ) ( )
2
, 0 ~

N T
a
returns depart from

iid normal
'
Prof. Doron Avramov
145
MLE vs. MOM
As shown earlier, the asymptotic distribution using
either MLE or MOM is normal with a zero mean but
distinct variance covariance matrices.
( ) ( )
1
, 0 ~

N T
MLE
( ) ( )
2
, 0 ~

N T
MOM
'
Prof. Doron Avramov
146
Distribution of the SR Estimate:
The Delta Method
We will show that
We use the Delta method to derive
The delta method is based upon the first order Taylor
approximation.
( ) ( )
2
, 0 ~
SR
a
N SR R S T
2
SR
'
Prof. Doron Avramov
147
Distribution of the SR Estimate:
The Delta Method
The first-order TA is
The derivative is estimated at
( ) ( ) ( )
( ) ( ) ( )
1 2
2 1
1 1
1 2
2 1
1 1
'
'
+ =

SR
SR SR
SR
SR SR
'
Prof. Doron Avramov
148
Distribution of the Sample SR
'
Prof. Doron Avramov
149
( ) ( ) | | ( ) ( )
( ) ( )( )
( ) ( ) | | ( ) ( ) | |
2
'
'

0

SR SR E SR SR VAR
E VAR
MOREOVER
E
SR
E SR SR E
=
(
=
=
(
=
The Variance of the SR
Continue:
( )( )
( )( )

=
=
=
=
(
SR SR
SR
E
SR
SR SR
E
'
'
'
'
'

'
Prof. Doron Avramov
150
First Derivatives of the SR
The SR is formulated as
Let us derive:
( )
5 . 0
2
f
r
SR

=
( ) ( )
( )
3 2
5 . 0
2
2
2
2
1
1

f
f
r
r
SR
SR

=

=
'
Prof. Doron Avramov
151
The Distribution of SR under MLE
Continue:
,
( )
|
|
\
|
|
\
|
+
|
|
\
|
|
|
\
|
|
|
\
|

=
2
4
2
3
1
'
'
2
1
1 , 0 ~
2 , 0
0 ,
2
,
1
R S N R S SR T
r
SR SR
f

|
|
|
|
\
|
|
|
\
|

3
2
1
f
r
'
Prof. Doron Avramov
152
The Delta Method in General
Here is the general application of the delta method
If:
Then let be some function of :
where is the vector of derivatives of
with respect to
( ) ( )
, 0 ~
N T
( ) d

( ) ( ) ( ) ( ) ( ) ( )
'
, 0 ~

D D N d d T
( ) D
( ) d
'
Prof. Doron Avramov
153
Hypothesis Testing
Does the S&P index outperform the Rf?
H0: SR=0
H1: Otherwise
Under the null there is no outperformance.
Thus, under the null
) 1 , 0 ( ~
1 , 0 ( ~
2
2
N
R S
R S T
R S N R S T
+
+
'
Prof. Doron Avramov
154
Lecture Notes in
The Efficient Frontier and the
Tangency Portfolio
'
Prof. Doron Avramov
155
Testing Asset Pricing Models
Central to this course is the introduction of test
statistics to examine the validity of asset pricing
models.
There are time series as well as cross sectional
asset pricing tests.
'
Prof. Doron Avramov
156
Testing Asset Pricing Models
Time series tests correspond to only those cases
where the factors are portfolio spreads, such as
excess return on the market portfolio as well as the
SMB (small minus big), the HML (high minus low),
and the WML (winner minus loser) portfolios.
The cross sectional tests apply to both portfolio and
non-portfolio based factors.
Consumption growth in the CCAPM is a good
example of a factor which is not a return spread.
'
Prof. Doron Avramov
157
Time Series Tests and the Tangency
Portfolio
Interestingly, time series tests are directly linked to
the notion of the tangency portfolio and the efficient
frontier.
Here is the efficient frontier, in which the tangency
portfolio is denoted by T.
T
'
Prof. Doron Avramov
158
Economic Interpretation of the Time
Series Tests
Testing the validity of the CAPM entails the time
series regressions:
The CAPM says: H0:
H1: Otherwise
Nt
e
mt N N
e
Nt
t
e
mt
e
t
r r
r r

+ + =
+ + =
K K K K K K K
1 1 1 1
N
= = = K
2 1
'
Prof. Doron Avramov
159
Economic Interpretation of the Time
Series Tests
The null is equivalent to the hypothesis that the
market portfolio is the tangency portfolio.
Of course, even if the model is valid the market
portfolio WILL NEVER lie on the estimated
frontier.
This is due to sampling errors.
The question is whether the market portfolio is
close enough, statistically, to the tangency portfolio.
'
Prof. Doron Avramov
160
What about Multi-Factor Models?
The CAPM is a one-factor model.
There are several extensions to the CAPM.
The multivariate version is given by the K-factor
model:
'
Prof. Doron Avramov
161
Testing Multifactor Models
The null is again: H0:
H1: Otherwise
In the multi-factor context, the hypothesis is that
some optimal combination of the factors is the
tangency portfolio.
Nt K NK N N N
e
Nt
t K K
e
t
f f f r
f f f r

+ + + + + =
+ + + + + =
K
K K K K K K K K K K K K K K K
K
2 2 1 1
1 1 2 12 1 11 1 1
N
= = = K
2 1
'
Prof. Doron Avramov
162
The Efficient Frontier: Investable Assets
Consider N risky assets whose returns at time t are:
The expected value of return is denoted by:
(
(
(
Nt
t
N
t
R
R
R K
1
1
( )
( )
( )

=
(
(
(
=
(
(
(
=
N Nt
t
t
R E
R E
R E K K
1 1
'
Prof. Doron Avramov
163
The Covariance Matrix
The variance covariance matrix is denoted by:
( )( ) | |
(
(
(
(
(
= =
2
2
2
2
1
2
1
,
, , ,
, , ,
'
N
N
N
t t
R R E V

K K K K K
K K K K K K
K K K
K K K
'
Prof. Doron Avramov
164
Creating a Portfolio
A portfolio is investing is the N assets.
The return of the portfolio is:
The expected return of the portfolio is:
(
(
(
N
N
w
w
w K
1
1
N N p
R w R w R w R + + + = K
2 2 1 1
( ) ( ) ( ) ( )
=
= = + + + =
= + + + =
N
i
i i N N
N N p
w w w w w
R E w R E w R E w R E
1
2 2 1 1
2 2 1 1
' K
K
'
Prof. Doron Avramov
165
The variance of the portfolio is:
'
Prof. Doron Avramov
166
2 2
2 2 1 1
2 2
2
2
2
2 12 2 1
N N N N N N
N N
w w w w w
w w w w w

+ + + +
+
+
+ + + +
K
M
K
( )
N N p p
w w w w w R VAR
1 1 12 2 1
2
1
2
1
2
K + + = =
Thus
where is the coefficient of correlation.
Using matrix notation:
= =
=
N
i
N
i
ij j i j i p
w w
1 1
2

1 1
1 1
2
'

=
N N N N
P
w V w
ij
'
Prof. Doron Avramov
167
The Case of Two Risky Assets
To illustrate, let us consider two risky assets:
We know:
Let us check:
So it works!
2 2 1 1
R w R w R
p
+ =
( )
2
2
2
2 12 2 1
2
1
2
1
2
2 w w w w R VAR
pt p
+ + = =
( )
2
2
2
2 12 2 1
2
1
2
1
2
1
2
2 12
12
2
1
2 1
2
2
,
,
,

w w w w
w
w
w w
p
+ + =
|
|
\
|
|
|
\
|
=
'
Prof. Doron Avramov
168
Dominance of the Covariance
When the number of assets is large, the
covariances define the portfolios rate of return.
To illustrate, assume that all assets have the same
volatility and pairwise correlations.
Then an equal weight portfolios variation is
cov
) 1 (
2
2
2 2
=
(
+
=

N
p
N
N N N
'
Prof. Doron Avramov
169
The Efficient Frontier: Excluding
Riskfree Asset
The optimization program:
min
s.t
where is the Greek letter iota ,
and where is the expected
return target set by the investor.
Vw w'
1 ' = w
p
w = '
(
(
(
1
1
1
K
N
'
Prof. Doron Avramov
170
The Efficient Frontier: Excluding
Riskfree Asset
Using the Lagrange setup:
( ) ( )

1
2
1
1
2 1
2 1
0
' ' 1 '
2
1

+ =
= =
+ + =
V V w
Vw
w
L
w w Vw w L
p
'
Prof. Doron Avramov
171
The Efficient Frontier: Without
Riskfree Asset
Let
The optimal portfolio is:
( )
( )

1 1
1 1
2
1
1
1
1
1
'
'
'

=
=
=
=
=
=
aV cV
d
h
aV bV
d
g
a bc d
V c
V b
V a
p
N
N
N
h g w
1
1
1
*
+ =
'
Prof. Doron Avramov
172
Examine the Optimal Solution
That is, once you specify the expected return target,
the optimal portfolio follows immediately.
Let us check whether the sum of weights is equal
to 1.
'
Prof. Doron Avramov
173
Indeed, .
( ) ( )
( ) ( ) 0
1
' '
1
'
1
1
' '
1
'
' ' '
1 1
2 1 1
= = =
= = = =
+ =

ac ac
d
V a V c
d
h
d
d
a bc
d
V a V b
d
g
h g w
p

1 ' = w
'
Prof. Doron Avramov
174
Let us now check whether the expected return on
the portfolio is equal to .
Recall:
So we need to show
p
( )

' ' ' h g w
h g w
p
p
+ =
+ =
1 '
0 '
=
=
h
g
'
Prof. Doron Avramov
175
Try it yourself: prove that .
( ) ( ) 0
1
' '
1
'
1 1
= = =

ab ab
d
V a V b
d
g
1 ' = h
'
Prof. Doron Avramov
176
Efficient Frontier
The optimization program also delivers the shape of
the frontier.
inefficient part
A
'
Prof. Doron Avramov
177
Efficient Frontier
Where point A stands for the Global Minimum
Variance Portfolio (GMVP).
The efficient frontier reflects the investment
opportunities; this is the supply side.
Points below A are inefficient since they are being
dominated by other more attractive portfolios.
'
Prof. Doron Avramov
178
The Notion of Dominance
If
and there is at least one strong inequality, then
portfolio B dominates portfolio A.
A B
A B

'
Prof. Doron Avramov
179
The Efficient Frontier with Riskfree
Asset
In practice, there is no really a riskfree asset. Why?
Credit risk.
Inflation risk.
Interest rate risk and the horizon effect.
'
Prof. Doron Avramov
180
Setting the Optimization in the Presence
of Riskfree Asset
The optimal solution is given by:
min
s.t or,
and the tangency portfolio takes the form:
The tangency portfolio is investing all the funds in
risky assets.
Vw w'
( )
p f
R w w = + ' 1 '
p
e
f
w R = + '
( )
( )
e
e
f
f
V
V
R V
R V
w

1
1
1
1
*
' '

=

=
'
Prof. Doron Avramov
181
The Investment Opportunities
However, the investor could select any point in the
line emerging from the riskfree rate and touching the
efficient frontier in point T.
The location depends on the attitude toward risk.
T
'
Prof. Doron Avramov
182
Fund Separation
Interestingly, all investors in the economy will mix
the tangency portfolio with a riskfree asset.
The mix depends on preferences.
But the proportion of risky assets will be equal
across the board.
One way to test the CAPM is indeed to examine
whether all investors hold the same proportions of
risky assets.
Obviously they dont!
'
Prof. Doron Avramov
183
General Equilibrium
The efficient frontier reflects the supply side.
What about the demand?
The demand side can be represented by a set of
indifference curves.
'
Prof. Doron Avramov
184
General Equilibrium
Why is the slope of indifference curve positive?
The equilibrium obtains when the indifference
curve tangents the efficient frontier
No riskfree asset:
A
'
Prof. Doron Avramov
185
General Equilibrium
With riskfree asset:
A
'
Prof. Doron Avramov
186
Maximize Utility / Certainty
Equivalent Return
In the presence of a riskfree security, the tangency
point can be found by maximizing a utility function
of the form
where is the relative risk aversion.
2
2
1
p p
U =
'
Prof. Doron Avramov
187
Maximize Utility / Certainty
Equivalent Return
Notice that utility is equal to expected return minus
a penalty factor.
The penalty factor positively depends on the risk
aversion (demand) and variance (supply).
'
Prof. Doron Avramov
188
Utility Maximization
Notice that here we get the same tangency portfolio
where reflects the fraction of weights invested in
risky assets. The rest is invested in the riskfree asset.
( )
e
e
e
f
V w
Vw
w
U
Vw w w R w U

1 *
1
0
'
2
1
'
=
= =
+ =
e
e
V
V
w

1
1
*
'

=
*
w
'
Prof. Doron Avramov
189
Mixing the Risky and Riskfree
Assets
So if then all funds (100%) are
invested in risky assets.
If then some fraction is invested in
riskfree asset.
If then the investor borrows money
to leverage his/her equity position.
e
V w
1 *
1

=
e
V
1
'

=
e
V
1
'

>
e
V
1
'

<
'
Prof. Doron Avramov
190
The Exponential Utility Function
The exponential utility function is of the form
where
Notice that
That is, the marginal utility is positive but
diminishes with an increasing wealth.
( ) ( ) W W U = exp 0 >
( ) ( )
( ) ( ) 0 exp ' '
0 exp '
2
< =
> =
W W U
W W U

'
Prof. Doron Avramov
191
The Exponential Utility Function
Indeed, the exponential preferences belong to the
class of constant absolute risk aversion (CARA).
For comparison, power preferences belong to the
class of constant relative risk aversion (CRRA).
= =
'
' '
U
U
ARA
'
Prof. Doron Avramov
192
Exponential: The Optimization
Mechanism
The investor maximizes the expected value of the
exponential utility where the decision variable is the
set of weights w and subject to the wealth evolution.
That is
( ) | |
t t
W W U E
1 +
w
max
t s.
( )
e
t f t t
R w R W W
1 1
' 1
+ +
+ + =
'
Prof. Doron Avramov
193
Mechanism
Let us assume that
Then
mean variance
( ) | | Vw w W w R W N W
t
e
f t t
' , ' 1 ~
2
1
+ +
+
( ) V N R
e e
t
, ~
1

+
'
Prof. Doron Avramov
194
Mechanism
It is known that for
( ) | |
|
\
|
+ =
2 2
2
1
exp exp
x x
a a ax E
( )
2
, ~
x x
N x
'
Prof. Doron Avramov
195
Mechanism
Thus,
( ) | | ( ) ( ) ( )
( )
( ) ( )
|
\
|
|
\
|
+ =
|
\
|
+ + + =
|
\
|
+ =
+ + +
Vw w W w W R W
Vw w W w R W
W VAR W E W E
t
e
t f t
t
e
f t
t t t
'
2
1
' exp 1 exp
'
2
1
' 1 exp
2
1
exp exp
2 2
1
2
1 1

'
Prof. Doron Avramov
196
Mechanism
Notice that
where is the relative risk aversion coefficient.
t
W =
'
Prof. Doron Avramov
197
Mechanism
So the investor ultimately maximizes
The optimal solution is .
The tangency portfolio is the same as before
e
V w
1 *
1

=
w
max
(
Vw w w
e
'
2
1
'
e
e
V
V
w

1
1
*
'

=
'
Prof. Doron Avramov
198
Mechanism
Conclusion:
The joint assumption of exponential utility and
normally distributed stock return leads to the well-
known mean variance solution.
'
Prof. Doron Avramov
199
Quadratic Preferences
The quadratic utility function is of the form
where
Notice that the first derivative is positive for
( )
2
2
W
b
W a W U + = 0 > b
0
1
2
2
< =
b
W
U
bW
W
U
W
b
1
<
'
Prof. Doron Avramov
200
The utility function looks like
( ) W U
W
b
1
'
Prof. Doron Avramov
201
It has a diminishing part which makes no sense
because we always prefer higher than lower wealth
Utility is thus restricted to the positive slope part
Notice that
The optimization formulation is given by
bW
bW
W
U
U
RRA
= = =
1 '
' '
w
max
t s.
( )
e
t ft t t
R w R W W
1 1
' 1
+ +
+ + =
( ) | |
t t
W W U E
1 +
'
Prof. Doron Avramov
202
Avramov and Chordia (2006 JFE) show that the
optimization could be formulated as
The solution takes the form
( ) w V w
R
w
e e
ft
e
1
'
'
1
2
1
'

+
|
\
|

w
max
e e
e
ft
V
V
R w

1 '
1
*
1
1
+
|
|
\
|
=
'
Prof. Doron Avramov
203
The tangency portfolio is
The only difference from the previously presented
competing specifications is the composition of risky
and riskfree assets.
e
e
V
V
w

1
1
*
'

=
'
Prof. Doron Avramov
204
The Sharpe Ratio of the Tangency
Portfolio
Notice that is actually the squared
Sharpe Ratio of the tangency portfolio.
Let us prove it
e e
V
1 '
( )
( )
2
1
1 '
* *' 2
1
1 '
*' *' *'
1
1
*
'
'
1
'
e
e e
TP
e
e e
e
f
e
f TP
e
e
V
V
w V w
V
V
w R w w R
V
V
w

= =
= = + =
=
'
Prof. Doron Avramov
205
The Sharpe Ratio of the Tangency
Portfolio
Thus,
Notice that TP is a subscript for the tangency
portfolio.
( )
e e
TP
f TP
TP
V
R
SR
1 '
2
2
2
=
=
'
Prof. Doron Avramov
206
Lecture Notes in
Testing Asset Pricing Models:
Time Series Perspective
'
Prof. Doron Avramov
207
Why Caring about Asset
Pricing Models?
An essential question that arises is why would both
academics and practitioners invest huge resources in
developing and testing asset pricing models.
It turns out that pricing models have crucial roles in
various applications in financial economics both
asset pricing as well as corporate finance.
In the following, I list five major applications.
'
Prof. Doron Avramov
208
1 Common Risk Factors
Pricing models characterize the risk profile of a firm.
In particular, systematic risk is no longer stock return
volatility rather it is the loadings on risk factors.
For instance, in the single factor CAPM the market
beta or the co-variation with the market
characterizes the systematic risk of the firm.
'
Prof. Doron Avramov
209
Likewise, in the single factor (C)CAPM the
consumption growth beta or the co-variation with
consumption growth characterizes the systematic risk
of the firm.
In the multi-factor Fama-French (FF) model there are
three sources of risk the market beta, the SMB beta,
and the HML beta.
'
Prof. Doron Avramov
210
Under FF, other things being equal (ceteris paribus), a
firm is riskier if its loading on SMB beta is higher.
Under FF, other things being equal (ceteris paribus), a
firm is riskier if its loading on HML beta is higher.
'
Prof. Doron Avramov
211
2 Moments for Asset
Allocation
Pricing models deliver moments for asset allocation.
For instance, the tangency portfolio takes on the form
e
e
TP
V
V
w

1
1
'

=
'
Prof. Doron Avramov
212
2 Asset Allocation
Under the CAPM, the vector of expected returns and the
covariance matrix are given by:
where is the covariance matrix of the residuals in the
time-series asset pricing regression.
We denoted by the residual covariance matrix in the
case wherein the off diagonal elements are zeroed out.
e
m
e
=
+ =
2
'
m
V
'
Prof. Doron Avramov
213
2 Asset Allocation
The corresponding quantities under the FF model are
where is the covariance matrix of the factors.
+ =
+ + =
'

F
HML HML SML SML
e
m MKT
e
V
F
'
Prof. Doron Avramov
214
3 Discount Factors
Expected return is the discount factor, commonly
denoted by k, in present value formulas in general
and firm evaluation in particular:
In practical applications, expected returns are
typically assumed to be constant over time, an
unrealistic assumption.
( )
=
+
=
T
t
t
t
k
CF
PV
1
1
'
Prof. Doron Avramov
215
3 Discount Factors
Indeed, thus far we have examined models with
constant beta and constant risk premiums
where is a K-vector of risk premiums.
When factors are return spreads the risk premium is
the mean of the factor.
Later we will consider models with time varying
factor loadings.
' =
e
'
Prof. Doron Avramov
216
4 Benchmarks
Factors in asset pricing models serve as benchmarks
for evaluating performance of active investments.
In particular, performance is the intercept (alpha) in
the time series regression of excess fund returns on a
set of benchmarks (typically four benchmarks in
mutual funds and more so in hedge funds):
t t WML t HML
t SMB
e
t MKT MKT
e
t
WML HML
SMB r r

+ + +
+ + =
,
'
Prof. Doron Avramov
217
5 Corporate Finance
There is a plethora of studies in corporate finance that
use asset pricing models to risk adjust asset returns.
Here are several examples:
o Examining the long run performance of IPO firm.
o Examining the long run performance of SEO firms.
o Analyzing abnormal performance of stocks going
through splits and reverse splits.
'
Prof. Doron Avramov
218
5 Corporate Finance
o Analyzing mergers and acquisitions
o Analyzing the impact of change in board of
directors.
o Studying the impact of corporate governance on
the cross section of average returns.
o Studying the long run impact of stock/bond
repurchase.
'
Prof. Doron Avramov
219
Time Series Tests
Time series tests are designated to examine the validity
of models in which factors are portfolio based, or
factors that are return spreads.
Example: the market factor is the return difference
between the market portfolio and the riskfree asset.
Consumption growth is not a return spread.
Thus, the consumption CAPM cannot be tested using
time series regressions, unless you form a factor
mimicking portfolio (FMP) for consumption growth.
'
Prof. Doron Avramov
220
Time Series Tests
FMP is a convex combination of asset returns
having the maximal correlation with consumption
growth.
The statistical time series tests have an appealing
economic interpretation. In particular:
Testing the CAPM amounts to testing whether the
market portfolio is the tangency portfolio.
Testing multi-factor models amounts to testing
whether some optimal combination of the factors is
the tangency portfolio.
'
Prof. Doron Avramov
221
Testing the CAPM
Run the time series regression:
The null hypothesis is:
Nt
e
mt N N
e
Nt
t
e
mt
e
t
r r
r r

+ + =
+ + =
M
1 1 1 1
0 :
2 1 0
= = = =
N
H K
'
Prof. Doron Avramov
222
Testing the CAPM
In the following, I will introduce four times series test
statistics:
WALD.
Likelihood Ratio.
GRS (Gibbons, Ross, and Shanken (1989)).
GMM.
'
Prof. Doron Avramov
223
The Distribution of .
Recall, is asset mispricing.
The time series regressions can be rewritten using a
vector form as:
Let us assume that
for t=1,2,3,,T
Let be the set of all
parameters.
1 1 1
1
1
1 NX
t
X
e
mt
NX
NX
NX
e
t
r r + + =
|
\
|

NxN
iid
Nx
t
N , 0 ~
1
( ) ( )' ' , ' , ' vech =

'
Prof. Doron Avramov
224
Under normality, the likelihood function for is
where c is the constant of integration (recall the
integral of a probability distribution function is
unity).
( ) ( ) ( )
(
e
mt
e
t
e
mt
e
t t
r r r r c L
1
'
2
1
2
1
exp
t
'
Prof. Doron Avramov
225
Moreover, the IID assumption suggests that
Taking the natural log from both sides yields
( )
( ) ( )
(
T
t
e
mt
e
t
e
mt
e
t
T
T
N
r r r r
c L
1
1
'
2
2 1
2
1
exp
, , ,

K
( ) ( ) ( ) ( )

T
t
e
mt
e
t
e
mt
e
t
r r r r
T
L
1
1
'
2
1
ln
2
ln
'
Prof. Doron Avramov
226
Asymptotically, we have
where
( ) ) ( , 0 ~
N
( )
1
2
'
ln

(

=

E
'
Prof. Doron Avramov
227
Let us estimate the parameters
( )
( )
( )
( )
( )
1
1
' 1 1
1
1
1
1
2
1
2
ln
ln
ln
=

=
+ =

T
t
t t
e
mt
T
t
e
mt
e
t
T
t
e
mt
e
t
T L
r r r
L
r r
L

'
Prof. Doron Avramov
228
Solving for the first order conditions yields
( )( )
( )
=
=

=
=
T
t
e
m
e
mt
T
t
e
m
e
mt
e e
t
e
m
e
r
r r
1
2
1

'
Prof. Doron Avramov
229
Moreover,
=
=
=
=
=
=
T
t
e
mt
e
m
T
t
e
t
e
T
t
t t
r
T
r
T
T
1
1
1
'
1
1

'
Prof. Doron Avramov
230
Recall our objective is to find the variance-
covariance matrix of .
Standard errors could be found using the information
matrix.
'
Prof. Doron Avramov
231
The information matrix is constructed as follows
( )
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
(
(
(
(
(
(
(
(
=
'
ln
,
'
ln
,
'
ln
'
ln
,
'
ln
,
'
ln
'
ln
,
'
ln
,
'
ln
2 2 2
2 2 2
2 2 2
L L L
L L L
L L L
E I

'
Prof. Doron Avramov
232
The Distribution of the Parameters
Try to establish yourself the information matrix.
Notice that and are independent of -
thus, your can ignore the second derivatives with
respect to in the information matrix if your
objective is to find the distribution of and .
If you aim to derive the distribution of then
focus on the bottom right block of the information
matrix.
'
Prof. Doron Avramov
233
We get:
Moreover,
|
|
\
|
(
(
|
|
\
|
+
2
1
1
, ~
m
e
m
T
N

|
|
\
|

2
1 1
, ~
m
T
N

( ) , 2 ~
T W T
'
Prof. Doron Avramov
234
Notice that W (x,y) stands for the Wishart
distribution with x=T-2 degrees of freedom and a
parameter matrix .
= y
'
Prof. Doron Avramov
235
The Wald Test
Recall, if
then
Here we test
where
The Wald statistic is
( ) , ~ N X
( ) ( ) ( ) N X X
2 1
~
:
0
:
1
0
H
H
( )
, 0 ~
0
N
H
( ) N
2 1
~
'
'
Prof. Doron Avramov
236
The Wald Test
which becomes:
where is the Sharpe ratio of the market factor.
2
1
1
1
2
1
'
'
1
m
m
e
m
R S
T T J
+
=
(
(
|
|
\
|
+ =

m
R S
'
Prof. Doron Avramov
237
Algorithm for Implementation
The algorithm for implementing the statistic is as
follows:
Run separate regressions for the test assets on the
common factor:
where
1 1 2
2
1
1
1
1 2
1
2
1
Tx
N
x
N
Tx
Tx
e
N
Tx x
Tx
Tx
e
t
X r
X r

+ =
+ =
M
| |' ,
, 1
, 1
1
2
i i i
e
mT
e
m
Tx
r
r
X
=
(
(
(
= M
'
Prof. Doron Avramov
238
Retain the estimated regression intercepts
and
Compute the residual covariance matrix
| |'
, ,
,

2 1 N
K =
(
=

N
TxN
, ,
1
K

= '
1
T
'
Prof. Doron Avramov
239
Compute the sample mean and the sample variance
of the factor.
Compute J1.
'
Prof. Doron Avramov
240
The Likelihood Ratio Test
We run the unrestricted and restricted specifications:
un:
res:
Using MLE, we get:
t
e
mt
e
t
r r + + =
* *
t
e
mt
e
t
r r + =
( ) , 0 ~ N
t
( )
* *
, 0 ~ N
t
'
Prof. Doron Avramov
241
( )
( )
|
|
\
|
+
=
=
=
=
=
, 1 ~

1 1
, ~
'

1
*
2 2
*
1
* * *
1
2
1
*
T W T
T
N
T
r
r r
m m
T
t
t t
T
t
e
mt
T
t
e
mt
e
t

'
Prof. Doron Avramov
242
where again W is the Wishart distribution, this time
with T-1 degrees of freedom.
'
Prof. Doron Avramov
243
The LR Test
Using some algebra, one can show that
Thus,
( ) ( ) | |
| | ( ) N T LR J
T
L L LR
2 *
2
* *
~
ln
ln 2
ln
ln
2
ln ln
= =
= =
|
|
\
|

|
\
|
= 1 exp
2
1
T
J
T J
|
\
|
+ = 1 ln
1
2
T
J
T J
'
Prof. Doron Avramov
244
GRS (1989)
Theorem: let
let where
and let A and X be independent then
( ) , 0 ~
1
N X
Nx
1 ,
1
~ '
1
+
+
N N
F X A X
N
N
( ) , ~
1
W A
Nx
N
'
Prof. Doron Avramov
245
GRS (1989)
In our context:
Then:
This is a finite-sample test.
( )
( )
2
, ~
, 0 ~
1
0
2
1
2
=
=
(
(
|
|
\
|
+ =

T
where
W T A
N T X
H
m
m
( ) 1 , ~
'
1
1
1
1
2
3

(
(
|
|
\
|
+
|
\
|

=

N T N F
N
N T
J
m
m

'
Prof. Doron Avramov
246
GMM
I will directly give the statistic without derivation:
where
( ) ( ) ) ( ~
' '
2
1
1
1 '
4
0
N R D S D R T J
H
T T T
=

( )
N
m
e
m
e
m
T
NxN
NxN
N
N Nx
I D
I R
(
(
+
=
(
=
2
2
2

,
, 1
0 ,

e
m
'
Prof. Doron Avramov
247
GMM
Assume no serial correlation but heteroskedasticity:
Under homoskedasticity and serially uncorrelated
moment conditions: J4=J1.
That is, the GMM statistic boils down to the WALD.
( )
| |
'
1
' '
, 1

1
e
mt t
T
t
t t t t T
r x
where
x x
T
S
=
=

=

'
Prof. Doron Avramov
248
The Multi-Factor Version of
Asset Pricing Tests
J2 follows as described earlier.
where is the mean vector of the factor based
return spreads.
( ) ( ) N T J
F r
F F F
Nx
t
Kx
t
NxK
Nx
Nx
e
t

~
'

1
1
1
1 '
1
1 1
1
1
+ =
+ + =
( )
( ) K N T N F F F
F
N
K N T
J

+

=
,
1
1
1 '
3
~
'

1
F
'
Prof. Doron Avramov
249
The Multi-Factor Version of
Asset Pricing Tests
is the variance covariance matrix of the factors.
For instance, considering the Fama-French model:
(
(
(
=
HML
SMB
e
m
F
(
(
(
(
=
2
, , ,
,
2
,
, ,
2

,
HML SMB HML m HML

HML SMB SMB m SMB
HML m SMB m m
F

'
Prof. Doron Avramov
250
The Current State of Asset Pricing
Models
o The CAPM has been rejected in asset pricing
tests.
o The Fama-French model is not a big success.
o Conditional versions of the CAPM and CCAPM
display some improvement.
o Should decision-makers abandon a rejected
CAPM?
'
Prof. Doron Avramov
251
Should a Rejected CAPM be
Abandoned?
Not necessarily!
Assume that expected stock return is given by
where
You estimate using the sample mean and CAPM:
'
Prof. Doron Avramov
252
( )
f m i f i i
R R + + =
0
i
( )
=
=
T
t
it i
R
T
1
1
1
( )
( )
f m i f i
R R + =
2
Mean Squared Error (MSE)
The quality of estimates is evaluated based on
the Mean Squared Error (MSE)
'
Prof. Doron Avramov
253
( ) ( )
( )
( ) ( )
( )
2
2 2
2
1 1
i i
i i
E MSE
E MSE

=
=
MSE, Bias, and Noise of
Estimates
Notice that
Of course, the sample mean is unbiased thus
However, the CAPM is rejected, thus
'
Prof. Doron Avramov
254
( ) estimate Var bias MSE + =
2
( ) ( )
( )
1 1
i
Var MSE =
( ) ( )
( )
2 2 2
i i
Var MSE + =
The Bias-Variance Tradeoff
It might be the case that is significantly
lower than - thus even when the CAPM is
rejected, still zeroing out could produce a smaller
mean square error.
'
Prof. Doron Avramov
255
( )
( )
2
i
Var
( )
( )
1
i
Var
i
When is the Rejected CAPM Superior?

'
Prof. Doron Avramov
256
( )
( ) ( ) ( ) ( ) | |
( )
( ) ( )
e
m i i
i m i i i
Var Var
R
T
R
T
Var

1 1
2
2 2 2 2 1
=
+ = =
Using variance decomposition
'
Prof. Doron Avramov
257
( )
( ) ( ) | | ( ) ( )
( )
( )
( ) ( )
( ) | |
(
(
|
|
\
|
=
+ =
+
(
(
|
|
\
|
= +
(
=
+ = =
2
2
2 2 2
2
2
2
2
2
2
2
2
2
m
e
m
m
m i i m
m
i
m
e
m
i
e
m i
m
i
e
m
e
m
e
m i
e
m
e
m i
e
m i i
E SR
where
SR
T
T
E
T
Var
T
E
E Var Var E Var Var

Then
where is the R squared in the market regression.
Since is small -- the ratio of the variance
estimates is smaller than 1.
'
Prof. Doron Avramov
258
( )
( )
( )
( )
( )
( )
( )
2 2
2
2 2 2
1
2
1
R R SR
R
SR
Var
Var
m
i
m i i m
i
i
+ =
+
=
2
R
m
SR
Example
Let
For what values of it is sill preferred to use
the CAPM?
Find such that the MSE of the CAPM is smaller.
'
Prof. Doron Avramov
259
( )
3 . 0
05 . 0
01 . 0
2
2
2
=
=
|
|
\
|
=
R
E
R
m
e
m
i
0
i
Example
'
Prof. Doron Avramov
260
( )
( )
( )
( )
% 528 . 1 01528 . 0
665 . 0 01 . 0
60
1
1
01 . 0
60
1
3 . 0 7 . 0 05 . 0
1 1
2
2
1
2
2 2
1
2
= <
<
<
+ +
< + + =
i
i
i
i
m
MSE
R R SR
MSE
MSE
Economic versus Statistical

Factors
Factors such as the market portfolio, SMB, HML,
WML, liquidity, credit risk, as well as bond based
factors are pre-specified.
Such factors are considered to be economically
based.
For instance, Fama and French argue that SMB and
HML factors are proxying for underlying state
variables in the economy.
'
Prof. Doron Avramov
261
Economic versus Statistical
Factors
Statistical factors are derived using econometric
procedures on the covariance matrix of stock return.
Two prominent methods are the factor analysis and
the principal component analysis (PCA).
Such methods are used to extract common factors.
The first factor typically has a strong (about 96%)
correlation with the market portfolio.
Later, I will explain the PCA.
'
Prof. Doron Avramov
262
The Economics of Time Series
Test Statistics
Let us summarize the first three test statistics:
|
\
|
+ = 1 ln
1
2
T
J
T J
2
1
1
'
m
R S
T J
+
=

2
1
3
'
1
m
R S N
N T
J
+

=

'
Prof. Doron Avramov
263
The Economics of the Time
Series Tests
The J4 statistic, the GMM based asset pricing test, is
actually a Wald test, just like J1, except that the
covariance matrix of asset mispricing takes account
of heteroskedasticity and often even potential serial
correlation.
Notice that all test statistics depend on the quantity

'
'
Prof. Doron Avramov
264
The Economics of the Time
Series Tests
GRS show that this quantity has a very insightful
representation.
Let us provide the steps.
'
Prof. Doron Avramov
265
Consider an investment universe that consists of N+1
assets - the N test assets as well as the market
portfolio.
The expected return vector of the N+1 assets is given
by
where is the estimated expected excess return on
the market portfolio and is the estimated expected
excess return on the N test assets.
( )
' '
1
1 1
1 1
(
=
+
xN
e
x
e
m
x N

e
m
Understanding the Quantity .
'
'
Prof. Doron Avramov
266
The variance covariance matrix of the N+1 assets is
given by
where
is the estimated variance of the market factor.
is the N-vector of market loadings and is the
covariance matrix of the N test assets.
( ) ( )
(
(
=
+ +
V
m
m m
N x N
'
2
2 2
1 1

2
'
'
Prof. Doron Avramov
267
Notice that the covariance matrix of the N test assets
is
The squared tangency portfolio of the N+1 assets is
+ =
'

2
m
V

'

1 2
=
TP
R S
'
'
Prof. Doron Avramov
268
Notice also that the inverse of the covariance matrix
is
Thus, the squared Sharpe ratio of the tangency
portfolio could be represented as
( )
(
(

+
=

'
'
1
1 1
1
2
1 m
1
'
'
Prof. Doron Avramov
269
( ) ( )
2 2 1
1 2 2
1
'
2
2

'
'
m TP
m TP
e
m
e e
m
e
m
e
m
TP
R S R S
or
R S R S
R S
=
+ =
(
+
|
|
\
|
=
'
'
Prof. Doron Avramov
270
In words, the quantity is the difference
between the squared Sharpe ratio based on the N+1
assets and the squared Sharpe ratio of the market
portfolio.
If the CAPM is correct then these two Sharpe ratios
are identical in population, but not identical in sample
due to estimation errors.

'
'
'
Prof. Doron Avramov
271
The test statistic examines how close the two sample
Sharpe ratios are.
Under the CAPM, the extra N test assets do not add
anything to improving the risk return tradeoff.
The geometric description of is given in
the next slide.

'
'
'
Prof. Doron Avramov
272
2
2
2
1
1
'
M
TP
2
Rf
r
'
'
Prof. Doron Avramov
273
So we can rewrite the previously derived test
statistics as
'
( )
( ) 1 , ~
1

1
~
1

2
2 2
3
2
2
2 2
1

+

=
+

=
N T N F
R S
R S R S
N
N T
J
N
R S
R S R S
T J
m
m TP
m
m TP
'
Prof. Doron Avramov
274
Asset Pricing Models with
Time Varying Beta
We consider for simplicity only the one factor
CAPM extensions follow the same vein.
Let us model beta variation with the lagged dividend
yield or any other macro variable again for
simplicity we consider only one information,
predictive, macro, or lagged variable.
'
Prof. Doron Avramov
275
Asset Pricing Models with
Time Varying Beta
Typically, the set of predictive variables contains the
dividend yield, the term spread, the default spread,
the yield on a T-bill, inflation, lagged market return,
market volatility, market illiquidity, etc.
'
Prof. Doron Avramov
276
Conditional Models
Here is a conditional asset pricing specification:
0 ) , cov(
) | (
1
1
1 1 0
=
=
+ + =
+ =
+ + =
t it
e
m t
e
mt
t t t
t i i it
it
e
mt it i
e
it
z r E
bz a z
z
r r

'
Prof. Doron Avramov
277
Conditional Asset Pricing
Models
Substituting beta back into the asset pricing equation
yields.
Interestingly, the one factor conditional CAPM
becomes a two factor unconditional model the first
factor is the market portfolio, while the second is the
interaction of the market with the lagged variable.
it t
e
mt i
e
mt i i
e
it
z r r r + + + =
1 1 0
'
Prof. Doron Avramov
278
Conditional Asset Pricing
Models
You can use the statistics J1 through J4 to test such
models.
If we have K factors and M predictive variables then
the K-conditional factor model becomes a K+KM-
unconditional factor model.
If you only scale the market beta, as is typically the
case, we have an M+K unconditional factor model.
'
Prof. Doron Avramov
279
Conditional Moments
Suppose you are at time t what is the discount factor
for time t+2?
2 1 2 1 2 0 2 + + + + +
+ + + =
it t
e
mt i
e
mt i i
e
it
z r r r
( )
2 1 2 1 2 0 + + + +
+ + + + + =
it t t
e
mt i
e
mt i i
bz a r r
2 1 2 1 2 1 2 1 2 0 + + + + + +
+ + + + + =
it t
e
mt i t
e
mt i
e
mt i
e
mt i i
r z br r a r
'
Prof. Doron Avramov
280
Conditional Moments
| |
t
e
m i
e
m i
e
m i i t
e
it
z b a z r E
1 1 0 2
+ + + =
+
( )
t
e
m i
e
m i i i
z b a
1 1 0
+ + + =
Notice that here the discount factor, or the conditional
expected return, is no longer constant through time.
Rather, it varies with the macro variable.
'
Prof. Doron Avramov
281
Conditional Moments
Could you derive a general formula in particular
you are at time t what is the expected return for time
t+T as a function of the model parameters as well as
?
t
z
'
Prof. Doron Avramov
282
Conditional Moments
Next, the conditional covariance matrix the
covariance at time t+1 given is given by
where and are the N-asset versions of
and respectively.
| | ( )( ) + + + =
+
2
'
1 0 1 0 1 m t t t
e
t
z z z r V
0
0 i
1 i
t
z
'
Prof. Doron Avramov
283
Conditional Moments
Could you derive a general formula in particular
you are at time t what is the conditional covariance
matrix for time t+T as a function of the model
parameters as well as ?
Could you derive general expressions for the
conditional moments of cumulative return?
t
z
'
Prof. Doron Avramov
284
Conditional versus
Unconditional Models
There are different ways to model beta variation.
Here we used lagged predictive variables; other
applications include using firm level variables such
as size and book market to scale beta as well as
modeling beta as an autoregressive process.
'
Prof. Doron Avramov
285
Conditional versus
Unconditional Models
You can also model time variation in the risk
premiums in addition to or instead of beta variations.
Asset pricing tests show that conditional models
typically outperform their unconditional
counterparts.
'
Prof. Doron Avramov
286
Different Ways to Model Beta
Variation
The base case: beta is constant, or time invariant.
Case II: beta varies with macro conditions
t t t
t i i it
bz a z
z

+ + =
+ =
1
1 1 0
'
Prof. Doron Avramov
287
Variation
Case III: beta varies with firm-level size and the
book-to-market ratio
Case IV: beta is some function of both macro and
firm-level variables as well as their interactions:
1 , 2 1 , 1 0
+ + =
t i i t i i i it
bm size
( )
1 , 1 , 1
, ,

=
t i t i t it
bm size z f
'
Prof. Doron Avramov
288
Variation
Case V: beta follows an auto-regressive AR(1) process
it t i it
v b a + + =
1 ,

'
Prof. Doron Avramov
289
Single vs. Multiple Factors
Notice that we describe the case of a single factor
single macro variable.
We can expand the specification to include more
factors and more macro and firm-level variables.
Even if we expand the number of factors it is
common to model variation only in the market beta,
while the other risk loadings are constant.
Some scholars model time variations in all factor
loadings.
'
Prof. Doron Avramov
290
Testing Conditional Models
You can implement the J1-J4 test statistics only to
those cases where beta is either constant or it varies
with macro variables.
Those specifications involving firm-level
characteristics require cross sectional tests.
The last specification (AR(1)) requires filtering
methods involving state space representations.
'
Prof. Doron Avramov
291
Lecture Notes in
GMVP and Tracking Error
Volatility
'
Prof. Doron Avramov
292
GMVP
Of particular interest to academics and practitioners is
the Global Minimum Volatility Portfolio.
For two distinct reasons:
1. No need to estimate the notoriously difficult to
estimate .
2. Low volatility stocks have been found to
outperform high volatility stocks.
'
Prof. Doron Avramov
293
GMVP Optimization
min
s.t
Solution:
No analytical solution in the presence of portfolio
constraints such as no short selling.
Vw w'
1 ' = w

1
1
'

=
V
V
w
GMVP
'
Prof. Doron Avramov
294
GMVP Optimization
Ex ante, the GMVP is the lowest volatility portfolio
among all efficient portfolios.
Ex ante, it is also the lowest mean portfolio, but ex
post it performs reasonably well in delivering high
payoffs.
'
Prof. Doron Avramov
295
The Tracking Error Volatility
(TEV) Portfolio
Actively managed funds are often evaluated based on
their ability to achieve high return subject to some
constraint on their Tracking Error Volatility (TEV).
In that context, a managed portfolio can be
decomposed into both passive and active components.
TEV is the volatility of the active component.
'
Prof. Doron Avramov
296
The Tracking Error Volatility
(TEV) Portfolio
The passive component is the benchmark portfolio.
The benchmark portfolio changes with the
investment objective.
For instance, if you invest in TA 100 stocks the
proper benchmark would be the TA100 index.
If you invest in corporate bonds traded in TASE the
benchmark could be TelBond 60.
'
Prof. Doron Avramov
297
Tracking Error Volatility: The
Benchmark and Active Portfolios
Let q be the vector of weights of the benchmark
portfolio.
Then the expected return and variance of the
benchmark portfolio are given by
where, as usual, and V are the vector of expected
return and the covariance matrix of stock returns.
Vq q
q
B
B
'
'
2
=
=

'
Prof. Doron Avramov
298
Tracking Error Volatility: The
Benchmark and Active Portfolios
The matrix can be estimated in different methods
most prominent of which will be discussed here.
The active fund manager attempts to outperform this
benchmark.
Let x be the vector of deviations from the benchmark,
or the active part of the managed portfolio.
Of course, the sum of all the components of x, by
construction, must be equal to zero.
V
'
Prof. Doron Avramov
299
The Mathematics of TEV
So the fund manager invests w=q+x in stocks, q is the
passive part of the portfolio and x is the active part.
Notice that is the tracking error variance.
Also notice that the expected return and volatility of
the chosen portfolio are
Vx x'
2
=
2 2 2
' 2
'

+ + =
+ =
Vx w
x
B p
B p
'
Prof. Doron Avramov
300
The optimization problem is formulated as
=
=
Vx x
x
t s
x
x
'
0 '
. .
' max
'
Prof. Doron Avramov
301
The resulting active part of the portfolio, x, is given by
where
|
\
|
=

c
a
V
e
x
1
e c b a
c V
a V
b V
=
=
=
=
/
'
'
'
2
1
1
1

'
Prof. Doron Avramov
302
Questions
Is the sum of the x components equal to zero?
Prove that if the benchmark is the GMVP then all x
components are equal to zero.
What if the benchmark is another efficient portfolio
does this result still hold?
Does the active part of the portfolio depend on the
benchmark composition?
'
Prof. Doron Avramov
303
Caveats about the Minimum TEV
Portfolio
Richard Roll (distinguished UCLA professor) points
out that the solution is independent on the benchmark.
Put differently, the active part of the portfolio x is
totally independent of the passive part q.
Of course, the overall portfolio q+x is impacted by q.
The unexpected result is that the active manager pays
no attention to the assigned benchmark. So it does not
really matter if the benchmark is S&P or any other
index.
'
Prof. Doron Avramov
304
TEV with total Volatility
Constraint (based on Jorion an Expert in
Risk Management)
Given the drawbacks underlying the TEV portfolio we
add one more constraint on the total portfolio volatility.
The derived active portfolio displays two advantages.
First, its composition does depend on the benchmark.
Second, the systematic volatility of the portfolio is
controlled by the investor.
'
Prof. Doron Avramov
305
TEV with total volatility
constraint
The optimization is formulated as
Home assignment: derive the optimal solution.
2
) ( )' (
'
0 '
. .
' max
p
x
x q V x q
Vx x
x
t s
x
= + +
=
=
'
Prof. Doron Avramov
306
Lecture Notes in
Estimating the Large Scale
Covariance Matrix
'
Prof. Doron Avramov
307
Estimating the Covariance
Matrix:
There are various applications in financial economics
which use the covariance matrix as an essential input.
The Global Minimum Variance Portfolio, the
minimum tracking error volatility portfolio, the mean
variance efficient frontier, and asset pricing tests are
good examples.
In what follows I will present the most prominent
estimation methods of the covariance matrix.
'
Prof. Doron Avramov
308
The Sample Covariance Matrix
(Denoted S)
This method uses sample estimates.
Need to estimate N(N+1)/2 parameters which is a lot.
You can use excel to estimate all variances and co-
variances which is tedious and inefficient.
Here is a much more efficient method.
Consider T monthly returns on N risky assets.
We can display those returns in a T by N matrix R.
'
Prof. Doron Avramov
309
Estimate the mean return of the N assets and denote
the N-vector of the mean estimates by .
Next, compute the deviations of the return
observations from their sample means:
where is a T vector of ones. Then the sample
covariance matrix is estimated as

= '
T
R E
T

= E E
T
S '
1
'
Prof. Doron Avramov
310
The Sample Covariance Matrix
(Denoted S)
The Equal Correlation Based
Covariance Matrix (Denoted F):
Estimate all N(N-1)/2 pair-wise correlations between
any two securities and take the average.
Let be that average correlation, let be the
estimated variance of asset i, and let be the
estimated variance of asset j, both estimates are the i-th
and j-th elements of the diagonal of S.
Then the matrix F follows as
ii
s
jj
s
ii
s i i F = ) , (
jj ii
s s j i F

= ) , (
'
Prof. Doron Avramov
311
The Factor Based Covariance
Matrix:
Consider the time series regression
where is an N vector of returns at time t and
is a set of K factors. Factor means are denoted by
Notice that the mean return is given by
t t t
e F r + + =
t
r
F
+ =
t
F
F
'
Prof. Doron Avramov
312
Matrix
Thus deviations from the means are given by
The factor based covariance matrix is estimated by
Here, is an N by K matrix of factor loadings and
is a diagonal matrix with each element represents the
idio syncratic variance of each of the assets.
t F t t
e F r + = ) (

+ = '
FF
V

'
Prof. Doron Avramov
313
Matrix Number of Parameters
This procedure requires the estimation of NK betas as
well as K variances of the factors, K(K-1)/2
correlations of those factors, and N firm specific
variances.
Overall, you need to estimate NK+K+K(K-1)/2+N
parameters, which is considerably less than N(N+1)/2
since K is much smaller than N.
For instance, using a single factor model the number
of parameters to be estimated is only 2N+1.
'
Prof. Doron Avramov
314
Steps for Estimating the Factor
Based Covariance Matrix
1) Run the MULTIVARIATE regression of stock
returns on asset pricing factors
where
( ) ( ) N T N K K T N T
N K
K T N T N T
E F E F R
+ +

+ = + + =
1 1 1 1
' ' '
(
| |
| |

,
,
=
= F F
T
(
'
Prof. Doron Avramov
315
2) Estimate
and retain only.
( ) | | ( ) | |
' ' ' '
1
'
1
= = =

F F F R R F F F
( ( ( ( ( (
)
'
Prof. Doron Avramov
316
3) Estimate the covariance matrix of E
4) Let be a diagonal matrix with the
component being equal to the component
of .
N T T N N N
E E
T

= '
1
( ) th i i ,
( ) th i i ,
'
Prof. Doron Avramov
317
5) Compute:
where is the mean return of the factors.
6) Estimate:
K
f
T K T K T
F V

=
1
'
1
K T T K
K K
F
V V
T

= '
1
'
Prof. Doron Avramov
318
7)
This is the estimated covariance matrix of stock
returns.
8) Notice there is no need to run N individual
regressions! Use multivariate specifications.
+ =
N N
N K
K K
F
K N
N N
V '

'
Prof. Doron Avramov
319
A Shrinkage Approach - Based on
a Paper by Ledoit and Wolf (LW)
There is a well perceived paper (among Wall Street
quants) by LW demonstrating an alternative approach
to estimating the covariance matrix.
It had been claimed to deliver superior performance in
reducing tracking errors relative to benchmarks as well
as producing higher Sharpe ratios.
Here are the formal details.
'
Prof. Doron Avramov
320
A Shrinkage Approach - Based on
a Paper by Ledoit and Wolf (LW)
Let S be the sample covariance matrix, let F be the
equal correlation based covariance matrix, and let be
the shrinkage intensity. S and F were derived earlier.
The operational shrinkage estimator of the covariance
matrix is given by
Notice that F is the shrinkage target.
S F V + =

*) 1 ( *
_

'
Prof. Doron Avramov
321
The Shrinkage Intensity
LW propose the following shrinkage intensity, based on
optimization:
where T is the sample size and k is given as
and where with being the (i,j)
component of S and is the (i,j)
component of F.
(
)
`
1 , min , 0 max *
T
k

= k
= =
=
N
i
N
j
ij ij
s f
1 1
2
) (
ij
s
ij
f
322
The Shrinkage Intensity
and with and being the time t return on asset i
and the time-series average of return on asset i,
respectively.
= =
=
N
i
N
j
ij
1 1

=
=
T
t
ij j jt
i
it ij
s r r r r
T
1
2
_ _
} ) )( {(
1

= = =
|
|
\
|
+ + =
N
i
N
i
N
i j j
ij jj
jj
ii
ij ii
ii
jj
ii
s
s
s
s
1 1 , 1
, ,
_
2

=
)
`

)
`
=
T
t
ij j jt i it ii i it ij ii
s r r r r s r r
T
1
_ _
2
_
,
) )( ( ) (
1
=
)
`

)
`
=
T
t
ij j jt i it jj
j
jt ij jj
s r r r r s r r
T
1
_ _
2
_
,
) )( ( ) (
1
it
r
_
i
r
'
Prof. Doron Avramov
323
The Shrinkage Intensity A
Nave Method
If you get overwhelmed by the derivation of the
shrinkage intensity it would still be useful to use a
nave shrinkage approach, which often even works
better. For instance, you can take equal weights:
S F V
2
1
2
1
_
+ =
'
Prof. Doron Avramov
324
Backtesting
We have proposed several methods for estimating the
covariance matrix.
Which one dominates?
We can backtest all specifications.
That is, we can run a horse race across the various
models searching for the best performer.
There are two primary methods for backtesting
rolling versus recursive schemes.
'
Prof. Doron Avramov
325
The Rolling Scheme
You define the first estimation window.
It is well perceived to use the first 60 sample
observations as the first estimation window.
Based on those 60 observations derive the GMVP
under each of the following methods:
'
Prof. Doron Avramov
326
Competing Covariance Estimates
The sample based covariance matrix
The equal correlation based covariance matrix
Factor model using the market as the only factor
Factor model using the Fama French three factors
Factor model using the Fama French plus Momentum
factors
The LW covariance matrix either the full or the
nave method.
'
Prof. Doron Avramov
327
Out of Sample Returns
Then given the GMVPs compute the actual returns on
each of the derived strategies.
For instance, if the derived strategy at time t is
then the realized return at time t+1 would be
where is the realized return at time t+1 on all the
N investable assets.
t
w
1 1 ,
'
+ +
=
t t t p
R w R
1 + t
R
'
Prof. Doron Avramov
328
Suppose you rebalance every six months derive the
out of sample returns also for the following 5 months
Then at time t+6 you re-derive the GMVPs.
2 2 ,
'
+ +
=
t t t p
R w R
3 3 ,
'
+ +
=
t t t p
R w R
4 4 ,
'
+ +
=
t t t p
R w R
5 5 ,
'
+ +
=
t t t p
R w R
6 6 ,
'
+ +
=
t t t p
R w R
'
Prof. Doron Avramov
329
The Recursive Scheme
A recursive scheme is using an expanding window.
That is, you first estimate the GMVPs based on the
first 60 observations, then based on 66 observations,
and so on, while in the rolling scheme you always use
the last 60 observations.
Pros: the recursive scheme uses more observations.
Cons: since the covariance matrix may be time
varying perhaps you better drop initial observations.
'
Prof. Doron Avramov
330
So you generate out of sample returns on each of
the strategies starting from time t+1 till the end of
sample, which we typically denote by T.
Next, you can analyze the out of sample returns.
For instance, you can form the table on the next
page and examine which specification has been able
to deliver the best performance.
'
Prof. Doron Avramov
331
Rolling Scheme Recursive Scheme
S F MKT FF FF+MOM LW S F MKT FF FF+MOM LW
Mean
STD
SR
SP (5%)
alpha
IR
'
Prof. Doron Avramov
332
In the above table:
Mean is the simple mean of the out of sample
returns
STD is the volatility of those returns
SR is the associated Sharpe ratio obtained by
dividing the difference between the mean return and
the mean risk free rate by STD.
SP is the shortfall probability with a 5% threshold
applied to the monthly returns.
'
Prof. Doron Avramov
333
In the above table:
alpha is the intercept in the regression of out of
sample EXCESS returns on the contemporaneous
market factor (market return minus the riskfree rate).
IR is the information ratio obtained by dividing
alpha by the standard deviation of the regression
error, not the STD above.
Of course, higher SR, higher alpha, higher IR are
associated with better performance.
'
Prof. Doron Avramov
334
Lecture Notes in
Principal Component Analysis
(PCA)
'
Prof. Doron Avramov
335
PCA
The aim is to extract K common factors to
summarize the information of a panel of rank N.
In particular, we have a TN panel of stock returns
where T is the time dimension and N (<T) is the
number of firms of course K<< N
The PCA is an operation on the sample covariance
matrix of stock returns.
| |
N
N T
R R R , ,
1
K =
'
Prof. Doron Avramov
336
Principal Component Analysis
(PCA)
To understand the PCA let us master the notion of
eigen-vector and eigen-value.
An eigen-vector of a squared matrix is a non zero
vector which, when multiplied by the matrix, yields
a vector parallel to the origin.
'
Prof. Doron Avramov
337
Principal Component Analysis (PCA)
To illustrate:
Let
Hence
is the first eigen-value. is the
corresponding eigen vector
(
=
2 , 1
1 , 2
A
1 1
3
3
3
x Ax =
(
=
3
1
=
(
=
1
1
1
x
'
Prof. Doron Avramov
338
1
x
Eigen Values and Eigen Vectors
How to find eigen values and eigen vectors?
Set det(B)=0 and solve
(
=
(
= =
2 , 1
1 , 2
, 0
0 ,
A I A B
( )
1 , 3
0 3 4 det
2 1
2
= =
= + =

B
'
Prof. Doron Avramov
339
Principal Component Analysis (PCA)
Let ,
hence
is the second eigen-value. is the
corresponding eigen vector.
(
=
2 , 1
1 , 2
A
1
2
=
(
=
3
3
2
x
2 2
1
3
3
x Ax =
(
=
'
Prof. Doron Avramov
340
2
x
Eigen Values and Eigen Vectors
Finding the eigen vector
1 , 1
3
2 , 1
1 , 2
= =
(
=
(
y x
y
x
y
x
3 , 1
1
2 , 1
1 , 2
= =
(
=
(
y x
y
x
y
x
'
Prof. Doron Avramov
341
PCA
You have returns on N stocks for T periods
| |
R R
T
V
R R R R
let
R R R R
R R
N
N N N
T
N
T
~
'
~
1
~
, ,
~
,
~ ~
~ ~
, ,
2 1
1 1 1
1 1
1
=
=
= =

K
K
K

'
Prof. Doron Avramov
342
PCA
Extract K-eigen vectors corresponding to the largest
K-eigen values.
Each of the eigen vector is an N by 1 vector.
The extraction mechanism is as follows.
The first eigen vector is obtained as
1
'
1
w V w
1
1
'
1
= w w
1
max
w
t s.
'
Prof. Doron Avramov
343
PCA
is an eigen vector since
is therefore the highest eigen value.
1
w
1
'
1 1
1 1 1
w V w
Moreover
w w V
=
=
'
Prof. Doron Avramov
344
PCA
Extracting the second eigen vector
t s.
2
max
w
2
'
2
w V w
0
1
2
'
1
2
'
2
=
=
w w
w w
'
Prof. Doron Avramov
345
PCA
The optimization yields:
The second eigen value is smaller than the first due
to the presence of one extra constraint in the
optimization the orthogonality constraint.
1 2
'
2 2
2 2 2

< =
=
w V w
w w V
'
Prof. Doron Avramov
346
PCA
The K-th eigen vector is derived as
t s.
max
K K
w V w
'
0
0
0
1
1
'
2
'
1
'
'
=
=
=
=
K K
K
K
K K
w w
w w
w w
w w
M
'
Prof. Doron Avramov
347
PCA
The optimization yields:
1 2 1
'

< < < < =
=
K
K K K K
K K K
w V w
w w V
'
Prof. Doron Avramov
348
PCA
Then - each of the K-eigen vectors delivers a unique
asset pricing factor.
Simply, multiply excess stock returns by the eigen
vectors:
1 1
1
1
1
1
=
=
N
K
N T
e
T
K
N
N T
e
T
w R F
w R F
M
'
Prof. Doron Avramov
349
PCA
Recall, the basic idea here is to replace the original
set of N variables with a lower dimensional set of
K-factors (K<<N).
The contribution of the j-th eigen vector to explain
the covariance matrix of stock returns is
=
N
i
i
j
1

'
Prof. Doron Avramov
350
PCA
Typically the first three eigen vectors explain over
and above 95% of the covariance matrix.
What does it mean to explain the covariance
matrix? Coming up soon!
'
Prof. Doron Avramov
351
Understanding the PCA:
Digging Deeper
The covariance matrix can be decomposed as
| |
(
(
(
(
(
(
=
'
'
, , 0
0 ,
, ,
1
, 1
1
N
N
w
w
w w V M
K
M
K
K
O
'
Prof. Doron Avramov
352
Digging Deeper
If some of the are either zero or negative
the covariance matrix is not properly defined -- it is
not positive definite.
In fixed income analysis there are three prominent
eigen vectors, or three factors.
The first factor stands for the term structure level,
the second for the term structure slope, and the third
for the curvature of the term structure.
s
'
Prof. Doron Avramov
353
Digging Deeper
In equity analysis, the first few (up to three)
principal components are prominent.
Others are around zero.
The attempt is to replace the sample covariance
matrix by the matrix which mostly summarizes
the information in the sample covariance matrix.
V
~
'
Prof. Doron Avramov
354
Digging Deeper
The matrix is given by
Of course, the dimension of is N by N.
However, its rank is K, thus the matrix is not
invertible.
| |
(
(
(
(
(
(
'
'
, , 0
0 ,
, ,
~
1
, 1
1
k
k
N N
w
w
w w V M
K
M
K
K
O
V
~
V
~
'
Prof. Doron Avramov
355
This is the same as asking: what does it mean to
explain the sample covariance matrix?
Is close to ? V
~
V
'
Prof. Doron Avramov
356
Let us represent returns as
M
t Kt K t t t
f f f r
1 1 2 12 1 11 1 1
+ + + + + = K
t
r
1
~
Nt Kt NK t N t N N Nt
f f f r + + + + + = K
2 2 1 1
Nt
r
~
T t , , 1K =
T t , , 1K =
Is close to ? V
~
V
'
Prof. Doron Avramov
357
where are the principal component based
factors and are the exposures of firm i to
those factors.
Kt t
f f K
1
iK i
K
1
Is close to ? V
~
V
'
Prof. Doron Avramov
358
is closed enough to - if
1. The variances of the residuals cannot be
dramatically reduced by adding more factors.
2. The pairwise cross-section correlations of the
residuals cannot be considerably reduced by
adding more factors.
Is close to ? V
~
V
V
~
V
'
Prof. Doron Avramov
359
The Contribution of the PC-S
to Explain Portfolio Variation
Let be the exposures of firm i to
the K common factors.
Let be an N-vector of portfolio weights:
Recall, R is a TN matrix of stock returns.
| |
'
1
1
, ,
iK i
K
i
K =
p
w
| |
Np p p p
w w w w , , ,
2 1
K =
'
'
Prof. Doron Avramov
360
The portfolios rate of return is
Moreover, the portfolio time t return is given by
1 1
=
N
p
N T
T
p
w R R
1
1
'
2 2 1 1
+ + + =
N
t
N
p Nt Np t p t p pt
r w r w r w r w R K
'
Prof. Doron Avramov
361
We can approximate the portfolios rate of return as:
( )
Kt K t t p pt
f f f w R
1 2 12 1 11 1
~
+ + + = K
( )
( )
Kt NK t N t N Np
Kt K t t p
f f f w
f f f w

+ + + +
+
+ + + +
K
K K K K K K K K K K K K K
K
2 2 1 1
2 2 22 1 21 2
'
Prof. Doron Avramov
362
Thus,
1
1
'
1
2
1
'
2
1
1
1
'
1
2 2 1 1
~
=
=
=
+ + + =
N
K
N
p K
N
N
p
N
N
p
Kt K t t pt
w
w
w
where
f f f R

M
K
'
Prof. Doron Avramov
363
Notice that
are the K loadings on the common factors,
or they are the risk exposures, while
are the K-realizations of the factors at time t.
K
K
1
Kt t
f f K
1
'
Prof. Doron Avramov
364
Explain the Portfolio Variance
The actual variation is
The approximated variation is
Both quantities are quite similar.
p p pt
w V w R
) (
' 2
=
) var( ) var( ) var( )
~
(
2
2
2
2 1
2
1
2
Kt K t t pt
f f f R + + + = K
'
Prof. Doron Avramov
365
Explaining the Portfolio
Variance
The contribution of the i-th PC to the overall portfolios
variance is:
=
k
j
j j
i i
f
f
1
2
2
) var(
) var(
'
Prof. Doron Avramov
366
Asy. PCA: What if N>T?
Then create a TT matrix and extract K eigen
vectors those eigen vectors are the factors
and is the T-vector of cross sectional (across-
stocks) mean of returns.
'
'
1
1
N
N
T
N T N T
R E
where
E E
N
V

=
=

'
Prof. Doron Avramov
367
Other Applications
PCA can be implemented in a host of other
applications.
For instance, you want to predict economic growth
with many predictors, say Mwhere M is large.
You have a panel of TM predictors, where T is the
time-series dimension.
'
Prof. Doron Avramov
368
Other Applications
In a matrix form
where is the m-the predictor realized at time t.
(
(
(
TM T T
M
M T
Z Z Z
Z Z Z
Z
, , ,
, , ,
2 1
1 12 11
K
M
K
tm
Z
'
Prof. Doron Avramov
369
Other Applications
If T>M compute the covariance matrix of Z then
extract K principal components such that you
summarize the M-dimension of the predictors with a
smaller subspace of order K<<M.
You extract eigen vectors.
1 1
1
, ,
M
K
M
w w K
'
Prof. Doron Avramov
370
Other Applications
Then you construct K predictors:
1 1
1
1
1
1
=
=
M
K
M T
T
K
M
M T
T
w Z Z
w Z Z
M
'
Prof. Doron Avramov
371
What if M>T?
If M>T then you extract K eigen vectors from the
TT matrix.
In this case, the K-predictors are the extracted K
eigen vectors.
Be careful of a look-ahead bias in real time
prediction.
'
Prof. Doron Avramov
372
The Number of Factors in PCA
An open question is: how many factors/eigen vectors
should be extracted?
Here is a good mechanism: set - the highest
number of factors.
Run the following multivariate regression for
'
Prof. Doron Avramov
373
max
K
max
, , 2 , 1 K K K =
N T
N K
K T N T N T
E F R

+ + + = ' '
1 1

Estimate first the residual covariance matrix and then
the average of residual variances:
Where is the sum of diagonal elements in
'
Prof. Doron Avramov
374
( )
( )
N
V tr
K
E E
K T
V
N N
'
1
1
2
=

=
( ) V tr

Compute for each chosen K
and pick K which minimizes this criterion.
'
Prof. Doron Avramov
375
( ) ( ) ( )
|
\
|
+
+
+ =
T N
NT
NT
T N
K K K K PC ln

max
2 2

Lecture Notes in
The Black-Litterman (BL) Approach
for Estimating Mean Returns:
Basics, Extensions, and
Incorporating Market Anomalies
'
Prof. Doron Avramov
376
Estimating Mean Returns
The Bayesian Perspective
Thus far, we have been dealing with classical, also termed
frequentist, econometrics.
Indeed, GMM, MLE, etc. are all classical methods.
The competing perspective is the Bayesian.
The BL approach is Bayesian.
Theoretically, the Bayesian approach is the most appealing
for decision making, such as asset allocation, security
selection, and policy making in general.
Major advantages: accounting for estimation risk, model
risk, informative priors, and it is numerically tractable.
'
Prof. Doron Avramov
377
To explain the difference between classical and Bayesian
methods, assume that you observe the market returns over T
periods:
The classical approach computes the sample mean return,
which is a stochastic (random) variable, and then specifies a
distribution for the mean return.
mT m m
R R R , , ,
2 1
K
'
Prof. Doron Avramov
378
For instance, assume that
Then the sample mean is distributed as
( )
2
, ~
m mt
N R
'
Prof. Doron Avramov
379
|
|
\
|
T
N r
m
2
, ~


The Bayesian approach is very different the unobserved
actual mean return is the stochastic variable.
The econometrician specifies both prior as well as likelihood
(data based) distributions for the mean return:
Prior: Likelihood:
( )
2
, ~
p p
N ( )
2
, ~
L L
N
'
Prof. Doron Avramov
380
The Posterior Distribution
The inference is then based on the posterior distribution
which combines the prior and the likelihood.
In particular, given the above specified prior and likelihood
based distributions, the posterior density is given by
( )
2
~
,
~
~ N
'
Prof. Doron Avramov
381
The Posterior Distribution
The mean and variance of the posterior are
( )
L p
w w
1 1
1
~
+ =
1
2 2
2
1 1
~

|
|
\
|
+ =
L p

'
Prof. Doron Avramov
382
The Posterior Mean Return
The posterior mean return is a weighted average of the prior
mean and the sample mean with weights proportional to the
precision of the prior versus the sample means, where
precision defined as the reciprocal of the variance.
In particular,
2 2
2
1
1 1
1
L p
p
w

+
=
2 2
2
1 2
1 1
1
1
L p
L
w w

+
= =
;
'
Prof. Doron Avramov
383
It is notoriously difficult to propose good estimates for
mean returns.
The sample means are quite noisy thus asset pricing
models -even if misspecified -could give a good guidance.
To illustrate, you consider a K-factor model (factors are
portfolio spreads) and you run the time series regression
1 1
2
1
2 1
1
1
1
1
+ + + + + =
N
t Kt
N
K t
N
t
N
N
N
e
t
e f f f r K
'
Prof. Doron Avramov
384
Then the estimated excess mean return is given by
where are the sample estimates of factor
loadings, and are the sample estimates of the
factor mean returns.
K
f K f f
e

2 1
2 1
+ + + = K
K

,
2 1
K
K
f f f

,
2 1
K
'
Prof. Doron Avramov
385
The BL Mean Returns
The BL approach combines a model (CAPM) with some
views, either relative or absolute, about expected returns.
The BL vector of mean returns is given by
(
+
|
\
|

(
+
|
\
|
=
1
1
1
1
1 1
1
1
1
1 1
1
' '
K
v
K K K N
N
eq
N N
N K K K K N
N N
N
BL
P P P
'
Prof. Doron Avramov
386
Understanding the BL Formulation
We need to understand the essence of the following
parameters, which characterize the mean return vector
Starting from the matrix you can choose any of the
specifications derived in the previous meetings either the
sample covariance matrix, or the equal correlation, or an
asset pricing based covariance, or you could rely on the LW
shrinkage approach either the complex or the nave one.
v eq
P , , , , ,
'
Prof. Doron Avramov
387
Constructing Equilibrium
Expected Returns
The , which is the equilibrium vector of expected
return, is constructed as follows.
Generate , the N 1 vector denoting the weights of
any of the N securities in the market portfolio based on
market capitalization. Of course, the sum of weights must
be unity.
Then, the price of risk is where and
are the expected return and variance of the market portfolio.
Later, we will justify this choice for the price of risk.
eq
MKT
2
m
m
Rf

=
2
m
'
Prof. Doron Avramov
388
Expected Returns
One could pick a range of values for going from 1.5 to
2.5 and examine performance of each choice.
If you work with monthly observations, then switching to the
annual frequency does not change as both the numerator
and denominator are multiplied by 12.
Having at hand both and , the equilibrium return
vector is given by
MKT

MKT
eq
=
'
Prof. Doron Avramov
389
Expected Returns
This vector is called neutral mean or equilibrium expected
return.
To understand why, notice that if you have a utility function
that generates the tangency portfolio of the form
then using as the vector of excess returns on the N assets
would deliver as the tangency portfolio.
e
e
TP
w

1
1
'

=
eq
MKT
'
Prof. Doron Avramov
390
What if you Directly Apply
the CAPM?
The question being would you get the same vector of
equilibrium mean return if you directly use the CAPM?
Yes, if
'
Prof. Doron Avramov
391
The CAPM based Expected
Returns
Under the CAPM the vector of excess returns is given by
( ) ( ) ( )
MKT
e
m
m
MKT
N
e
m
MKT
m
MKT
e e
m
e
m
e
e
m
N N
e
w
w
CAPM
w w r r r r
=
= = =
=

2
1
2 2 2
1 1
:
' , cov , cov
'
Prof. Doron Avramov
392
What if you use Directly the
CAPM?
Since and
then
( )
MKT
e e
m
w
'
= ( )
MKT
e e
m
w r r
'
=
eq
MKT
m
e
m
e
w
= =
2
'
Prof. Doron Avramov
393
What if you use Directly the
CAPM?
So indeed, if you use (i) the sample covariance matrix, rather
than any other specification, as well as (ii)
then the BL equilibrium expected returns and expected
returns based on the CAPM are identical.
2
m
m
Rf

=
'
Prof. Doron Avramov
394
The P Matrix: Absolute Views
In the BL approach the investor/econometrician forms some
views about expected returns as described below.
P is defined as that matrix which identifies the assets
involved in the views.
To illustrate, consider two "absolute" views only.
The first view says that stock 3 has an expected return of 5%
while the second says that stock 5 will deliver 12%.
'
Prof. Doron Avramov
395
The P Matrix: Absolute Views
In general the number of views is K.
In our case K=2.
Then P is a 2 N matrix.
The first row is all zero except for the fifth entry which is
one.
Likewise, the second row is all zero except for the fifth entry
which is one.
'
Prof. Doron Avramov
396
The P Matrix: Relative Views
Let us consider now two "relative views".
Here we could incorporate market anomalies into the BL
paradigm.
Market anomalies are cross sectional patterns in stock
returns unexplained by the CAPM.
Example: price momentum, earnings momentum, value, size,
accruals, credit risk, dispersion, and volatility.
'
Prof. Doron Avramov
397
Black-Litterman: Momentum and
Value Effects
Let us focus on price momentum and the value effects.
Assume that both momentum and value investing outperform.
The first row of P corresponds to momentum investing.
The second row corresponds to value investing.
Both the first and second rows contain N elements.
'
Prof. Doron Avramov
398
Winner, Loser, Value, and
Growth Stocks
Winner stocks are the top 10% performers during the past six
months.
Loser stocks are the bottom 10% performers during the past six
months.
Value stocks are 10% of the stocks having the highest book-to-
market ratio.
Growth stocks are 10% of the stocks having the lowest book-
to-market ratios.
'
Prof. Doron Avramov
399
Momentum and Value Payoffs
The momentum payoff is a return spread return on an
equal weighted portfolio of winner stocks minus return on
equal weighted portfolio of loser stocks.
The value payoff is also a return spread the return
differential between equal weighted portfolios of value and
growth stocks.
'
Prof. Doron Avramov
400
Back to the P Matrix
Suppose that the investment universe consists of 100 stocks
The first row gets the value 0.1 if the corresponding stock is a
winner (there are 10 winners in a universe of 100 stocks).
It gets the value -0.1 if the corresponding stock is a loser (there
are 10 losers).
Otherwise it gets the value zero.
The same idea applies to value investing.
Of course, since we have relative views here (e.g., return on
winners minus return on losers) then the sums of the first row
and the sum of the second row are both zero.
'
Prof. Doron Avramov
401
Back to the P Matrix
More generally, if N stocks establish the investment universe
and moreover momentum and value are based on deciles (the
return difference between the top and bottom deciles) then
the winner stock is getting 10/N
while the loser stock gets -10/N.
The same applies to value versus growth stocks.
Rule: the sum of the row corresponding to absolute views is
one, while the sum of the row corresponding to relative
views is zero.
'
Prof. Doron Avramov
402
Computing the Vector
It is the K 1 vector of K views on expected returns.
Using the absolute views above
Using the relative views above, the first element is the
payoff to momentum trading strategy (sample mean); the
second element is the payoff to value investing (sample
mean).
v
| |' 0.05,0.12 =
v
'
Prof. Doron Avramov
403
The Matrix
is a K K covariance matrix expressing uncertainty
about views.
It is typically assumed to be diagonal.
In the absolute views case described above denotes
uncertainty about the first view while denotes
uncertainty about the second view both are at the
discretion of the econometrician/investor.
( ) 1 , 1
( ) 2 , 2
'
Prof. Doron Avramov
404
The Matrix
In the relative views described above: denotes
uncertainty about momentum. This could be the sample
variance of the momentum payoff.
denotes uncertainty about the value payoff. This is the
could be the sample variance of the value payoff.
( ) 1 , 1
( ) 2 , 2
'
Prof. Doron Avramov
405
Deciding Upon
There are many debates among professionals about the right
value of .
From a conceptual perspective it should be 1/T where T
denotes the sample size.
You can pick
You can also use other values and examine how they
perform in real-time investment decisions.
1 . 0 =
'
Prof. Doron Avramov
406
Maximizing Sharpe Ratio
The remaining task is to run the maximization program
such that each of the w elements is bounded below by 0 and
subject to some agreed upon upper bound, as well as the sum
of the w elements is equal to one.
w w
w
BL
w
'
'
max

'
Prof. Doron Avramov
407
Extending the BL Model to Incorporate Sample
Moments
Consider a sample of size T, e.g., T=60 monthly
observations.
Let us estimate the mean and covariance (V) of our N assets
based on the sample.
Then the vector of expected return that serves as an input for
asset allocation is given by
where
| | | |
sample sample BL sample
T V T V
1 1
1
1 1
) / ( ) / (

+ + =
| |
1
1 1
' ) (

+ = P P
'
Prof. Doron Avramov
408
Class Notes in
Risk Management:
Down Side Risk Measures
'
Prof. Doron Avramov
409
Downside Risk
Downside risk is the financial risk associated
with losses.
Downside risk measures quantify the risk of
losses, whereas volatility measures are both
about the upside and downside outcomes.
That is, volatility treats symmetrically up and
down moves (relative to the mean).
Or volatility is about the entire distribution while
down side risk concentrates on the left tail.
'
Prof. Doron Avramov
410
Downside Risk
Example of downside risk measures
Value at Risk (VaR)
Expected Shortfall
Semi-variance
Maximum drawdown
Downside Beta
Shortfall probability
We will discuss below all these measures.
'
Prof. Doron Avramov
411
Value at Risk (VaR)
The says that there is a 5% chance that the
realized return, denoted by R, will be less than .
More generally,
% 95
VaR
% 95
VaR
% 95
VaR
( )
=
1
Pr VaR R
% 5
'
Prof. Doron Avramov
412
Value at Risk (VaR)
( ) ( ) ( )
1
1
1
1

+ = =

VaR
VaR
where
, the critical value, is the inverse cumulative
distribution function of the standard normal evaluated at .
Let =5% and assume that
The critical value is
( )
1
( )
2
, ~ N R
( ) 64 . 1 05 . 0
1
=
'
Prof. Doron Avramov
413
Therefore
Value at Risk (VaR)
Check:
If
Then
( )
2
, ~ N R
( )
( ) ( ) 05 . 0 64 . 1 64 . 1 Pr
64 . 1
Pr
Pr Pr
1
= = < =
|
\
|

<
=
|
\
|

=

z
R
VaR R
VaR R
( ) ( )
64 . 1
1
1
+ = + =

VaR
'
Prof. Doron Avramov
414
Example: The US Equity Premium
Suppose:
That is to say that we are 95% sure that the future equity
premium wont decline more than 25%.
If we would like to be 97.5% sure the price is that the
threshold loss is higher.
To illustrate,
( )
( ) 25 . 0 20 . 0 64 . 1 08 . 0
20 . 0 , 08 . 0 ~
95 . 0
2
= = VaR
N R
( ) 31 . 0 20 . 0 96 . 1 08 . 0
975 . 0
= = VaR
'
Prof. Doron Avramov
415
VaR of a Portfolio
Evidently, the VaR of a portfolio is not necessarily lower
than the combination of individual VaR-s which is
apparently at odds with the notion of diversification.
However, recall that VaR is a downside risk measure while
volatility which diminishes at the portfolio level is a
symmetric measure.
'
Prof. Doron Avramov
416
Backtesting the VaR
The VaR requires the specification of the exact distribution
and its parameters (e.g., mean and variance).
Typically the normal distribution is chosen.
Mean could be the sample average.
Volatility estimates could follow ARCH, GARCH,
EGARCH, stochastic volatility, and realized volatility, all
of which are described later in this course.
We can examine the validity of VaR using backtesting.
'
Prof. Doron Avramov
417
Backtesting the VaR
Assume that stock returns are normally distributed with
mean and variance that vary over time
The sample is of length T.
Receipt for backtesting is as follows.
Use the first, say, sixty monthly observations to estimate
the mean and volatility and compute the VaR.
If the return in month 61 is below the VaR set an indicator
function I to be equal to one; otherwise, it is zero.
T t , , 2 , 1 K =
( )
2
, ~
t t t
N r
'
Prof. Doron Avramov
418
Backtesting the VaR
Repeat this process using either a rolling or recursive
schemes and compute the fraction of time when the next
period return is below the VaR.
If =5% - only 5% of the returns should be below the
computed VaR.
Suppose we get 5.5% of the time is it a bad model or just
a bad luck?
'
Prof. Doron Avramov
419
Model Verification Based on
Failure Rates
To answer that question let us discuss another example
which requires a similar test statistic.
Suppose that Y analysts are making predictions about the
market direction for the upcoming year. The analysts
forecast whether market is going to be up or down.
After the year passes you count the number of wrong
analysts. An analyst is wrong if he/she predicts up move
when the market is down, or predict down move when the
market is up.
'
Prof. Doron Avramov
420
Model Verification Based on
Failure Rates
Suppose that the number of wrong analysts is x.
Thus, the fraction of wrong analysts is P=x/Y this is the
failure rate.
'
Prof. Doron Avramov
421
The Test Statistic
The hypothesis to be tested is
Under the null hypothesis it follows that
Otherwise H
P P H
:
:
1
0 0
=
( ) ( )
x y
x
P P
x
y
x f

|
|
\
|
=
0 0
1
'
Prof. Doron Avramov
422
The Test Statistic
Notice that
( )
( ) ( )
( )
( ) 1 , 0 ~
1
1
0 0
0
0 0
0
N
y P P
y P x
Z
Thus
y P P x VaR
y P x E
=
=
=
'
Prof. Doron Avramov
423
Back to Backtesting VaR: A
Real Life Example
In its 1998 annual report, JP Morgan explains: In 1998,
daily revenue fell short of the downside (95%VaR) band
on 20 trading days (out of 252) or more than 5% of the
time (2525%=12.6).
Is the difference just a bad luck or something more
systematic? We can test the hypothesis that it is a bad luck.
Otherwise H
x H
:
6 . 12 :
1
0
=
14 . 2
252 95 . 0 05 . 0
6 . 12 20
=

= Z
'
Prof. Doron Avramov
424
Back to Backtesting VaR: A
Real Life Example
Notice that you reject the null since 2.14 is higher than the
critical value of 1.96.
That suggests that JPM should search for a better model.
They did find out that the problem was that the actual
revenue departed from the normal distribution.
'
Prof. Doron Avramov
425
Expected Shortfall (ES): Truncated
Distribution
ES is the expected value of the loss conditional upon the
event that the actual return is below the VaR.
The ES is formulated as
| |

=
1 1
| VaR R R E ES
'
Prof. Doron Avramov
426
Expected Shortfall (ES) and the
Truncated Normal Distribution
Assume that returns are normally distributed:
( )
2
, ~ N R
( ) | |
( ) ( )
1
1
1
1
|
+ =
+ =
ES
R R E ES
'
Prof. Doron Avramov
427
where is the pdf of the standard normal density
e.g
This formula for ES is about the expected value of a
truncated normally distributed random variable.
( )
( ) 0.103961 64 . 1 =
'
Prof. Doron Avramov
428
Proof:
since
( )
( )
( )
( )
( ) ( )
1
0
0
1
2
|
, ~
= VaR x x E
N x
( )

1
1
0

=

=
VaR
'
Prof. Doron Avramov
429
Expected Shortfall: Example
Example
( )
( )
40 . 0
25 . 0
06 . 0
20 . 0 08 . 0
025 . 0
96 . 1
20 . 0 08 . 0
32 . 0
05 . 0
10 . 0
20 . 0 08 . 0
05 . 0
64 . 1
20 . 0 08 . 0
% 20
% 8
% 5 . 97
% 95
= +
+ =
= +
+ =
=
=
ES
ES
'
Prof. Doron Avramov
430
Previously, we got that the VaRs corresponding to those
parameters are 25% and 31%.
Now the expected losses are higher, 32% and 40%.
Why?
The first lower figures (VaR) are unconditional in nature
relying on the entire distribution.
In contrast, the higher ES figures are conditional on the
existence of shortfall realized return is below the VaR.
'
Prof. Doron Avramov
431
Expected Shortfall in
Decision Making
The mean variance paradigm minimizes portfolio volatility
subject to an expected return target.
Suppose you attempt to minimize ES instead subject to
expected return target.
'
Prof. Doron Avramov
432
Expected Shortfall with
Normal Returns
If stock returns are normally distributed then the ES chosen
portfolio would be identical to that based on the mean
variance paradigm.
No need to go through optimization to prove that assertion.
Just look at the expression for ES under normality to
quickly realize that you need to minimize the volatility of
the portfolio subject to an expected return target.
'
Prof. Doron Avramov
433
Target Semi Variance
Variance treats equally downside risk and upside potential.
The semi-variance, just like the VaR, looks at the downside.
The target semi-variance is defined as:
where h is some target level.
For instance,
Unlike the variance,
f
R h =
( ) ( ) | |
2
0 , min h r E h =
( )
2
2
= r E
'
Prof. Doron Avramov
434
The target semi-variance:
I. Picks a target level as a reference point instead of the
mean.
II. Gives weight only to negative deviations from a
reference point.
'
Prof. Doron Avramov
435
Notice that if
where
and are the PDF and CDF of a variable,
respectively
Of course if
then
( )
2
, ~ N r
( )
|
\
|

(
(
+
|
\
|

+
|
\
|

=

h h h h
h 1
2
2 2
( ) 1 , 0 N
= h
( )
2
2
= h
'
Prof. Doron Avramov
436
Maximum Drawdown (MD)
The MD (M) over a given investment horizon is the largest
M-month loss of all possible M-month continuous periods
over the entire horizon.
Useful for an investor who does not know the entry/exit
point and is concerned about the worst outcome.
It helps determine the investment risk.
'
Prof. Doron Avramov
437
Down Size Beta
I will introduce three distinct measures of downsize beta
each of which is valid and captures the down side of
investment payoffs.
Displayed are the population betas.
Taking the formulations into the sample simply replace the
expected value by the sample mean.
'
Prof. Doron Avramov
438
Downside Beta
The numerator in the equation is referred to as the co-semi-
variance of returns and is the covariance of returns below
on the market portfolio with return in excess of on
security i.
It is argued that risk is often perceived as downside
deviations below a target level by market participants and
the risk-free rate is a replacement for average equity market
returns.
( ) ( ) | | | |
( ) | |
2
) 1 (
) 0 , min(
) 0 , min(
f m
f m f i
im
R R E
R R R R E

=
f
R
f
R
'
Prof. Doron Avramov
439
Downside Beta
where and are security i and market average return
respectively.
One can modify the down side beta as follows:
( ) ( ) | | | |
( ) | |
2
) 2 (
0 , min
0 , min
m m
m m i i
im
R E
R R E

=
i
( ) | | ( ) | | | |
( ) | |
2
) 3 (
0 , min
0 , min 0 , min
m m
m m i i
im
R E
R R E

=
'
Prof. Doron Avramov
440
Shortfall Probability
We now turn to understand the notion of shortfall
probability.
While VaR specifies upfront the probability of
undesired outcome and then finds the threshold level,
shortfall probability gives a threshold level and seeks
for the probability that the outcome is below that
threshold.
We will thoroughly study the implications of shortfall
probability for long horizon investment decisions.
'
Prof. Doron Avramov
441
Shortfall Probability in Long
Horizon Asset Management
Let us denote by R the cumulative return on the
investment over several years (say T years).
Rather than finding the distribution of R we analyze
the distribution of
which is the continuously compounded return over the
investment horizon.
( ) R r + = 1 ln
'
Prof. Doron Avramov
442
The investment value after T years is
Dividing both sides of the equation by we get
Thus
( )( ) ( )
T T
R R R V V + + + = 1 1 1
2 1 0
K
0
V
( )( ) ( )
T
T
R R R
V
V
+ + + = 1 1 1
2 1
0
K
( )( ) ( )
T
R R R R + + + = + 1 1 1 1
2 1
K
'
Prof. Doron Avramov
443
Taking natural log from both sides we get
Assuming that
Then using properties of the normal distribution, we
get
T
r r r r + + + = K
2 1
( )
2
, ~ T T N r
'
Prof. Doron Avramov
444
( ) T t N r
IID
t
,..., 1 , ~
2
=
Shortfall Probability and Long Horizon
The normality assumption for log return implies the log
normal distribution for the cumulative return more later.
Let us understand the concept of shortfall probability.
We ask: what is the probability that the investment yields a
return smaller than a threshold level (e.g., the riskfree rate)?
To answer this question we need to compute the value of a
riskfree investment over the T year period.
'
Prof. Doron Avramov
445
The value of such a riskfree investment is
where is the continuously compounded risk free rate.
( )
T
f r
R V V
f
+ = 1
0
( )
f
Tr V exp
0
=
f
r
'
Prof. Doron Avramov
446
Essentially we ask: what is the probability that
This is equivalent to asking what is the probability
that
This, in turn, is equivalent to asking what is the
probability that
rf T
V V <
0 0
V
V
V
V
rf
T
<
|
|
\
|
<
|
|
\
|
0 0
ln ln
V
V
V
V
rf
T
'
Prof. Doron Avramov
447
So we need to work out
Subtracting and dividing by both sides of
the inequality we get
We can denote this probability by
( )
f
Tr r p <
T
T
|
|
\
|
|
|
\
|

<

f
r
T z P
|
|
\
|
|
|
\
|

=

f
r
T Shortfall
probabilit
y
'
Prof. Doron Avramov
448
Shortfall Probability and long horizon
Typically which means the probability
diminishes the larger T.
Notice that the shortfall probability can be written as a
function of the Sharpe ratio of log returns:
<
f
r
( ) SR T SP =
'
Prof. Doron Avramov
449
Example
Take r=0.04, =0.08, and =0.2 per year. What is the
Shortfall Probability for investment horizons of 1, 2, 5, 10,
and 20 years?
Use the excel normdist function.
If T=1 SP=0.42
If T=2 SP=0.39
If T=5 SP=0.33
If T=10 SP=0.26
If T=20 SP=0.19
'
Prof. Doron Avramov
450
Cost of Insuring against Shortfall
Let us now understand the mathematics of insuring against
shortfall.
Without loss of generality let us assume that
The investment value at time T is a given by the random
variable
1
0
= V
T
V
'
Prof. Doron Avramov
451
Once we insure against shortfall the investment value
after T years becomes
If you get
If you get
T
V ( )
f T
Tr V exp >
( )
f T
Tr V exp < ( )
f
Tr exp
'
Prof. Doron Avramov
452
So you essentially buy an insurance policy that pays 0
if
Pays if
You ultimately need to price a contract with terminal
payoff given by
( )
f T
Tr V exp >
( )
f T
Tr V exp < ( )
T f
V Tr exp
( ) { }
T f
V Tr exp , 0 max
'
Prof. Doron Avramov
453
This is a European put option expiring in T years with
a. S=1
b. .
c. Riskfree rate given by
d. Volatility given by
e. Dividend yield given by
( )
f
Tr K exp =
f
r
0 =
'
Prof. Doron Avramov
454
The B&S formula indicates that
Which, given the parameter outlined above, becomes
( ) ( ) ( ) ( )
1 2
exp exp d N T S d N Tr K Put
f
=
|
\
|

|
\
|
= T N T N Put
2
1
2
1
'
Prof. Doron Avramov
455
To show the pricing formula of the put use the following:
and
while
'
Prof. Doron Avramov
456
d
ln S / K r T
T
1
2
1
2
=
+ + ( ) ( )
d d T
2 1
=
( ) ( )
( ) ( )
2 2
1 1
1
1
d N d N
d N d N
=
=
The B&S option-pricing model gives the current put
price P as
where
and is
( ) ( )
2 1
d N d N Put =
1 2
1
2
d d
T
d
=
=

( ) d N
{ } d z prob <
'
Prof. Doron Avramov
457
For (per year)
2 . 0 =
P T (years)
0.08 1
0.18 5
0.25 10
0.35 20
0.42 30
0.52 50
'
Prof. Doron Avramov
458
We have found out that the cost of the insurance increases
in T, even when the probability of shortfall decreases in T
(as long as the Sharpe ratio is positive).
To get some idea about this apparently surprising outcome
it would be essential to discuss the expected value of the
investment payoff given the shortfall event.
It is a great opportunity to understand down side risk when
the underlying distribution is log normal rather than
normal.
'
Prof. Doron Avramov
459
The Expected Value of
Cumulative Return
Two proper questions emerge at this stage:
1. What is the expected value of cumulative return during
the investment horizon ?
2. What is the conditional expectation conditional on
shortfall ?
We assume, without loss of generality, that the initial
invested wealth is one.
) (
T
V E
| | ) exp( |
f T T
Tr V V E <
'
Prof. Doron Avramov
460
The Expected Value of
Cumulative Return
Notice that, given the tools we have acquired thus far,
finding the conditional expectation is a nontrivial task
since is not normally distributed rather it is log-
normally distributed since.
Thus, let us first display some properties of the log normal
distribution.
) , ( ~ ) ln(
2
T T N V
T
T
V
'
Prof. Doron Avramov
461
The Log Normal Distribution
Suppose that x has the log normal distribution. Then the
parameters and are, respectively, the mean and the
standard deviation of the variables natural logarithm,
which means
where z is a standard normal variable.
The probability density function of a log-normal
distribution is
z
e x
+
=
'
Prof. Doron Avramov
462
The Log Normal Distribution
If x is a log-normally distributed variable, its expected
value and variance are given by
( )
( )
0 ,
2
1
, ,
2
2
2
ln
>
x e
x
x f
x
x

| |
| | ( ) ( ) | | ( )
2
2
2
1
1 1
2 2 2
2
x E e e e x Var
e x E
= =
=
+
+

'
Prof. Doron Avramov
463
The Mean and Variance of
Using moments of the log normal distribution, the mean
and variance of are
Next, we aim to find the conditional mean.
( ) 1 ) exp( ) 2 exp( ) (
)
2
1
exp( ) (
2 2
2
+ =
+ =

T T T V Var
T T V E
T
T
T
V
T
V
'
Prof. Doron Avramov
464
The Mean of a Variable that has the
Truncated Log Normal Distribution
where is a normalizing constant.
Let us make change of variables:
( )
( )
=
c
ln
2
1
2
2 x
1
x c) > x | E(x c F dx e
x

( ) c F 1/
dt e = dx and e = x t
- ln(x)
t t

+ +
=
'
Prof. Doron Avramov
465
then:
( )
( )
( )
( ) ( )
( )
( )
( )
( )
+ +
+ +
=
=
c ln
2
1
5 . 0
c ln
5 . 0
2
1
c ln
2
1
c ln
2
1
2
2
2 2
2
2
2
1
2
1
2
1
2 x
1
c) > x | E(x c F
dt e e
dt e
dt e
dt e e
t
t
t t
t
t
'
Prof. Doron Avramov
466
Let us make another change of variables:
The integral is the complement CDF of the standard normal
random variable.
( )
( )
( )

c ln
2
1
5 . 0
2
2
2
1
c) > x | E(x c F dv e e
v
dt = dv - t = v
'
Prof. Doron Avramov
467
Thus, the formula is reduced to:
( )
( )
( )
( )
( )
|
|
\
|
+ +
=
(
|
|
\
|

=
+
+

2
5 . 0
2
5 . 0
LN
ln
ln
1 c) > x | (x E c F
2
2
c
e
c
e
'
Prof. Doron Avramov
468
In the same way we can show that:
( )
( )
( )
( )
( )
( )
( )
( )
( )
( )
|
|
\
|

=
=
=

=
+

+ +
(

2
5 . 0
ln
2
1
5 . 0
ln
2
1
5 . 0
ln
2
1
ln
2
1
0
LN
ln
2
1
2
1
2
1
c) x | (x E c F
2
2
2
2
2
2
2
c
e
dv e e dt e e
dt e dx e
x
x
c
v
c
t
c
t t
x
c
'
Prof. Doron Avramov
469
Punch Lines
( )
( )
|
\
|
+
|
|
\
| + +
= >

c
c
ln
ln
(x) E c) x | (x E
2
LN LN
( )
( )
|
\
|

|
|
\
|

c
c
ln
ln
(x) E c) x | (x E
2
LN LN
'
Prof. Doron Avramov
470
The Expected Value given Shortfall
The expected value given shortfall is
or
Notice that the denominator is the shortfall probability.
| |
|
\
|

|
|
\
|
=
|
\
|
+

T
T Tr
T
T T Tr
e shortfall V E
f
f
T T
T
2
2
1
2
|
| |
( ) ( )
( ) SR T
SR T
e shortfall V E
T T
T

+
=
|
\
|
+

2
2
1
|
'
Prof. Doron Avramov
471
The Expected Value given Shortfall
Thus,
Which means that the shortfall probability times the expected
value given shortfall is equal to the unconditional expected
value times a factor smaller than one.
That factor diminishes with higher Sharpe ratio and/or with
higher volatility.
| | ( ) ( ) + = SR T V E shortfall V E shortfall ob
t T
] [ | ) ( Pr
'
Prof. Doron Avramov
472
The Horizon Effect
Numerical example
Lets take . For different values of
the conditional expectation over horizon T looks like:
% 10 %, 5 = =
f
r
f
r >
'
Prof. Doron Avramov
473
The Horizon Effect
Previously we have shown that even when the shortfall
probability diminishes with the investment horizon, the cost
of insuring against shortfall rises.
Notice that the insured amount is
The expected value of that insured amount given shortfall
sharply rises with the investment horizon, which explains the
increasing value of the put option.
( )
T f
V Tr exp
'
Prof. Doron Avramov
474
Value at Risk with Log Normal
Distribution
We have analyzed VaR when quantities of interest are
normally distributed.
It is challenging to extend the analysis to the case wherein the
log normal distribution is considered.
Analytics follow.
'
Prof. Doron Avramov
475
VaR with Log Normal Distribution
We are looking for threshold, VaR, such that
Then in order to find the threshold we need to calculate
quantile of lognormal distribution:
where is as defined earlier.
( ) ( )
0 0 0
Pr V VaR CDF V Var V V
T
= < =
( )
2 1
0
, ; T T CDF V VaR

=
( )
1
0

+
=
T T
e V
( )
1
'
Prof. Doron Avramov
476
Specifically,
( ) ( ) ( ) ( )
0 0 0 0
ln ln Pr Pr V VaR V V V Var V V
T T
< = < =
( ) ( )
|
\
|

<

T
T V VaR
T
T V V
T 0 0
ln ln
Pr
( )
|
\
|

=

T
T V VaR
0
ln
'
Prof. Doron Avramov
477
and then
That is to say that there is a % probability that the investment
value at time T will be below that VaR.
( )
( )
( )

1
0
1
0
ln
=
=
|
\
|

T T
e V VaR
T
T V VaR
'
Prof. Doron Avramov
478
VaR with t Distribution
Suppose now that stock returns have a t
distribution with degrees of freedom and
expected return and volatility given by and
The pdf of stock return is formulated as
'
Prof. Doron Avramov
479
( )
( )
2
1
2
) (
1
2 1 , 2
1
, , |
+
|
|
\
|

+
=
v
v
x
v B v
v x f

Let -standardized r.v distribution with .
Than,
Partial Expectation
'
Prof. Doron Avramov
480
( ) ( )
( )
=
|
|
\
|
+
= =
+

dx
v
x
v B v
dx v x f x z X X PE
v
z z
2
1
2
1
2 1 , 2
1
, |
( )
( )
( )
1
, 1
1
2 1 , 2
1
| 1
1
2 1 , 2
1
2
1
2
1
2
2
1
2
+
=
|
|
\
|
+
=
(
(
(
|
|
\
|
+
=
+
+
v
z v
v z f
v
z
v
v
v B v
v
x
v
v
v B v
v
z
v

=
x
Y ( ) v x F , 1 , 0 |
And thus
Where is x quantile of
Partial Expectation
'
Prof. Doron Avramov
481
( )
( )
1
,
2
1
1
1
+
=

v
q v
q
v q f
ES
T

( ) x VaR q
x
= 1
n
t T ~
The sum of independent t-distributed random variables is not t-
distributed. So we have no nice formula for expected shortfall
in the long run. However, it can be approximated by normal
with zero mean variance:
Long Run Return when Periodic
Return has the t- Distribution
'
Prof. Doron Avramov
482
|
|
\
|

2
, ~
.
v
v
N r
approx
for 2 v
Approximation makes sense for large vs when t coincides
with normal distribution.
However, simulation studies show that for sufficient
number of periods this approximation works well enough.
'
Prof. Doron Avramov
483
However simulations shows that for sufficient number of
periods this approximation works well enough.
Let . The next graphs show normal
curve fit to the sum of t r.v.s (over T periods); sample
estimates vs. predicted parameters are includes
'
Prof. Doron Avramov
484
05 . 0 ; 01 . 0 = =
t t

'
Prof. Doron Avramov
485
-3 -2 -1 0 1 2 3
0
100
200
300
400
500
600
700
800
900
10 = T
274 . 0
1 . 0
=
=
N
N
273 . 0
102 . 0
=
=
3 = v
-4 -3 -2 -1 0 1 2 3 4 5
0
50
100
150
200
250
300
350
400
'
Prof. Doron Avramov
486
100 = T
866 . 0
1
=
=
N
N
856 . 0
011 . 1
=
=
3 = v
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
0
50
100
150
200
250
300
350
'
Prof. Doron Avramov
487
10 = T
177 . 0
1 . 0
=
=
N
N
174 . 0
099 . 0
=
=
10 = v
-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5
0
50
100
150
200
250
300
350
'
Prof. Doron Avramov
488
100 = T
559 . 0
1
=
=
N
N
566 . 0
994 . 0
=
=
10 = v
Estimating VaR and ES using EVT can be done in two
ways:
The idea of sub-periods;
The idea based on exceedances over a threshold.
Both ideas are based on assumption that extreme events of
many distributions (normal, student-t etc.) are distributed
with one of the extreme value distributions family. There is
a lot of material on an issue and jut summarized the most
important points.
VaR Expected Shortfall using Extreme
Value Theorem (EVT) - Without
Distribution Assumption
'
Prof. Doron Avramov
489
Divide the sample into m non-overlapping subsamples (to
preserve iid assumption) of length n:
Then using minimum of each sub-period estimate GEV
density of the following form
Sub-Periods
'
Prof. Doron Avramov
490
{ }
mn mn n n n
r r r r r r K K K K
1 2 1 1
| | |
+ +
( )
( )
where
e
e
z GEV
z
z
e
z
,
0 ,
1
1

=

+

0 ,

=
x
z
Using estimations of and we can compute VaR
Here we used the assumption that returns are iid and then
single period return is related to return over sub-period n as
follows
Sub-Periods
'
Prof. Doron Avramov
491
( )
( ) | |
( ) | |

)
+
+
=

1 ln ln
0 , 1 1 ln
n
n
VaR
n n
n
n
n
n
n
0 ,
n
( ) ( ) ( ) | |
n
n
t
n n n
c r P c r P c r P

=
< = < = <
1
n n
,
n
The result of the EVT are also relevant for the task of
converting short-term VaR into long-term VaR. Assume
applies to a single-period return for large R.
Then we have for a T-period return
It follows that a multi-period VaR forecast of a fat tailed
return distribution under the iid assumption is given by
Long Term VaR
'
Prof. Doron Avramov
492
1
1
~

|
\
|

R T R r P
T t
t
( ) ( )

1
1
VaR T VaR
T
=

( )
1
~

R R r P
Exceedances over a Threshold
'
Prof. Doron Avramov
493
( ) ( ) h l y h l P y F
h
> = |
for
h l y
F
0
( )
( ) ( )
( ) y F
y F y h F
y F
h
+
=
1
) 1 (
For a given threshold h conditional excess distribution
function on losses is defined as
In terms of F this can be written as
For a large class of underlying distribution functions F the
conditional excess distribution function , for h large,
is well approximated by generalized Pareto distribution:
For
( ) y F
h
Exceedances over a Threshold
'
Prof. Doron Avramov
494
( )
\
|
+
=
y
e
y
y GP
1
0 , 1 1
1
0 ,
0
| | h l z
F
, 0
if
and
| | , 0
if
0 <
Using this model VaR and ES can be expressed
analytically:
First, using (1) we can isolate
Then is replaced by GP CDF and F(h) by the sample
estimate , where n is the total number of
observations and is the number of exceedances above
the threshold h:
( ) y F
h
( )
n
N n
h
h
N
( ) l F
Long Horizon VaR
'
Prof. Doron Avramov
495
( ) ( ) | | ( ) ( ) h F y F h F l F
h
+ = 1
( ) ( ) ( )

1
1 1

+ = h l
n
N
l F
h
Next
And
is mean excess function for GP distribution which
equals to
Long Horizon VaR
'
Prof. Doron Avramov
496
( ) ( )
(
(
|
|
\
|
+ = =

1
1

h
N
n
h F VaR
( ) ( ) ( ) ( ) ( ) VaR l VaR l E VaR ES > + = |
Long Horizon VaR
'
Prof. Doron Avramov
497
( )
( )
+
+
=
1 1
h VaR
ES
Given a s sample and some threshold h, parameters , and
can be estimated using MLE.

Class Notes in
Testing the Black&Scholes
Formula
'
Prof. Doron Avramov
498
d
ln S / K r T
T
1
2
1
2
=
+ + ( ) ( )
d d T
2 1
=
C S,K, ,r,T, Se N d Ke N d
- T -rT
( ) = ( ) ( )

1 2
P S,K, ,r,T, Ke N d Se N d
-rT - T
( ) = ( ) ( )

2 1
Option Pricing
The B&S call Option price is given by
The put Option price is
Where and
'
Prof. Doron Avramov
499
Option Pricing
There are six inputs required:
S- Current price of the underlying asset.
K- Exercise/Strike price.
r- Continuously compounded riskfree rate.
T- Time to expiration.
- Volatility.
- Continuously compounded dividend yield.
'
Prof. Doron Avramov
500
The B&S Economy
The B&S formula is derived under several assumptions:
The stock price follows a geometric Brownian motion
(continuous path and continuous time).
The dividend is paid continuously and uniformly over
time.
The interest rate is constant over time.
'
Prof. Doron Avramov
501
The B&S Economy
The underlying asset volatility is constant over time
and it does not change with the option maturity or with
the strike price.
You can short sell or long any amount of the stock.
You can borrow and lend in the riskfree rate.
There are no transactions costs.
'
Prof. Doron Avramov
502
Testing the B&S Formula
Mark Rubinstein analyzes call options that are deep out of
the money.
He considers matched pairs: options with the same striking
price, on the same underlying asset (stock), but with
different time to maturity (expiration date).
He examines overall 373 pairs.
If B&S is correct then the implied volatility (IV) of the
matched pair is equal. Time to maturity plays no role.
'
Prof. Doron Avramov
503
Testing the B&S Formula
However, Rubinstein finds that our of the 373 examined
matched pairs shorter maturity options had higher IV.
Under the null the expected value of such an outcome is
373/2=186.5.
Is the difference statistically significant?
'
Prof. Doron Avramov
504
The Failure Rate based Test
Statistic
Use the failure rate test developed earlier to show that time to
expiration does play a major role.
That is to say that the constant volatility assumption is
strongly violated in the data.
'
Prof. Doron Avramov
505
The Volatility Smile for Foreign
Currency Options
'
Prof. Doron Avramov
506
Implied
Volatility
Strike
Price
Implied Distribution for
Foreign Currency Options
Both tails are heavier than the lognormal distribution.
It is also more peaked than the lognormal distribution.
'
Prof. Doron Avramov
507
The Volatility Smile for Equity
Options
'
Prof. Doron Avramov
508
Implied
Volatility
Strike
Price
Implied Distribution for Equity
Options
The left tail is heavier and the right tail is less heavy than the
lognormal distribution.
'
Prof. Doron Avramov
509
Ways of Characterizing the
Volatility Smiles
Plot implied volatility against K/S0 (The volatility smile is
then more stable).
Plot implied volatility against K/F0 (Traders usually define an
option as at-the-money when K equals the forward price, F0,
not when it equals the spot price S0).
Plot implied volatility against delta of the option (This
approach allows the volatility smile to be applied to some non-
standard options. At-the money is defined as a call with a delta
of 0.5 or a put with a delta of 0.5. These are referred to as 50-
delta options).
'
Prof. Doron Avramov
510
Possible Causes of Volatility
Smile
Asset price exhibits jumps rather than continuous changes.
Volatility for asset price is stochastic:
- In the case of an exchange rate volatility is not heavily
correlated with the exchange rate. The effect of a
stochastic volatility is to create a symmetrical smile.
- In the case of equities volatility is negatively related to s
stock prices because of the impact of leverage. This is
consistent with the skew that is observed in practice.
'
Prof. Doron Avramov
511
Volatility Term Structure
In addition to calculating a volatility smile, traders also
calculate a volatility term structure.
This shows the variation of implied volatility with the time to
maturity of the option.
The volatility term structure tends to be downward sloping
when volatility is high and upward sloping when it is low
'
Prof. Doron Avramov
512
'
Prof. Doron Avramov
513
Example of a Volatility Surface

K/S
0

0.90 0.95 1.00 1.05 1.10
1 mnth 14.2 13.0 12.0 13.1 14.5
3 mnth 14.0 13.0 12.0 13.1 14.2
6 mnth 14.1 13.3 12.5 13.4 14.3
1 year 14.7 14.0 13.5 14.0 14.8
2 year 15.0 14.4 14.0 14.5 15.1
5 year 14.8 14.6 14.4 14.7 15.0

Class Notes in
Time Varying Volatility Models
'
Prof. Doron Avramov
514
Volatility Models
We describe several volatility models commonly applied in
analyzing quantities of interest in finance and economics:
ARCH
GARCH
EGARCH
Stochastic Volatility
Realized and implied Volatility
'
Prof. Doron Avramov
515
Volatility Models
All such models attempt to capture the empirical evidence
that volatility is time varying (rather than constant) as well
as persistent.
The EGARCH captures the asymmetric response of
volatility to advancing versus diminishing markets.
In particular, volatility tends to be higher (lower) during
down (up) markets.
'
Prof. Doron Avramov
516
where
ARCH(1)
( ) 1 , 0 ~ N e
t
( ) ( )
2 2 2
1
2
1 t t t t t t
e E E = =

2
1
2
+ =
=
+ =
t t
t t t
t t
w
e
r

so
is the conditional variance.
2
t
'
Prof. Doron Avramov
517
ARCH (1)
( ) ( )
2
1
2
+ =
t t
w E E
( )
2 2
1
=
t
E
is the unconditional variance.
( )
2
1
+ =
t
E w
( ) ( )
2
1
2
1
+ =
t t
e E E w
( )
2
1
+ =
t
E w
( ) ( ) ( )
2
2
2
1
2 2

= = =
t t t
E E E
( )

=
1
2
w
E
t
'
Prof. Doron Avramov
518
ARCH (1)
( )
( )
( ) | |
( ) | |
2
2 2
1
4
1
2
2
4
4
t t t
t t
t
t
e E E
E E
E
E
+
= =
Fat tail?
( )( ) | |
( ) | |
( )
( ) ( )
( )
( ) ( )
3 3
3
3
2
2
4
2
2
4
2
2 2
1
4 4
1
=
=
+
=
t
t
t
t
t t t
t t t
E
E
E
E
e E E
e E E
'
Prof. Doron Avramov
519
ARCH (1)
( ) ( ) ( ) 0
2
2 4 2
=
t t t
E E Var
The last step follows because:
so
Yes fat tail!
( )
( )
1
2
2
4
t
t
E
E
'
Prof. Doron Avramov
520
where
GARCH(1,1)
( ) 1 , 0 ~ N e
t
( ) ( ) ( )
2
1
2
1
2
2
1
2
1
2

+ + =
+ + =
=
+ =
t t t
t t t
t t t
t t
E E w E
w
e
r

'
Prof. Doron Avramov
521
GARCH(1,1)
( )( )
3
3 2 1
1 1 3
1
1
2 2
4
2
2 2 2
>

+ +
=

=
+ + =

w
'
Prof. Doron Avramov
522
where
The first component
is the absolute value of a normally distribution variable
minus its expectation.
EGARCH
( ) 1 , 0 ~ N e
t
( ) ( )
2
1
1
1
1
1
2
ln
2
ln

+
|
|
\
|
+ =
=
+ =
t
t
t
t
t
t
t t t
t t
w
e
r

2 2
1
1
1
=
|
|
\
|
t
t
t
e
1 t
e
'
Prof. Doron Avramov
523
The second component is
Notice that the two normal shocks behave differently.
The first produces a symmetric rise in the log conditional
variance.
The second creates an asymmetric effect, in that, the log
conditional variance rises following a negative shock.
EGARCH
1
1
1
=
t
t
t
e
'
Prof. Doron Avramov
524
More formally if then the log conditional variance
rises by .
If then the log conditional variance rises by
This produces the asymmetric volatility effect volatility is
higher during down market and lower during up market.
EGARCH
0
1
<
t
e
+
0
1
>
t
e

'
Prof. Doron Avramov
525
Stochastic Volatility (SV)
There is a variety of SV models.
A popular one follows the dynamics
where
where
Notice, that unlike ARCH, GARCH, and EGARCH, here
the volatility itself has a stochastic component.
t t t t
r + =
( ) 1 , 0 ~ N
t
( ) ( )
t t t t
v + + =
1 1 0
ln ln
( )
( ) 0 , cov
1 , 0 ~
=
t t
t
v
N v
'
Prof. Doron Avramov
526
Realized Volatility
The realized volatility (RV) is a very tractable way to
measure volatility.
It essentially requires no parametric modeling approach.
Suppose you observe daily observations within a trading
month on the market portfolio.
RV is the average of the squared daily returns within that
month.
'
Prof. Doron Avramov
527
Realized Volatility
Of course, volatility varies on the monthly frequency but it
is assumed to be constant within the days of that particular
month.
If you observe intra-day returns (available for large US
firms) then daily RV is the sum of squared of five minute
returns.
'
Prof. Doron Avramov
528
d
ln S / K r T
T
1
2
1
2
=
+ + ( ) ( )
d d T
2 1
=
C S,K, ,r,T, Se N d Ke N d
- T -rT
( ) = ( ) ( )

1 2
P S,K, ,r,T, Ke N d Se N d
-rT - T
( ) = ( ) ( )

2 1
Implied Volatility (IV)
The B&S call Option price is given by
The put Option price is
Where and
'
Prof. Doron Avramov
529
Implied Volatility (IV)
In the traditional option pricing practice one inserts into the
formula all the six parameters, i.e., the stock price, the strike
price, the time to expiration, the cc riskfree rate, the cc
dividend yield, and stock return volatility.
IV is that volatility that if inserted into the B&S formula
would yield the market price of the call or put option.
As noted earlier, IV in not constant across maturities or
across strike prices.
'
Prof. Doron Avramov
530
Class Notes in
Stock Return Predictability,
Model Selection, and Model
Combination
'
Prof. Doron Avramov
531
Return Predictability
If log returns are IID there is no way you can deliver better
prediction for stock return than the current mean return.
That is, if
where is the continuously compounded return and is
the set of information available at time t.
t t
r + =
where
( )
2
, 0 ~ N
iid
t
| |
| |
2
1
1
|
|
=
=
+
+
t t
t t
I r Var
I r E
t
r
t
I
'
Prof. Doron Avramov
532
Return Predictability
Also note that the variance of a two-period return
is equal to
That is, variance grows linearly with the investment horizon,
while volatility grows in the rate square root.
( ) ( ) ( )
2
1 1
2 , cov 2 = + +
+ + t t t t
r r r Var r Var
1 +
+
t t
r r
'
Prof. Doron Avramov
533
Variance Ratio Tests
However, is it really the case?
Perhaps stock returns are auto-correlated, or
then:
( ) 0 , cov
1

+ t t
r r
( )
( )
( ) ( ) ( )
( )
t
t t t t
t
t t
r Var
r r r Var r Var
r Var
r r Var
VR
2
, cov 2
2
1 1 1
2
+ + +
+ +
=
+
=
( )
( ) ( )

+ = + =
+
+
1
, cov
1
1
1
t t
t t
r r
r r
'
Prof. Doron Avramov
534
Test:
The test statistic is
0 :
0 :
1
0
H
H
( ) ( ) 1 , 0 1
2
N VR T
d

'
Prof. Doron Avramov
535
More generally,
no auto correlation
Otherwise H
VR H
g
:
1 :
1
0
=
( )
( )
s
g
s
t
g t t t
g
g
s
r gVar
r r r Var
VR
=
+ +
|
|
\
|
+ =
+ + +
=
1
1
1 2 1
K
'
Prof. Doron Avramov
536
Variance Ratios
Test statistic:
e.g
( )
(
(
|
|
\
|

+
1
1
2
1 4 , 0 1
g
s
d
g
g
s
N VR T
2 = g
( ) | | 1 , 0 1
2
N VR T
d

3 = g
( )
(

9
20
, 0 1
3
N VR T
d
'
Prof. Doron Avramov
537
Predictive Variables
In the previous specification, we used lagged returns to
forecast future returns or future volatility.
You can use a bunch of other predictive variables, such as:
The term spread.
The default spread.
Inflation.
The aggregate dividend yield
'
Prof. Doron Avramov
538
Predictive Variables
The aggregate book-to-market ratio.
The market volatility.
The market illiquidity
'
Prof. Doron Avramov
539
Predictive Regressions
To examine whether stock returns are predictable, we can
run a predictive regression.
This is the regression of future excess log or gross return on
predictive variables.
It is formulated as:
1 2 2 1 1 1 + +
+ + + + + =
t Mt M t t t
z b z b z b a r K
'
Prof. Doron Avramov
540
Predictive Regressions
To examine whether either of the macro variables can predict
future returns test whether either of the slope coefficients is
different from zero.
Use the t-statistic or F-statistic for the regression R-squared.
There is a small sample bias if (i) the predictive variables are
highly persistent, (ii) the contemporaneous correlation
between the predictive regression residual and the innovation
of the predictor is high, or (iii) the sample is small.
'
Prof. Doron Avramov
541
Long Horizon Predictive
Regressions
The dependent variable is the sum of log excess return over
the investment horizon, which is K periods.
Since the residuals are auto correlated compute the standard
errors for the slope coefficient accounting for serial
correlation and often for heteroskedasticity.
For instance you can use the Newey-West correction.
K t t t K t t
z b a r
+ + + +
+ + =
, 1 , 1
'
'
Prof. Doron Avramov
542
Newey-West Correction
Rewriting the long horizon regression
| |
| | ' , '
, 1
' '
, 1
'
, 1
b a
z x
x r
t t
K t t t K t t
=
=
+ =
+ + + +

'
Prof. Doron Avramov
543
Newey-West Correction
The estimation error of the regression coefficient is
represented by
where is the Newey-West given by serially correlated
adjusted estimator
( ) ( ) S X X Var
'
1
=
S

+ =

=

=
T
j t
j t t
K
K j
T K
j K
S
1
1

'
Prof. Doron Avramov
( )
Var
544
Regressions
Tradeoff:
Higher K better coverage of dependence.
But we loose degrees of freedom.
Feasible solution:
3
1
T K
'
Prof. Doron Avramov
545
Regressions
E.g K=1:
Here we have no serial correlation.
1 , 0 , 1 = = = j j j
=
=
T
t
t
T
S
1
2
1

'
Prof. Doron Avramov
546
Regressions
E.g K=2:
2 , 1 , 0 , 1 , 2 = = = = = j j j j j

=

=
=
+
+ + =
T
t
t t
T
t
t
T
t
t t
T T T
S
2
1
1
2
1
1
1
1
2
1 1 1
2
1
+ =

=
=
+
T
t
T
t
t t t
T
1
1
1
1
2
1

'
Prof. Doron Avramov
547
In the Presence of
Heteroskedasticity
( )
| |

+
=

=
=
|
\
|
|
\
|
=
1
1
'
1
1
'
1
1
'
1
1 1
K T
t
j t j t t t
K
K j
T
t
t t
T
t
t t
x x
T K
j K
S
x x
T
S x x
T T
Var

'
Prof. Doron Avramov
548
Out of Sample Predictability
There is ample evidence of in-sample predictability, but little
evidence of out-of-sample predictability.
Consider the two specifications for the stock return evolution
Which one dominates? If then there is predictability
otherwise, there is no.
:
1
M
t t t
bz a r + + =
1
:
2
M
t t
r + =
1
M
'
Prof. Doron Avramov
549
Out of Sample Predictability
One way to test predictability is to compute the out of sample
:
Where is the return forecast assuming the presence of
predictability, and is the sample mean (no predictability).
Can compute the MSE (Mean Square Error) for both models.
2
R
( )
( )
=
=
=
T
t
t t
T
t
t t
OOS
r r
r r
R
1
2
1
2
1 ,
2
1
1 ,
t
r
t
r
'
Prof. Doron Avramov
550
Model Selection
When M variable are potential candidates for predicting stock
returns there are linear combinations of predictive models.
In the extreme, the model that drops all predictors is the no-
predictability or IID model.
The one that retains all predictors is the all inclusive model.
Which model to use?
One idea (bad) is to implement model selection criteria.
M
2
'
Prof. Doron Avramov
551
Model Selection Criteria
where L is the maximized value of the likelihood function.
Bayesian posterior probability
( ) L m AIC ln 2 2 =
( ) ( )
( ) ( )
1
1
1
1
1 1
ln 2 ln
2 2 2 2

=

=
=
m T
m
R R
m T
T
R R
L T m BIC
'
Prof. Doron Avramov
552
Model Selection
Notice that all criteria are a combination of goodness of fit
and a penalty factor.
You choose only one model and disregard all others.
Model selection criteria have been shown to exhibit very
poor out of sample predictive power.
'
Prof. Doron Avramov
553
Model Combination
The other approach is to combine models.
Bayesian model averaging (BMA) computes posterior
probabilities for each model then it uses the posterior
probabilities as weights to compute the weighted model.
There are more naive combinations.
Such combination methods produce quite robust predictors
not only in sample but also out of sample.
'
Prof. Doron Avramov
554

Avramov Doron Financial Econometrics

Transféré par

Informations du document

Description originale:

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Avramov Doron Financial Econometrics

Transféré par

Droits d'auteur :

Formats disponibles

Lecture Notes

Positive skewness means non-zero probability for large payoffs.

Sample estimates of the Skewness and Kurtosis are

sk is the skewness of the standardized

returns depart from

( ) ( )' ' , ' , ' vech =

HML SMB HML m HML

When is the Rejected CAPM Superior?

Economic versus Statistical

Understanding the Quantity .

Understanding the Quantity .

Understanding the Quantity .

Understanding the Quantity .

Understanding the Quantity .

Understanding the Quantity .

Understanding the Quantity .

' ' ' '

The Number of Factors in PCA

Estimating Mean Returns

Vous aimerez peut-être aussi