Financial Econometrics

Financial Econometrics
Lecture Notes
Prof. Doron Avramov

The Jerusalem School of Business Administration
The Hebrew University of Jerusalem
Introduction:
Why do we need a course in financial
econometrics?
Prof. Doron Avramov, The Jerusalem School of Business Administration, The Hebrew University of Jerusalem, Financial Econometrics
Syllabus: Motivation
The past few decades have been characterized by an
extraordinary growth in the use of quantitative methods and
analysis in financial markets for various asset classes; be it
equities, fixed income instruments, commodities, or derivative
securities.
Financial market participants, both academics and practitioners,
have been routinely using advanced econometric techniques in a
host of applications including portfolio management, risk
management, volatility modeling, and interest rate modeling, for
understanding pivotal issues in corporate finance and asset
pricing, as well as regulatory purposes, and more.
3
Syllabus: Objectives
This course attempts to provide a fairly deep understanding of
topical issues in asset pricing and deliver econometric methods.
More specifically, the purpose is twofold, both to develop
research agenda in financial economics and to comprehend
complex investment designs employed by practitioners.
The course targets advanced master level and PhD level

students in finance and economics.
Syllabus: Prerequisite #1
I will assume prior exposure to matrix algebra, distribution
theory, Ordinary Least Squares, Maximum Likelihood
Estimation, Method of Moments, and the Delta Method.
Syllabus: Prerequisite #2
I will also assume you have some skills in computer
programing beyond Excel.
MATLAB and R are the most recommended for this course. OCTAVE
could be used as well, as it is a free software, and is practically
identical to MATLAB when considering the scope of the course.
If you desire to use STATA, SAS, or other comparable tools, please
consult with the TA.
Syllabus: Grade Components

Assignments (36%): there will be two problem sets during the
term. You can form study groups to prepare the assignments up to four students per group. The main purpose of the
assignments is the implementation of the concepts studied in
class.
Class Participation (14%) - Attending AT LEAST 80% of the
sessions is mandatory. Otherwise, there you will not get credit
for this course.
Take-home final exam (50%): based on class material,
handouts, assignments, and readings.
7
Syllabus: Topics to be Covered - #1

Overview:
Matrix algebra
Regression analysis
Law of iterated expectations
Variance decomposition
Taylor approximation
Distribution theory
Hypothesis testing
OLS
MLE
8

Testing asset pricing models including the CAPM and multi
factor models: Time series analysis.
The econometrics of the mean-variance frontier
Estimating expected asset returns
Estimating the covariance matrix of asset returns.
Forming mean variance efficient portfolio, the Global

Minimum Volatility Portfolio, and the minimum Tracking
Error Volatility Portfolio.

9

The Sharpe ratio: estimation and distribution
The Delta method

The Black-Litterman approach for estimating expected
returns.
Principal component analysis.
10

Risk management and downside risk measures:
value at risk, shortfall probability, expected shortfall (also
known as C-VaR), target semi-variance, downside beta, and
drawdown.
Option pricing: testing the validity of the B&S formula

Model verification based on failure rates
11

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
12
Session #1 Overview
13
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
14
Financial Econometrics
In previous courses in finance and economics you had 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.
15
Plotting the Efficient Frontier
16
However, how could you

Practically Form an Efficient Portfolio?
Problem: there are far TOO many parameters to estimate.
For instance, investing in ten assets requires:
12 . . . . . . . . . . . . 1,10
1
2
1,2 , 22 . . . . . . . . . . . .
.
.....................
.
.....................
2
10
1,10 . . . . . . . . . . . 10
which is about ten estimates for expected return, ten for
volatility, and 45 for co-variances/correlations.
Overall, 65 estimates are required. That is a lot!!!
17
More generally,
if there are N investable assets, you need:
N estimates for means,
N estimates for 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.
18
Sample Mean and Volatility

The volatility estimate is less noisy than the mean.
I will later show that the standard error of the volatility
estimate is typically lower than that of the mean estimate.
We have T asset return observations:
1 , 2 , 3 , ,
First pass: the mean and volatility are estimated as

=
19
=1
Less Noise =
=1
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.
20
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?
21
Down Side Risk

For one, investors typically care more about the downside risk
of investments 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.
22
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.
23
Drawback in Mean-Variance
A
B
S1
S2
S3
S4
S5
5.00%
3.05%
8.25%
2.90%
15.00%
2.95%
10.50%
3.10%
20.05%
3.01%
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.
24
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.
25
Dispersion around the Mean

Assuming that is a zero mean random variable.
As ~(0, 2 ) the distribution takes the form:
When is small, the distribution is concentrated and vice

versa.
26
Probability Distribution Function

The Probability Density Function (pdf) of the normal
distribution is:
1
1 2
=

2
2
2
Note that =
1
2 2
, and specifically
if = 1, then =
1
2
The Cumulative Density Function (CDF), which is the integral

of the pdf, is 50% at that point: = 0.5
27
Probability Integral Transform

Assume that x is normally distributed what is the
distribution of y=F(x), where F is a cumulative density
Function?
Could one make a general statement here?
28
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.
29
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 approx. 95% confidence level (assuming ~(8%, (20%)2 )
30
Confidence Intervals for

Annual Excess Return on the Market
0.08 0.2 < < 0.08 + 0.2 = 68%
0.08 2 0.2 < < 0.08 + 2 0.2 = 95%
0.08 3 0.2 < < 0.08 + 3 0.2 = 99%
The probability that the realization is negative
< 0 =
31
0.08
0.2
<
00.08
0.2
= < 0.4 = 34.4%
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.
32
Skewness
The skewness can be negative (left tail) or positive (right tail).
33
Kurtosis
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.
34
The Essence of Skewness and Kurtosis

Positive skewness means nontrivial probability for large
payoffs.
Kurtosis is a measure for how thick the distributions tails are.
When is the skewness zero? In symmetric distributions.
35
Bivariate Normal Distribution

Bivariate normal:
2 ,
, , 2

The marginal densities of x and y are

~ , 2
~ , 2
What is the distribution of if is known?
36
Conditional Distribution
Conditional distribution:
2
2
~ + 2 , 2
When the correlation between and is positive and >

then the conditional expectation is higher than the
unconditional one.
37
Conditional Moments
If we have no information about , or if and are
uncorrelated, then the conditional and unconditional expected
return of are identical.
Otherwise, the expected return of y is x .
If = 0, then: =
meaning that the realization of does not say anything about .
38
Conditional Standard Deviation

Developing the conditional standard deviation further:
2 1
2 2
1 2
When goodness of fit is higher the conditional standard

deviation is lower.
That makes a great sense: the realization of x gives
substnantial information about y.
39
Multivariate Normal
1
1
2
= . ~ 1 ,
.
Define Z=AX+B.
40
Multivariate Normal
Let us make some transformations to end up with (0, ):
= 1 + 1 1 + 1 ,
1 + 1 0,
41
1
2
1 1 0,
Multivariate Normal
Consider an N-vector or returns which are normally
distributed:
1 ~ 1 ,
Then:
~ 0,
1
2
~ 0,
1
What does
mean? A few rules:
1
2
1
2
2
42
1
2
The Chi-Squared Distribution

If 1 ~(0,1) then 12 ~ 2 1
If 1~(0,1), 2~ (0,1) & 1 2 then:

12 + 22 ~ 2 2
43
More about the Chi-Squared

1 ~ 2
Moreover if 2 ~ 2
1 2
then 1 + 2 ~ 2 +
44
The F Distribution
Gibbons, Ross & Shanken (GRS) designated a finite sample asset
pricing test that has the F - Distribution.
If
1 ~ 2
2 ~ 2
1 2
then =
45
~ ,
The t Distribution
As noted earlier, 1 , 2 , , is a sample of length T of stock
returns which are normally distributed.
The sample mean and variance are
1 +
=
1
=
1
=1
The resulting t-value which has the t distribution with T-1 d.o.f
is
=
1
46
The t Distribution
The pdf of student-t is given by:
+1
2
2
1
, =
1+
2,1 2
where , is beta function and 1 is the number of

degrees of freedom
47
The Students t Distribution

is the number of
degrees of freedom.
When = + the tdistribution becomes
normal dist.
48
The t Distribution
The t-distribution is the sampling distribution of the t-value
when the sample consist of independently 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.
49
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 predictive regression
= + 1 1,1 + 2 2,1 +
50
= 1,2, ,
Rewriting the system in a matrix form

1
1
2
= .
.
1, 1,0 , 2,0
1, 1,1 , 2,1
.
=
31 = 1
.
2
1, 1,1 , 2,1
1
2
1 = .
.
yields
R= +
We will derive regression coefficients and their standard errors
using OLS, MLE, and Method of Moments.
51
Vectors and Matrices: some Rules

If A is a row vector: 1 = 11 , 12 , , 1
Then the transpose of A is a column vector: 1
11
12
= .
.
1
The identity matrix satisfies

= =
52
Multiplication of Matrices
22
22 22
11, 12
=
21, 22
22
11, 12
=
21, 22
11 11 + 12 21, 11 12 + 12 22
=
21 11 + 22 21, 21 12 + 22 22
22
,
11, 21
=
12, 22
() =
()1 = 1 1
The inverse of the matrix is 1 which satisfies:

1,0
1
22 22 = =
0,1
* Both A and B have to be invertible
53
Solving Large Scale Linear Equations

1 = 1
= 1
Of course, A has to be a square invertible matrix.
54
Linear Independence and Norm

The vectors 1 , , are linearly independent if there does not
exists scalars 1 , , such that
1 1 + 2 2 + + = 0
unless 1 = 2 = = 0.
In the context of financial economics we will consider N
assets (be it stocks or bonds) such that the payoff of each asset
is not merely a linear combination of the other N-1 assets.
Then the covariance matrix of asset returns is positive definite
(see next page) of rank N and hence is invertible.
The norm of a vector V is
=
55
A Positive Definite Matrix

An matrix
is called positive definite if

1 1
>0
for any nonzero vector V.

A non-positive definite matrix cannot be inverted and its
determinant is zero.
56
The Trace of a Matrix

12 1,
2
1,2 2 = 1 , ,
=
Let
1
1
,1, 2
Trace of a matrix is the sum of diagonal elements, therefore:
= 12 + 22 + + 2
57
Matrix Vectorization
58
1
2
This is the vectorization operator

it works for both square as well as
non square matrices.
The VECH Operator

Similar to
elements of .
but takes only the upper triangular
+1
1
2
59
1 1
22
=
2
Partitioned Matrices
A, a square nonsingular matrix, is partitioned as
11 , 12
1 1 1 2
=
21 , 22
2 1 2 2
Then the inverse of A is given by

1
1 , 1
1 1
11
12 22 21
11
12 22 21
12 22
1 =
1
1 , ( 1 )1
1
12 22 21
22
21 11 12
22 21 11
Check that 1 =
60
1 +2
Matrix Differentiation
Y is an M-vector. X is an N-vector. Then
1 1
1
,
,,
1 2
,
,,
1 2
And specifically, if
=
Then:
61
Let
1 1
62
Let
If C is symmetric, then
63
=
1 1
= +
= 2
Kronecker Product
It is given by
=
64
11 , 1
=
1 ,
Kronecker Product
For A,B square matrices, it follows that
= 1 1
65
Operations on Matrices: More Advanced

( ) = ()()
( + ) = () + ()
( + ) = () + ()
() = ( ) ()
( ) = ( )
= ( ) = ( ) = + ()
where
= (), = ( )( )
66
Using Matrix Notation

in a Portfolio Choice Context
The expectation and the variance of a portfolios rate of return in the
presence of three stocks are formulated as
= 1 1 + 2 2 + 3 3 =
2 = 12 12 + 22 22 + 32 32
+21 2 1 2 12 + 21 3 1 3 13 + 22 3 2 3 23
=
67
Using Matrix Notation

in a Portfolio Choice Context
where
31
68
1
= 2
3
31
1
= 2
3
12 , 12 , 13
= 21 , 22 , 23
33
31 , 32 , 32
The Expectation, Variance, and Covariance of

Sum of Random Variables
( + + ) = () + () + ()
+ + = 2 + 2 + 2
+2(, ) + 2(, ) + 2(, )
+ , + = , + ,
+(, ) + (, )
69
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.
70
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, +1 = |+1 = +1 [ ]
The conditional expectation of the price change is
[+1 | ] = +1 [ ] [ ] = 0
The quantity does not depend on the information.
71
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:
+1
1+
+2
2
1+
based on actual future dividends
The actual stock price: =

72
Variance Decomposition
=
>
What do we observe in the data? The opposite!!
73
Session 2(part a) - Taylor Approximations in

Financial Economics
74
TA in Finance
Several major applications in finance require the use of Taylor
series approximation.
The following table describes three applications.
75
TA in Finance: Major Applications
76
Expected Utility
Bond Pricing
Option Pricing
First Order
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:
1
= 0 + 0 0 +
1!
1
1
2
+ 0 0 + 0 0
2!
3!
It can also be written with
() =
77
notation:
( )
0
0
!
=0
Maximizing Expected Utility of Terminal Wealth

The invested wealth is , the investment horizon is periods.
The terminal wealth, the wealth in the end of the investment
horizon, is
+ = 1 +
where
1 + = 1 + +1 1 + +2 1 + +
Applying ln on both sides of the equation:
ln 1 + = ln 1 + +1 1 + +2 1 + + =
= ln 1 + +1 + ln 1 + +2 + + ln 1 + +
78
Transforming to Log Returns

Denote the log returns as:
+1 = ln ( 1 + +1 )
+2 = ln ( 1 + +2 )
.....
+ = ln ( 1 + + )
And:
= ln ( 1 + )
Hence the cumulative log return (CLR) over the investment
horizon is:
= +1 + +2 + + +
The terminal wealth as a function of CLR is therefore:
+ = (1 + ) = exp ( )
79
Power utility as a Function of Log Return

Assume that the investor has the power utility function (where
0 < < 1):
1
+ = +
Then the utility function of terminal wealth can be expressed

as a function of CLR
1
1
1
+ = + = exp ( ) = exp ( )
80
Power utility as a Function of Log Return

Note that is stochastic (unknown), since all future returns
are unknown
We can use the Taylor Approximation to express the utility as
a function of the moments of CLR.
We will approximate + around =
81
The Utility Function Approximation

(+ ) =
exp ( ) +
exp ( )( ) +
1
1
2
2
exp ( )( ) +
exp ( )( )3 3 +
2
6
1
exp ( )( )4 4 +
24
And the expected value of (+ ) is:
1
[(+ )]
exp ( )[1 + () 2 +
2
1
1
3
() + () 4 ]
6
24
82
Expected Utility
Let us assume that log return is normally distributed:
~ , 2
= 0, = 3 4
then:
2 2 4 4
exp ( ) 1 + +
2
8
The exact solution under normality is
2 2
+ =
exp +
2
Are solutions compatible?
83
Taylor Approximation Bond Pricing

Taylor approximation is also used in bond pricing.
The bond price is: 0 =
=1 1+
0
where 0 is the yield to maturity.

Assume that 0 changes to 1
The delta (the change) of the yield to maturity is written as:

= 1 0
84
Changes in Yields and Bond Pricing

Using Taylor approximation we get:
1
1
1 0 + 0 1 0 + 0 1 0
1!
2!
Therefore:
1 0
1
1
0 1 0 + 0 1 0
1!
2!
0 + 12 0
85
Duration and Convexity

Denote = 0 . Dividing by (0 ) yields
0 0
0
2 0
86
Instead of:
0
0
Instead of:
0
0
we can write -MD (Modified Duration).
we can write Con (Convexity).
The Approximated Bond Price Change

It is given by
2
+
2
What if the yield to maturity falls?
87
The Bond Price Change when Yields Fall

The change of the bond price is:
2
= +
2
According to duration - bond price should increase.
According to convexity - bond price should also increase.
Indeed, the bond price rises.
88
The Bond Price Change when Yields Increase

And what if the yield to maturity increases?
Again, the change of the bond price is:
2
= +
According to duration the bond price should decrease.
89
The Bond Price Change when 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 only second order.
90
Option Pricing
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?
1 2
2
( ) ( ) +
+
2 2
1
( ) + + 2
2
Focusing on the first order term this establishes the delta
neutral trading strategy.
91
Delta Neutral Strategy

Suppose that the underlying asset volatility increases.
Suppose further that 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.
92
Session 2(part b) - OLS, MLE, MOM
93
Ordinary Least Squares (OLS)

The goal is to estimate the regression parameters (least squares)
Using optimization can help us reach the goal.
Consider the regression = + :
1
= .
Different writing:
1, 11 , , 1
, = ,
1, 1 , ,
1
= .
= +
Or:
V=
Define a function
of the squares of the errors, that we want to minimize:
2 = = = + 2
=
=1
94
OLS First Order Conditions

Let us derive the function with respect to beta and consider the
first order condition:

= 2( ) 2 = 0
= ( )1
Recall: X and Y are observations.

Could we choose other to obtain smaller =
95
=1
The OLS Estimator

No! Because the optimization minimizes the quantity .
We know: = + = ( )1 ( + )
= ( )1 + ( )1
= + ( )1
Is the estimator unbiased for ?

= ( )1
= ( )1 = 0 , So it is indeed unbiased.
96
The Standard Errors of the OLS Estimates

What about the Standard Error of ?
( ) =
Reminder:
= ( )1
97
= ( )1
The Standard Errors of the OLS Estimates

Continuing:
( )1 ( )1 =

( )1 ( )1
= ( )1 ( )1 2 = ( )1 2
= ( )1 2
And we know that:

1
=
1
where is the number of explanatory variables in the regression specification.
2
Therefore, we can calculate
98
( )1
=
1
Maximum Likelihood Estimation (MLE)

We now turn to Maximum Likelihood as a powerful tool for both
estimating regression parameters as well as testing models.
Main Assumption of MLE in this context: the returns are normally

distributed and IID (Independent and Identically Distributed).
The goal is to estimate the distributions parameters that are
defined as = 2 .
Intuition: the data implies the most likely value of .
The MLE is an asymptotic procedure and it is a parametric

approach in that the distribution of the regression residuals must be
specified explicitly.
99
Implementing MLE
Assume that
~ (, 2 )
Let us estimate and 2 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
1
1 2
=
exp
2
2
2
2
100
The Joint Likelihood

We define the Likelihood function as the joint pdf
Following Bayes Rule:
= 1 , 2 , , = 1 2 2 3
Since returns are assumed IID - it follows that
= 1 2
2 2 2
1
exp
2
=1
Now take the natural log of the joint likelihood:
1
2
ln = ln 2 ln
2
2
2
101
=1
MLE: Sample Estimates

Derive the first order conditions

=
=1

=0
2
1
= 2+
2
2
2
=1
1
=
=1
=0
1
=
=1
Since 2 2 - the variance estimator is not unbiased.
102
MLE: The Information Matrix

Take second derivatives:
2
2
=
2
103
=1
=1
= 2
2
= 2
2
=0
2
= 4
2
=1
= 4
2
2
MLE: The Covariance Matrix

Set the information matrix =
The variance of is:
= ()1
Or put differently, the asymptotic distribution of is:
0, ()1
[The proof is beyond the scope of this course]
In our context, the covariance matrix is derived from the information matrix as follows:
2
2 , 0 "1" 2 , 0
,0
2 4
0, 4
0, 4
0,
2
2
104
To Summarize
Multiply by and get:
0, ()1
And more explicitly:
105
=1
=1
1
(
)2
0 2, 0
,
0 0,2 4
=1
The Sample Mean and Variance

Notice that when returns are normally distributed the
sample mean and variance are independent.
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 (2 4 ) by the variance of the mean ( 2 ) is smaller
than one as long as volatility is below 70%.
Clearly, the mean return estimate is more noisy because
typically, the volatility is far below 70%.
106
Method of Moments: departing from Normality

We know:
=
2 = 2
If:

=
2 2
Then:
0
=
0
That is, we set two momentum conditions.
107
Method of Moments (MOM)

There are two parameters: , 2
Stage 1: Moment Conditions =

Stage 2: estimation
1
108

2 2
= 0
=1
Method of Moments
Continue estimation:
1
= 0 = 1, ,
1
=
=1
2 = 0
=1
2
109
1
=
=1
MOM: Stage 3
=
2 ,

2 2
+ 4
2,
3
=
3 , 4 4
110
MOM: stage 4
Memo:
=

2 2
Stage 4: differentiate w.r.t. and take the expected value

=
111
1,
=
2 ,
0
1,
=
1
0,
0
1
MOM: The Covariance Matrix

Stage 5:
= 1
In this specific case we have: = , therefore:

=
112
The Covariance Matrices

Denote the MLE covariance matrix by 1 , and the MOM covariance
matrix by 2 :
2
2,
,
3
0
1
2
=
=
4 ,
0, 2
3 , 4 4
3 = = 3
(sk is the skewness of the standardized return)
4 = = 4
(kr is the kurtosis of the standardized return)
Sample estimates of the Skewness and Kurtosis are
1
3 =
113

=1
1
4 =
=1
Under Normality
the MOM Covariance Matrix Boils Down to
2
114
2,
=
3 = 0,
3 = 0
=
4
4
4 = 2
MOM: Estimating Regression Parameters

Let us run the time series regression
= + +
where:
= 1,
= ,
We know that | = 0
Given that | = 0, from the Law of Iterated Expectations (LIE)
it is possible to show that = = 0
Therefore there are two moment conditions (stage 1):

= 0
=

= 0
115
MOM: Estimating Regression Parameters

Stage 2: estimation:
1
=1
= 0
=1
=
=1
2
116
1,
=
1,
=1
1
=
MOM: Estimating Standard Errors

Estimation of the covariance matrix
Stage 3:
1
=
Stage 4:
1
=
117
()
()
=1
=1
()
1
=
1
=
=1
( )2
=1
( ) =
MOM: Estimating Standard Errors

Stage 5: the covariance matrix estimate is:
= ( 1 )1
( )2 ( )1
= ( )1
=1
Then, asymptotically we get

( ) (0, )
118
Session #3: Hypothesis Testing
119
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
120
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: = 0 means no performance
H1: Otherwise
121
Hypothesis Testing
Errors emerge if we reject H0 while it is true, or when we do
not reject H0 when it is wrong:
Dont reject H0
Reject H0
H0
Good decision
Type 1 error
H1
Type 2 error
Good decision
True state of world
122
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.

123
Skewness, Kurtosis & Bera Jarque Test Statistics

We aim to test normality of stock returns.
We use three distinct tests to examine normality.
124
Test I Skewness
TEST 1 - Skewness (third moment)
The setup for testing normality of stock return:
H0: , 2
H1: otherwise
Sample Skewness is
1
=
125
=1
6
0,
Test I Skewness
Multiplying by
,
6
we get
~
6
0,1
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 normality.
126
TEST 2 - Kurtosis
Kurtosis estimate is:
1
=
=1
24
3,
After transformation:
0
3 0,1
24
127
TEST 3 - Bera-Jarqua Test

The statistic is:
2
= +
3
6
24
2 2
Why 2 (2) ?
128
TEST 3 - Bera-Jarqua Test

If 1~(0,1) , 2 ~ (0,1) & 1 2 then:
12 + 22 2 2
and
24
3 are both standard normal and
they are independent random variables.
129
Chi Squared Test

In financial economics, the Chi square test is implemented
quite frequently in hypothesis testing.
Let us derive it.
Suppose that: =
0,
Then: =
130
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:

+
1 = 1 + 1
1
= +
131
Testing the CAPM

The expected excess return for asset i is given by:
)
( ) = + (
According to the CAPM, the intercept is restricted to be
zero for every test asset.
0 : 1 = 2 = = 0
So, the joint hypothesis test is:
1 :
While doing the estimation, we will never get = 0.

Rather, we examine whether the estimate is equal to zero
statistically. For example, the sample estimate could = 0.005
-- nevertheless this estimate could be insignificant, or it is
statistically equal to zero.
132
The Test Statistic

1
Estimate : = .
If: ,
0
and 0,
We get:
0
1
Later we will develop the exact analytics of this test.

133
Joint Hypothesis Test

You run the regression:
= 0 + 1 1 + 2 2 + , = 1,2, ,
134

=0
We have three orthogonal conditions: 1 = 0
2 = 0
Using matrix notation:
1
1
1, 11 , 21
.
.
1, 12 , 22
1 = . 3 =
1 = . 31 = 1 , 2 , 3
1, 1 , 2
1 = 3 31 + 1
135

Let us assume that: 0, 2 1 , 2 , 3 are iid
, 1 2
= 1
Joint hypothesis testing:
0 :
1 :
136
0 = 1, 2 = 0

1,0,0
1
Define: =
, =
0,0,1
0
The joint test becomes:
0 : =
0
0 H0 1
1,0,0
1 =
=
0,0,1
2
0
2
1 :
137

0 :
Returning to the testing:
1 :
= 0
,
H0
Under H0: ~ 0,
Chi squared:
~ 2 2
138

Yet, another example:
= 0 + 1 1 + 2 2 + 3 3 + 4 4 + 5 5 +
= 1,2,3, ,
1
= . ,6
1, 11 , , 51
61 = 0 , 1 , 6
1
.
=
,
.
1 = .
1, 1 , , 5
= +
139

Joint hypothesis test:
1
1
1
0 = 1, 1 = , 2 = , 3 = , 5 = 7
2
3
4
0 :
1 :
Here is a receipt:
1. =
2. =
3.
4. ,
140
1 2

1,0,0,0,0,0
0,1,0,0,0,0
5. = 0,0,1,0,0,0 , =
0,0,0,1,0,0
0,0,0,0,0,1
1
1
2
1
3
1
4
6. 0 : = 0
7. ,
0
8. 0,
9.
141
0 2 5

There are five degrees of freedom implied by the five
restrictions on 0 , 1 , 2 , 3 , 5
The chi-squared distribution is always positive.
142
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.
143
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.
144
Shrinkage Methods and the CAPM
50% CAPM : ( ) = (
)
+ = +
50% Data: 1 = 1 + 1
1
1
1
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:
1 The CAPM is true
=0
2 The CAPM is wrong 0
3 Mix (equally weights) between CAPM and the Data
145
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 model (coming up later) is also a type of
a shrinkage approach - it is more rigorous than the nave one
presented here.
The weights are on the model versus views on expected
returns, either absolute or relative.
146
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):
1 , 2 , ,
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 = , 2

The Sharpe ratio is equal to () =
147
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
0,
148

MOM
0, 2
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.

149
0,
0,
Distribution of the SR Estimate: The Delta Method
2
We will show that 0,
2
We use the Delta method to derive
The delta method is based upon the first order Taylor

approximation.
150
Distribution of the SR Estimate: The Delta Method

The first-order TA is
= 11
+

12
21
11
=

12
21
The derivative is estimated at
151
Distribution of the Sample SR

whereas
=
152
=0
The Variance of the SR

Continue:
153
First Derivatives of the SR

The SR is formulated as =
Let us derive:
2 0.5
1
=
1 2 0.5

2
=
=
2
2
2 3
154
The Distribution of SR under MLE

Continue:
1
=
,
2 3
2,
0,
0
2 4
2 3
1 2
0, 1 +
2
155
The Delta Method in General

Here is the general application of the delta method
If: ~ 0,
Then let be some function of :

0,
where is the vector of derivatives of with respect to
156
Hypothesis Testing
Does the S&P index outperform the Rf?
H0: SR=0
H1: Otherwise
Under the null there is no outperformance.
(0,1 + 122 )
Thus, under the null
1
2
1+ 2
157
(0,1)
Session #4: The Efficient Frontier and the

Tangency Portfolio
158
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.
In this course the focus is on time-series tests, while in the
companion course for PhD students Asset Pricing also cross
sectional tests are covered.
159
Testing Asset Pricing Models

Time series tests correspond to 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 nonportfolio based factors.
Consumption growth in the CCAPM is a good example of a
factor which is not a return spread.
160
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
161
Economic Interpretation of the Time Series Tests

Testing the validity of the CAPM entails the time series
regressions:
1
= 1 + 1
+ 1
= +
+
1 = 2 = =
The CAPM says: H0: 1 = 2 = =
H1: Otherwise
162
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 efficient frontier is
estimated.
The question is whether the market portfolio is close enough,
statistically, to the tangency portfolio.
163
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:
164
Testing Multifactor Models
1
= 1 + 11 1 + 12 2 + + 1 + 1
= + 1 1 + 2 2 + + +
The null is again:
H0: 1 = 2 = =
H1: Otherwise
In the multi-factor context, the hypothesis is that some
particular combination of the factors is the tangency portfolio.
165
The Efficient Frontier: Investable Assets

Consider N risky assets whose returns at time t are:
1
=
1
The expected value of return is denoted by:

1
1
=
= =
166
The Covariance Matrix

The variance covariance matrix is denoted by:
12 , , , 1
2
2 , , 2
= =
, 2
167
Creating a Portfolio
1
A portfolio is investing = is the N assets.
1
The return of the portfolio is: = 1 1 + 2 2 + +

The expected return of the portfolio is:
= 1 1 + 2 2 + + =
= 1 1 + 2 2 + + =
168
=1
The variance of the portfolio is:
2 = = 12 12 + 1 2 12 + 1 1
+1 2 12 + 22 22 + + 2 2
+
+
+1 1 + 2 2 + + 2 2
169
Thus 2 =
=1 =1
where is the coefficient of correlation.
Using matrix notation: 2 =

11
170
1 1
The Case of Two Risky Assets

To illustrate, let us consider two risky assets:
= 1 1 + 2 2
We know: 2 = = 12 12 + 21 2 12 + 22 22
Let us check:
2
= 1 , 2
12 , 12
12 , 22
1
2 2
2 2
=
+
2
1 2 12
1 1
2 2
2
So it works!
171
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
+ ( 1) 2
2
2
=
= cov
2
172
The Efficient Frontier: Excluding Riskfree Asset

The optimization program:
min

.
= 1
=
1
where is the Greek letter iota 1 = ,
1
and where is the expected return target set by the investor.
173
The Efficient Frontier: Excluding Riskfree Asset

Using the Lagrange setup:
1
= + 1 1 + 2
2
= 1 2 = 0
= 1 1 + 2 1
174
The Efficient Frontier: Without Riskfree Asset

Let
= 1
= 1
= 1
= 2
1
= 1 1
1
= 1 1
The optimal portfolio is: = +

1
175
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.
176

= +
1 1
1
1
2
= = = = 1
1 1
1
1
= = = 0
Indeed, = 1.
177

Let us now check whether the expected return on the portfolio
is equal to .
Recall:
= +
= +
So we need to show = 0
= 1
178

1 1
1
1
= = = 0
Try it yourself: prove that = 1.
179
Efficient Frontier
The optimization program also delivers the shape of the frontier.
180
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.
181
The Notion of Dominance

If
and there is at least one strong inequality, then portfolio B

dominates portfolio A.
182
The Efficient Frontier with Riskfree Asset

In practice, there is not really a riskfree asset. Why?
Credit risk (see Greece, Spain, Iceland).
Inflation risk.
Interest rate risk and the horizon effect.
183
Setting the Optimization

in the Presence of Riskfree Asset
The optimal solution is given by:
min
s. t + 1 = , + =
and the tangency portfolio takes the form:
1
1
= 1
= 1

The tangency portfolio is investing all the funds in risky assets.
184
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.
T
The location depends on the attitude towards risk.

185
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!
186
Equilibrium
The efficient frontier reflects
the supply side.
What about the demand?
The demand side can be
represented by a set of
indifference curves.
187
Equilibrium
What is the slope of indifference curve positive?
The equilibrium obtains when the indifference curve tangents
the efficient frontier
No riskfree asset:
188
Equilibrium
With riskfree asset:
189
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
1 2
=
2
where is the relative risk aversion.
190
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 the variance (supply).
191
Utility Maximization
= +
1

2
= = 0
1 1
Notice that here we get the same tangency portfolio

1
= 1

where reflects the fraction of weights invested in risky assets.
The rest is invested in the riskfree asset.
192
Mixing the Risky and Riskfree Assets
1 1
=
So if = 1 then all funds (100%) are invested in risky

assets.
If > 1 then some fraction is invested in riskfree asset.
If < 1 then the investor borrows money to leverage
his/her equity position.
193
The Exponential Utility Function

The exponential utility function is of the form
= exp where > 0
Notice that
= exp > 0
= 2 exp < 0
That is, the marginal utility is positive but it diminishes with

an increasing wealth.
194
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).
195
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
max
+1 |
. .
+1 = 1 + + +1
196
Let us assume that +1

,
Then +1 1 + + , 2
197

It is known that for , 2
exp
198
1 2 2
= exp +
2

Thus,
1 2
exp +1 = exp +1 + +1
2
1 2 2
= exp 1 + + +
2
= exp 1 +
199
1
exp
2


Notice that =
where is the relative risk aversion coefficient.
200

So the investor ultimately maximizes
1
max
The optimal solution is
1 1
.
The tangency portfolio is the same as before

1
= 1

201

Conclusion:
The joint assumption of exponential utility and normally
distributed stock return leads to the well-known mean
variance solution.
202
Quadratic Preferences
The quadratic utility function is of the form
=+
2
2
where > 0
= 1
2
= < 0
2
Notice that the first derivative is positive for <

203
The utility function looks like

204
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

max +1 |
. . +1 = 1 + + +1
205
Avramov and Chordia (2006 JFE) show that the optimization
could be formulated as
1
1
max
+
The solution takes the form

1
1
=

1
1+
206
The tangency portfolio is
= 1

The only difference from the previously presented competing

specifications is the composition of risky and riskfree assets.
207
The Sharpe Ratio of the Tangency Portfolio

1
Notice that
is actually the squared Sharpe Ratio of
the tangency portfolio.
Let us prove it
1
= 1

= + 1 = = 1

1
= = 1 2

208
The Sharpe Ratio of the Tangency Portfolio

Thus,
TP is a subscript for the tangency portfolio.
209
Session #5: Testing Asset Pricing Models:

Time Series Perspective
210
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.
211
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.
212

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.
213

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.
214
2 Moments for Asset Allocation

Pricing models deliver moments for asset allocation.
For instance, the tangency portfolio takes on the form
215
1
= 1

2 Asset Allocation
Under the CAPM, the vector of expected returns and the
covariance matrix are given by:
=
2
=
+
where is the covariance matrix of the residuals in the timeseries asset pricing regression.
We denoted by the residual covariance matrix in the case
wherein the off diagonal elements are zeroed out.
216
2 Asset Allocation
The corresponding quantities under the FF model are
+
=
+
=
where
217
is the covariance matrix of the factors.
3 Discount Factors
Expected return is the discount factor, commonly denoted by k,
in present value formulas in general and firm evaluation in
particular:
=
=1
1+
In practical applications, expected returns are typically

assumed to be constant over time, an unrealistic assumption.
218
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.
219
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):
= + ,
+
+ + +
220
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:
Examining the long run performance of IPO firm.
Examining the long run performance of SEO firms.

Analyzing abnormal performance of stocks going through dividend
initiation, dividend omission, splits and reverse splits.
221
5 Corporate Finance
Analyzing mergers and acquisitions
Analyzing the impact of change in board of directors.
Studying the impact of corporate governance on the cross section of
average returns.
Studying the long run impact of stock/bond repurchase.
222
Time Series Tests

Time series tests are designated to examine the validity of
asset pricng 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.
223
Time Series Tests

FMP is a convex combination of returns on some basis assets
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
particular combination of the factors is the tangency portfolio.
224
Testing the CAPM

Run the time series regression:
1
= 1 + 1
+ 1
= +
+
The null hypothesis is:
0 : 1 = 2 = = = 0
225
Testing the CAPM

In the following, I will introduce four times series test
statistics:
WALD.
Likelihood Ratio.
GRS (Gibbons, Ross, and Shanken (1989)).
GMM.
226
The Distribution of .
Recall, is asset mispricing.
The time series regressions can be rewritten using a vector

form as:
=
+

+
1
1 11
Let us assume that
~ 0,
for = 1,2,3, ,
Let = , ,
227
be the set of all parameters.
Under normality, the likelihood function for is
1
1
2
1
| =
exp

2
where is the constant of integration (recall the integral of a
probability distribution function is unity).
228
Moreover, the IID assumption suggests that
1 , 2 , , | =
1
exp
2
=1
Taking the natural log from both sides yields
ln ln
2
2 =1
229
Asymptotically, we have 0, ()
where
2
230
Let us estimate the parameters

=
=1
1
231
=1
1
1 +
2
=1
Solving for the first order conditions yields
232
=1
=1
Moreover,
1
=
233

=1
1
=
1
=
=1
=1
Recall our objective is to find the standard errors of .
Standard errors could be found using the information matrix.
234
The information matrix is constructed as follows
2

2
=

2

235
2
,
2
,
2
,

2
,

2
,

2
,

The Distribution of the Parameters

Try to establish yourself the information matrix.
Notice that and are independent of - thus, you can

ignore the second derivatives with respect to in the
information matrix if your objective is to find the distribution
of and/or .
If you aim to derive the distribution of then focus on the
bottom right block of the information matrix.
236
We get:
1
, 1 +
Moreover,
1 1
, 2

2,
237
Notice that (, ) stands for the Wishart distribution with
= 2 degrees of freedom and a parameter matrix = .
238
The Wald Test

Recall, if
,
then
1 2
Here we test
0 : = 0
0
where ~ 0,
1 : 0
The Wald statistic is 1 2
239
The Wald Test

which becomes:
1 = 1 +
2 1
1 =
2
1 +
where is the Sharpe ratio of the market factor.
240
Algorithm for Implementation

The algorithm for implementing the statistic is as follows:
Run separate regressions for the test assets on the common
factor:
1, 1
221
1
1
2 =
where
1,
= +
221
1
= ,
1
1 = 1 + 1
241

Retain the estimated regression intercepts
= 1 , 2 , ,
and = 1 , ,
Compute the residual covariance matrix

1
=
242

Compute the sample mean and the sample variance of the
factor.
Compute 1 .
243
The Likelihood Ratio Test

We run the unrestricted and restricted specifications:
un: = +
+
0,
res: =
+
0,
Using MLE, we get:
244

=
1
=
=1
2
=1

=1
1
1
,
2
2
+
1,
245

where again is the Wishart distribution, this time with 1
degrees of freedom.
246
The LR Test
= ln
ln = ln ln
2
2 = 2 = ln ln 2
Using some algebra, one can show that

2
1 = exp
1
Thus,
1
2 = ln
+1
247
GRS (1989)
Theorem: let
~ 0,
let A ~ , where
1
and let A and X be independent then

+ 1 1
,+1
248
GRS (1989)
In our context:
= 1+
2 2
0,
= ,
where
=2
Then:
1
3 =
1+
2 1
1 , 1
This is a finite-sample test.

249
GMM
I will directly give the statistic without derivation:
4 =
1 1 1
2 ()
where
, 0
1,
=
,
250
2
2
+
GMM
Assume no serial correlation but heteroskedasticity:
1
=
where

=1
= 1,
Under homoskedasticity and serially uncorrelated moment

conditions: 4 = 1 .
That is, the GMM statistic boils down to the WALD.

251
The Multi-Factor Version of Asset Pricing Tests

=
1 = 1 +
1 1
1

2 follows as described earlier.
3 =
1 + 1
1 1

where is the mean vector of the factor based return spreads.

252
The Multi-Factor Version of Asset Pricing Tests
is the variance covariance matrix of the factors.
For instance, considering the Fama-French model:

2
, ,
,
,
,
,
=
= , ,
2
, , ,,
253
The Current State of Asset Pricing Models

The CAPM has been rejected in asset pricing tests.
The Fama-French model is not a big success.

Conditional versions of the CAPM and CCAPM display some
improvement.
Should decision-makers abandon a rejected CAPM?
254
Should a Rejected CAPM be Abandoned?

Not so fast!
Assume that expected stock return is given by
= + +
where 0
You estimate using the sample mean and the mis-specified

CAPM:
255
=1
= +
Mean Squared Error (MSE)

The quality of estimates is evaluated based on
the Mean Squared Error (MSE)
256
1
2
2
2
MSE, Bias, and Noise of Estimates

Notice that = 2 +
Of course, the sample mean is unbiased thus

1
1
=
However, the CAPM is rejected, thus
2
2
2
= +
257
The Bias-Variance Tradeoff

It might be the case that 2 is significantly lower than
1 - thus even when the CAPM is rejected, still zeroing out
could produce a smaller mean square error.
258
When is the Rejected CAPM Superior?
259
1 2
1 2 2
= = + 2

Using variance decomposition

=
= |
=

=
+
1 2
|
+
2
2 + 2
where
260

Then

2
2 + 2
2
2
=
=
where 2 is the squared in the market regression.

Since is small -- the ratio of the variance estimates is
smaller than 1.
261
Example
Let
2 = 0.01
= 0.05
2 = 0.3
For what values of 0 it is sill preferred to use the CAPM?
Find such that the MSE of the CAPM is smaller.
262
Example
263
2
2
=
+
<1
1
1
2
0.05 0.7 + 0.3 +
<1
1
0.01
60
1
2
<
0.01 0.665
60
< 0.01528 = 1.528%
2
Economic versus Statistical Factors

Factors such as the market portfolio, SMB, HML, WML,
liquidity, credit risk, as well as bond based factors are prespecified.
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.
264
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.
265
The Economics of Time Series Test Statistics

Let us summarize the first three test statistics:
1
1 =
2
1 +
1
2 = ln
+1
1 1
3 =
2
1 +
266
The Economics of Time Series Test Statistics

The 4 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

1
267
The Economics of the Time Series Tests

GRS show that this quantity has a very insightful
representation.
Let us provide the steps.
268
Understanding the Quantity
Consider an investment universe that consists of + 1 assets the test assets as well as the market portfolio.
The expected return vector of the + 1 assets is given by
= ,
+1 1
11 1
where
is the estimated expected excess return on the
market portfolio and is the estimated expected excess return
on the test assets.
269
The variance covariance matrix of the + 1 assets is given by
+1 +1
2,
=
2,

where
2 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.
270
Notice that the covariance matrix of the N test assets is

2
=
+
The squared tangency portfolio of the + 1 assets is

2
= 1
271
Notice also that the inverse of the covariance matrix is
1 =
+ 1 ,
1 ,
1
1
Thus, the squared Sharpe ratio of the tangency portfolio could

be represented as
272

2
2
2
=
+ 1
or
2
2
1 =

273
In words, the 1 quantity is the difference between the

squared Sharpe ratio based on the + 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.
274
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 1 is given in the next slide.
275
2
1
1 = 12 22
276
So we can rewrite the previously derived test statistics as

2
2

2
1 =
~
2
1 +
2
2
1

3 =
~ , 1
2
1 +
277
Session #6: Asset Pricing Models with Time

Varying Beta
278
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.
279
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.
280
Conditional Models
Here is a conditional asset pricing specification:
= +
+
= 0 + 1 1
= + 1 +
(
|1 ) =
cov ( , ) = 0
281
Conditional Asset Pricing Models

Substituting beta back into the asset pricing equation yields.
= + 0
+ 1
1 +
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.
282
Conditional Asset Pricing Models

You can use the statistics 1 through 4 to test such models.
If we have factors and predictive variables then the
K-conditional factor model becomes a + - unconditional
factor model.
If you only scale the market beta, as is typically the case, we
have an + unconditional factor model.
283
Conditional Moments
Suppose you are at time what is the discount factor for time
+ 2?
+2
= + 0 +2
+ 1 +2
+1 + +2
= + 0 +2
+ 1 +2
+ + +1 + +2
= + 0 +2
+ 1 +2
+ 1 +2
+ 1 +2
+1 + +2
284
Conditional Moments
|
+2
= + 0
+ 1
+ 1
+
= + 0 + 1
1

Notice that here the discount factor, or the conditional expected

return, is no longer constant through time.
Rather, it varies with the macro variable.
285
Conditional Moments
Could you derive a general formula in particular you are at
time what is the expected return for time + as a function of
the model parameters as well as ?
286
Conditional Moments
Next, the conditional covariance matrix the covariance at
time + 1 given is given by
|
2 +
+1
= 0 + 1 0 + 1
where 0 and 1 are the N-asset versions of 0 and 1

respectively.
287
Conditional Moments
Could you derive a general formula in particular you are at
time what is the conditional covariance matrix for time +
as a function of the model parameters as well as ?
Could you derive general expressions for the conditional
moments of cumulative return?
288
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.
289
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.
290
Different Ways to Model Beta Variation

The base case: beta is constant, or time invariant.
Case II: beta varies with macro conditions

= 0 + 1 1
= + 1 +
291

Case III: beta varies with firm-level size and the
book-to-market ratio
= 0 + 1 ,1 + 2 ,1
Case IV: beta is some function of both macro and firm-level
variables as well as their interactions:
= 1 , ,1 , ,1
292

Case V: beta follows an auto-regressive AR(1) process
= + ,1 +
293
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.
294
Testing Conditional Models

You can implement the 1 4 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 Kalman filtering
methods involving state space representations.
295
Session #7: GMVP, Tracking Error Volatility,

Large scale covariance matrix
296
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.
297
GMVP Optimization

= 1
Solution: =
1
1
No analytical solution in the presence of portfolio constraints

such as no short selling.
298
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. That is
related to the volatility anomaly: low volatility stocks deliver
higher payoffs than high volatility stocks.
299
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.
300
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.
301
Tracking Error Volatility:

The Benchmark and Active Portfolios
Let be the vector of weights of the benchmark portfolio.
Then the expected return and variance of the benchmark
portfolio are given by
=
2 =
where, as usual, and are the vector of expected return and

the covariance matrix of stock returns.
302
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 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 , by construction, must
be equal to zero.
303
The Mathematics of TEV

So the fund manager invests = + in stocks, is the
passive part of the portfolio and x is the active part.
Notice that 2 = is the tracking error variance.
Also notice that the expected return and volatility of the chosen
portfolio are
= +
2 = = 2 + 2 + 2
304

The optimization problem is formulated as
max
. .
305
= 0
=

The resulting active part of the portfolio, , is given by
1
where
306
1 =
1 =
1 =
2 / =
Questions
Is the sum of the components equal to zero?
Prove that if the benchmark is the GMVP then all

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?
307
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 is totally
independent of the passive part .
Of course, the overall portfolio + is impacted by .
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.
308
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.
309
TEV with total volatility constraint

The optimization is formulated as
max
. .
= 0
=
( + ) ( + ) = 2
Home assignment: derive the optimal solution.
310
Estimating the Large Scale Covariance

Matrix
311
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.
312
The Sample Covariance Matrix (Denoted S)

This method uses sample estimates.
Need to estimate ( + 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 monthly returns on risky assets.

We can display those returns in a by matrix .
313
The Sample Covariance Matrix (Denoted S)

Estimate the mean return of the 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
=
314
The Equal Correlation Based

Covariance Matrix (Denoted F):
Estimate all ( 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 , 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
315
, =
, =
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

factors. Factor means are denoted by
Notice that the mean return is given by
= +
316
The Factor Based Covariance Matrix

Thus deviations from the means are given by
= +
The factor based covariance matrix is estimated by
= +
Here, is an by matrix of factor loadings and is a
diagonal matrix with each element represents the
idiosyncratic variance of each of the assets.
317
The Factor Based Covariance Matrix

Number of Parameters
This procedure requires the estimation of betas as well as
variances of the factors, ( 1)/2 correlations of those
factors, and firm specific variances.
Overall, you need to estimate + + ( 1)/2 +
parameters, which is considerably less than ( + 1)/2 since
is much smaller than .
For instance, using a single factor model the number of
parameters to be estimated is only 2 + 1.
318
Steps for Estimating

the Factor Based Covariance Matrix
1. Run the MULTIVARIATE regression of stock returns on asset
pricing factors
= + + =
+
11
+1
+1
where = ,
= ,
319

2. Estimate
=
= ,
and retain only.
320

3. Estimate the covariance matrix of
1
=

4. Let be a diagonal matrix with the (i,i)-th component being
equal to the (i,i)-th component of .
321

5. Compute = -
1 1
where is the mean return of the factors.

6. Estimate:
322

7.
This is the estimated covariance matrix of stock returns.

8. Notice there is no need to run N individual regressions! Use
multivariate specifications.
323
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.
324
A Shrinkage Approach
Based on a Paper by Ledoit and Wolf (LW)
Let be the sample covariance matrix, let be the equal
correlation based covariance matrix, and let be the shrinkage
intensity. and were derived earlier.
The operational shrinkage estimator of the covariance matrix
is given by = + (1 )
Notice that F is the shrinkage target.
325
The Shrinkage Intensity

LW propose the following shrinkage intensity, based on
optimization:
= max 0, min
,1
where is the sample size and is given as =
2
and where =
(
)
with being the (, )
=1 =1
component of and is the (, ) component of .
326
The Shrinkage Intensity
=1 =1
1
=
{( )(
=1
+
=1
1
=
1
=
=1 =1,
_
) }2
_ 2
)
)2
, +
,
_
) (
_
) (
=1
_
)
_
)
=1
and with and being the time return on asset and the time-series
average of return on asset , respectively.
327
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:
1
1
= +
2
2
328
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.
329
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:
330
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.
331
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
,+1 = +1
where +1 is the realized return at time + 1 on all the
investable assets.
332

Suppose you rebalance every six months derive the out of
sample returns also for the following 5 months
,+2 = +2
,+3 = +3
,+4 = +4
,+5 = +5
,+6 = +6
Then at time t+6 you re-derive the GMVPs.
333
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.
334

So you generate out of sample returns on each of the strategies
starting from time + 1 till the end of sample, which we
typically denote by .
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.
335

Rolling Scheme
S
MKT
FF
Recursive Scheme
FF+
MOM
LW
MKT
FF
FF+
MOM
Mean
STD
SR
SP
(5%)
alpha
IR
336
LW

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.
337

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.
338
High Freuqncy Data

Let denote the continuously compounded one-year return for
year , where are independent variables from , 2 (iid).
Based on annual observations, we can estimate and 2 :
1
1
2
2.
=
,
and
=1
=1
We have =
339
and
2 4
.
High Freuqncy Data continued

Now, partition each year to equally-separated intervals (e.g.,
5 minutes), so that we have returns for each year :
,1 , ,2 , , , so that =
=1 , .
Overall, we have observations, and denote: = ,
(the expected return of one interval).
Using the iid property, we can estimate: =
and =
And we have: = 2 , and also

2 = 2 2 .
340
2
.

2
The variance of and of
As usual, =
Therefore we have (also using the results from previous slide):

2
= =
2 2
2
.
The conclusion regarding : its variance is the same whether it

is estimated based on annual observations or whether it is
estimated using high-frequency observations.
There is no gain in using high-frequency observations.
We will see a different result for 2 .
341

2
The variance of and of
What is the variance of 2 based on high-frequency
observations?
As usual, 2 =

2 4
.
22 4
1 2 4
The variance of 2 is much smaller with the partitioning !

Moreover, the larger the partition is (i.e., is larger), the
smaller the variance of the variance is.
342
Session #8: Principal Component Analysis

(PCA)
343
PCA
The aim is to extract common factors to summarize the
information of a panel of rank .
In particular, we have a panel of stock returns where is
the time dimension and (< ) is the number of firms of
course <<
= 1 , ,
The PCA is an operation on the sample covariance matrix of

stock returns.
344
PCA
You have returns on stocks for periods
1 , ,
1 = 1 1 =
let
= 1 , 2 , ,
1
=
345
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

max
1 1
1
346
1 1 = 1
PCA
1 is an eigen vector since
1 = 1 1
Moreover
1 = 1 1
1 is therefore the highest eigen value.
347
PCA
Extracting the second eigen vector
max
2
348
2 2
2 2 = 1
1 2 = 0
PCA
The optimization yields:
2 = 2 2
2 = 2 2 < 1
The second eigen value is smaller than the first due to the
presence of one extra constraint in the optimization the
orthogonality constraint.
349
PCA
The K-th eigen vector is derived as
max

.
= 1
1 = 0
2 = 0
1 = 0
350
PCA
The optimization yields:
=
= < 1 < < 2 < 1
351
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
352
PCA
Recall, the basic idea here is to replace the original set of
variables with a lower dimensional set of K-factors ( << ).
The contribution of the -th eigen vector to explain the
covariance matrix of stock returns is
=1
353
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!
354
Understanding the PCA: Digging Deeper

The covariance matrix can be decomposed as
= 1 , ,
355
1, , 0

0, ,

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.
356

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.
357

The matrix is given by
= 1 , ,
1, , 0

0, ,
Of course, the dimension of is by .

However, its rank is , thus the matrix is not invertible.
358
Is close to ?
This is the same as asking: what does it mean to explain the
sample covariance matrix?
359
Is close to ?
Let us represent returns as
1 = 1 +
= +
360
11 1
12 2 +
1
+ 1
1 1
2 2 +
= 1, ,
= 1, ,
Is close to ?
where 1 are the principal component based factors and
1 are the exposures of firm i to those factors.
361
Is close to ?
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.
362
The Contribution of the PC-S

to Explain Portfolio Variation
Let
= 1 , , be the exposures of firm to the
common factors.
Let be an N-vector of portfolio weights:
= 1 , 2 , ,
Recall, is a matrix of stock returns.

363

The portfolios rate of return is
=
1
Moreover, the portfolio time t return is given by

= 1 1 + 2 2 + +
1 1
364

We can approximate the portfolios rate of return as:
= 1 11 1 + 12 2 + + 1
+2 21 1 + 22 2 + + 2
+
+ 1 1 + 2 2 + +
365

Thus, = 1 1 + 2 2 + +
where
1 = 1
2 =
1 1
2
1 1
1 1
366

Notice that
1 are the loadings on the common factors,
or they are the risk exposures, while 1
are the K-realizations of the factors at time t.
367
Explain the Portfolio Variance

The actual variation is
2 ( ) =
The approximated variation is

2 ( ) = 12 var ( 1 ) + 22 var ( 2 ) + + 2 var ( )
Both quantities are quite similar.
368
Explain the Portfolio Variance

The contribution of the i-th PC to the overall portfolios
variance is:
2 ( )
=1 ( )
369
Asy. PCA: What if N>T?

Then create a matrix and extract eigen vectors
those eigen vectors are the factors
1
=
where
=
1 1
and is the T-vector of cross sectional (across- stocks) mean of

returns.
370
Other Applications
PCA can be implemented in a host of other applications.
For instance, you want to predict economic growth with many

predictors, say where is large.
You have a panel of predictors, where is the timeseries dimension.
371
Other Applications
In a matrix form
11 , 12 , , 1
1 , 2 , ,
where is the m-the predictor realized at time .
372
Other Applications
If > compute the covariance matrix of then extract
principal components such that you summarize the
M-dimension of the predictors with a smaller subspace of order
<< .
You extract 1 , , eigen vectors.

1
373
Other Applications
Then you construct K predictors:
1 =
374
What if > ?
If > then you extract eigen vectors from the TT
matrix.
In this case, the -predictors are the extracted eigen vectors.
Be careful of a look-ahead bias in real time prediction.
375
The Number of Factors in PCA

An open question is: how many factors/eigen vectors should be
extracted?
Here is a good mechanism: set max - the highest number of
factors.
Run the following multivariate regression for
= 1,2, , max
= + +
376
1 1

Estimate first the residual covariance matrix and then the
average of residual variances:
1
=
where is the sum of diagonal elements in
377

Compute for each chosen
2
= + max
ln
and pick which minimizes this criterion.
378
Session #9: The Black-Litterman (BL)

Approach for Estimating Mean Returns:
Basics, Extensions, and Incorporating
Market Anomalies
379
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.
380

To explain the difference between classical and Bayesian
methods, assume that you observe the market returns over T
periods:
1 , 2 , ,
The classical approach computes the sample mean return,
which is a stochastic (random) variable, and then specifies a
distribution for the mean return.
381

For instance, assume that
2
,
Then the sample mean is distributed as

2
382

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: , 2
383
Likelihood: , 2

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
384

The Posterior Distribution
The mean and variance of the posterior are
= 1 + 1 1
385
1
1
= 2+ 2

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,
1 =
386
1
2
1
1
2+ 2

2 = 1 1 =
1
2
1
1
2+ 2


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 2 + + +
1
387

Then the estimated excess mean return is given by
= 1 1 + 2 2 + +
where 1 , 2 are the sample estimates of factor loadings,
and 1 , 2 are the sample estimates of the factor mean
returns.
388
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
=
389
1
11
+ 1
1
11
+ 1
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.
390
Constructing Equilibrium Expected Returns

The , which is the equilibrium vector of expected return, is
constructed as follows.
Generate , the 1 vector denoting the weights of any
of the securities in the market portfolio based on market
capitalization. Of course, the sum of weights must be unity.
Then, the price of risk is =
2
where and
are the
expected return and variance of the market portfolio.
Later, we will justify this choice for the price of risk.

391

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
=
392

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
1
= 1
then using as the vector of excess returns on the assets

would deliver as the tangency portfolio.
393
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
394
The CAPM based Expected Returns

Under the CAPM the vector of excess returns is given by
cov ,
cov ,
=
=
=
2
2
395

=
2
What if you use Directly the CAPM?

=
=
Since
and
then
396
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)

=
2
then the BL equilibrium expected returns and expected

returns based on the CAPM are identical.
397
The Matrix: Absolute Views

In the BL approach the investor/econometrician forms some
views about expected returns as described below.
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%.
398
The Matrix: Absolute Views

In general the number of views is .
In our case = 2.
Then is a 2 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.
399
The 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.
400
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 corresponds to momentum investing.
The second row corresponds to value investing.
Both the first and second rows contain elements.
401
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-tomarket ratio.
Growth stocks are 10% of the stocks having the lowest book-tomarket ratios.
402
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.
403
Back to the 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 sum of the first row
and the sum of the second row are both zero.
404
Back to the 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/
while the loser stock gets 10/.
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.
405
Computing the Vector

It is the 1 vector of views on expected returns.
Using the absolute views above = 0.05,0.12

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).
406
The Matrix
is a covariance matrix expressing uncertainty about
views.
It is typically assumed to be diagonal.
In the absolute views case described above 1,1 denotes
uncertainty about the first view while 2,2 denotes
uncertainty about the second view both are at the discretion
of the econometrician/investor.
407
The Matrix
In the relative views described above: 1,1 denotes
uncertainty about momentum. This could be the sample
variance of the momentum payoff.
2,2 denotes uncertainty about the value payoff. This is the
could be the sample variance of the value payoff.
408
Deciding Upon
There are many debates among professionals about the right
value of .
From a conceptual perspective it could be 1/ where denotes
the sample size.
You can pick = 0.01
You can also use other values and examine how they perform
in real-time investment decisions.
409
Maximizing Sharpe Ratio

The remaining task is to run the maximization program
max
Preferably, impose portfolio constraints, such that each of the

elements is bounded below by 0 and subject to some agreed
upon upper bound, as well as the sum of the elements is
equal to one.
410
Extending the BL Model

to Incorporate Sample Moments
Consider a sample of size , e.g., = 60 monthly observations.
Let us estimate the mean and the covariance () of our
assets based on the sample.
Then the vector of expected return that serves as an input for
asset allocation is given by
1
=
where
411
+ ( /)
1 1
1 + ( /)1
= ()1 + 1
Session #10: Risk Management:

Down Side Risk Measures
412
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.
413
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.
414
Value at Risk (VaR)

The 95% says that there is a 5% chance that the realized
return, denoted by , will be less than-95% .
More generally,
Pr 1 =
5%
95%
415
Value at Risk (VaR)

1
= 1 1 = + 1
where
1 , the critical value, is the inverse cumulative distribution
function of the standard normal evaluated at .
Let = 5% and assume that

, 2
The critical value is 1 0.05 = 1.64
416
Value at Risk (VaR)

Therefore
1 = + 1 = + 1.64
Check:
If
Then
417
, 2

Pr 1 = Pr
1.64
= Pr
<
= Pr < 1.64 = 1.64 = 0.05
Example: The US Equity Premium

Suppose:
0.95
0.08, 0.202
= 0.08 1.64 0.20 = 0.25
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,
0.975 = 0.08 1.96 0.20 = 0.31
418
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.
419
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.

420
Backtesting the VaR

Assume that stock returns are normally distributed with mean
and variance that vary over time
, 2 = 1,2, ,
The sample is of length .
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 to be equal to one; otherwise, it is zero.
421
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?
422
Model Verification Based on Failure Rates

To answer that question let us discuss another example
which requires a similar test statistic.
Suppose that 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.
423
Model Verification Based on Failure Rates

Suppose that the number of wrong analysts is .
Thus, the fraction of wrong analysts is = / this is the

failure rate.
424
The Test Statistic

The hypothesis to be tested is
0 : = 0
1 :
Under the null hypothesis it follows that

=
0 1 0
425
The Test Statistic

Notice that
= 0
= 0 1 0
Thus
=
426
0
0 1 0
0,1
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.
0 : = 12.6
1 :
427
20 12.6
0.05 0.95 252
= 2.14
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.
428
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
429
Expected Shortfall (ES)

and the Truncated Normal Distribution
Assume that returns are normally distributed:
, 2
1 = | + 1
1
430
1
= +

where is the pdf of the standard normal density
e.g 1.64 = 0.103961
This formula for ES is about the expected value of a truncated

normally distributed random variable.
431

Proof:
, 2
| 1
since
432
0
1
=
=
0
1
0 =
= 1
Expected Shortfall: Example

Example:
= 8%
= 20%
1.64
0.10
95% = 0.08 + 0.20
0.08 + 0.20
= 0.32
0.05
0.05
1.96
0.06
97.5% = 0.08 + 0.20
0.08 + 0.20
= 0.40
0.025
0.25
433

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.
434
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.
435
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.
436
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:

= min , 0
where is some target level.

For instance, =
Unlike the variance,
2 =
437

The target semi-variance:
1. Picks a target level as a reference point instead of the mean.
2. Gives weight only to negative deviations from a reference point.
438

Notice that if , 2
2
=
+ 2
+1
where
and are the PDF and CDF of a 0,1 variable,
respectively
Of course if =
then =
439
2
2
Maximum Drawdown (MD)

The MD (M) over a given investment horizon is the largest Mmonth 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.
440
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.
441
Downside Beta
1

min
min
, 0
, 0
The numerator in the equation is referred to as the co-semivariance of returns and is the covariance of returns below on
the market portfolio with return in excess of on security .
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.
442
Downside Beta
(2)
min , 0
=
min , 0 2
where and are security and market average return

respectively.
One can modify the down side beta as follows:
(3)
443
min , 0 min , 0
=
min , 0 2
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.
444
in Long Horizon Asset Management
Let us denote by the cumulative return on the investment
over several years (say years).
Rather than finding the distribution of R we analyze the
distribution of
= ln 1 +
which is the continuously compounded return over the
investment horizon.
445
The investment value after years is
= 0 1 + 1 1 + 2 1 +
Dividing both sides of the equation by 0 we get
= 1 + 1 1 + 2 1 +
0
Thus
1 + = 1 + 1 1 + 2 1 +
446
Taking natural log from both sides we get
= 1 + 2 + +
Assuming that
, 2
= 1, . . . ,
Then using properties of the normal distribution, we get

, 2
447
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 year period.
448

The value of such a riskfree investment is
= 0 1 +
= 0 exp
where is the continuously compounded risk free rate.
449

Essentially we ask: what is the probability that
<
This is equivalent to asking what is the probability that

<
0
0
This, in turn, is equivalent to asking what is the probability

that
ln
< ln
0
0
450

So we need to work out
<
Subtracting and dividing by both sides of the inequality
we get

<
We can denote this probability by

=
451

Typically < which means the probability diminishes the
larger .
Notice that the shortfall probability can be written as a
function of the Sharpe ratio of log returns:
=
452
Example
Take =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 norm.dist 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
453
Cost of Insuring against Shortfall

Let us now understand the mathematics of insuring against
shortfall.
Without loss of generality let us assume that
0 = 1
The investment value at time is a given by the random
variable
454

Once we insure against shortfall the investment value after T
years becomes
455
If > exp
you get
If < exp
you get exp

So you essentially buy an insurance policy that pays 0
if > exp
Pays exp if < exp

You ultimately need to price a contract with terminal payoff
given by
max 0, exp
456

This is a European put option expiring in years with
1. S=1
2. = exp .
3. Riskfree rate given by
4. Volatility given by
5. Dividend yield given by = 0
457

The B&S formula indicates that
= exp 2 exp 1
Which, given the parameter outlined above, becomes
1
1
=
2
2
458

To show the pricing formula of the put use the following:
1
d1 =
(/)+(+2 2 )

and 2 = 1
while 1 = 1 1
2 = 1 2
459

The B&S option-pricing model gives the current put price as
= 1 2
where
1 =
2 = 1
and is <
460

For = 0.2 (per year)
461
T (years)
0.08
0.18
10
0.25
20
0.35
30
0.42
50
0.52

We have found out that the cost of the insurance increases in ,
even when the probability of shortfall decreases in (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.
462
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
< exp ?
We assume, without loss of generality, that the initial invested

wealth is one.
463
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.
ln ~ , 2
Thus, let us first display some properties of the log normal
distribution.
464
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 is a standard normal variable.

The probability density function of a log-normal distribution is
465
The Log Normal Distribution

1
, , =
2 2
>0
2
If is a log-normally distributed variable, its expected value
and variance are given by
=
466
1
2
+ 2
2
2
2+
The Mean and Variance of

Using moments of the log normal distribution, the mean and
variance of are
1 2
( ) = exp ( + )
2
( ) = exp ( 2 + 2 ) exp ( 2 ) 1
Next, we aim to find the conditional mean.
467
The Mean of a Variable

that has the Truncated Log Normal Distribution
c E(x|x > c) =
1
2
x 2
where 1/ c is a normalizing constant.
Let us make change of variables:

(x)
= t x = et+ and dx = et+ dt
468

then:
c E xx>c =
=
=
=
469
2
1
+0.5 2
1
2 2
1
2 2 ++
1
2 2 + +0.5 2
+ =
1
2
2

Let us make another change of variables:
v = t dv = dt
c E(x|x > c) =
+0.5 2
1
2
1
2 2
The integral is the complement CDF of the standard normal

random variable.
470

Thus, the formula is reduced to:
+0.5 2
F c ELN x x > c =
1
2
+0.5
=
471

In the same way we can show that:
c ELN (x|x c) =
=
=
472
+0.5 2
+0.5 2
0 2

1
1
2
2
2
2
1
2 ++
2
+0.5 2
1
2
1
2
2
Punch Lines
+ + 2
ELN (x|x > c) = ELN (x)

+
ELN (x|x c) = ELN (x)

473
The Expected Value given Shortfall

The expected value given shortfall is
| =
1
+2 2
or
| =
474
1
+2 2
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.
475
The Horizon Effect

Numerical example
Lets take = 5%, = 10% . For different values of > the
conditional expectation over horizon looks like:
476
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
exp
The expected value of that insured amount given shortfall

sharply rises with the investment horizon, which explains the
increasing value of the put option.
477
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.
478
VaR with Log Normal Distribution

We are looking for threshold, VaR, such that
= Pr 0 < 0 = 0
Then in order to find the threshold we need to calculate
quantile of lognormal distribution:
= 0 1 ; , 2
+ 1
= 0
where 1 is as defined earlier.
479

Specifically,
= Pr 0 < 0 = Pr ln 0 < ln 0
= Pr
480
<

and then
= 0
= 1
1
+
That is to say that there is a % probability that the

investment value at time T will be below that VaR.
481
VaR with Distribution

Suppose now that stock returns have a distribution with
degrees of freedom and expected return and volatility given by
and
The pdf of stock return is formulated as
+1
2
)2
1
(
|, , =
1+
/2,1/2
482
Partial Expectation
Let =
-standardized
r.v distribution with |0,1, .
Than,
| =
, =
+1
2
2
1
1+
/2,1/2
1
2
2
1+
1
/2,1/2
+1
+1
2
2
1+
/2,1/2 1
483
|
=
+ 2
= ,
1
Partial Expectation
And thus
2
1 , + 1
=
1
1
Where = 1 is quantile of
484
Long Run Return

when Periodic Return has the -Distribution
The sum of independent t-distributed random variables is not
-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:
.
,
2
2
485
Long Run Return

Approximation makes sense for large vs when coincides with
normal distribution.
However, simulation studies show that for sufficient number of
periods this approximation works well enough.
486
Long Run Return

However simulations shows that for sufficient number of
periods this approximation works well enough.
Let = 0.01; = 0.05. The next graphs show normal curve fit
to the sum of r.v.s (over periods); sample estimates vs.
predicted parameters are includes
487
Long Run Return

= 10
900
800
= 0.102
= 0.273
700
600
= 0.1
= 0.274
500
=3
400
300
200
100
0
-3
488
-2
-1
Long Run Return

= 100
400
350
300
= 1.011
= 0.856
= 1
= 0.866
250
=3
200
150
100
50
0
-4
489
-3
-2
-1
Long Run Return

= 10
350
300
250
= 0.099
= 0.174
= 0.1
= 0.177
200
= 10
150
100
50
0
-0.6
490
-0.4
-0.2
0.2
0.4
0.6
0.8
Long Run Return

= 100
350
300
250
= 0.994
= 0.566
= 1
= 0.559
200
= 10
150
100
50
0
-1
491
-0.5
0.5
1.5
2.5
3.5
VaR Expected Shortfall using Extreme Value

Theorem (EVT) - Without Distribution Assumption
Estimating VaR and ES using EVT can be done in two ways:
The idea of sub-periods;
The idea based on exceedences 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.
492
Sub-Periods
Divide the sample into non-overlapping subsamples (to
preserve assumption) of length :
1 |+1 2 | |+1
Then using minimum of each sub-period estimate GEV density

of the following form
1+ 1/
, 0
=
, =
493
Sub-Periods
Using estimations of , and we can compute VaR
+
ln 1 1 , 0
=
+ ln ln 1
, 0
Here we used the assumption that returns are iid and then
single period return is related to return over sub-period n as
follows
< =
< = <
=1
494
Long Term VaR

The result of the EVT are also relevant for the task of
converting short-term VaR into long-term VaR.
Assume 1/ applies to a single-period return for
large R. Then we have for a -period return
1/
It follows that a multi-period VaR forecast of a fat tailed return

distribution under the iid assumption is given by
= 1/ 1
495
Exceedances over a Threshold

For a given threshold h conditional excess distribution function
on losses is defined as
= | > 0
In terms of this can be written as
(1)
496
+
1
Exceedances over a Threshold

For a large class of underlying distribution functions F the
conditional excess distribution function , for large, is
well approximated by generalized Pareto distribution:
1/
1 1+
, 0
=
1
, 0
For
0,
497
0, /
<0
Long Horizon VaR

Using this model VaR and ES can be expressed analytically:
First, using (1) we can isolate
= 1 +
Then is replaced by GP CDF and () by the sample
estimate
, where n is the total number of observations
and is the number of exceedances above the threshold :
1/
=1
1+
498
Long Horizon VaR

Next
=
=+
And
= + | >
is mean excess function for GP distribution which equals to
499
Long Horizon VaR

=
+
1+
1+
Given a s sample and some threshold h, parameters , and
can be estimated using MLE.
500
Session #11 (part a): Testing the

Black&Scholes Formula
501
Option Pricing
The B&S call Option price is given by
C(S, K, , r, T, ) = SeT N(d1 ) KerT N(d2 )
The put Option price is
P(S, K, , r, T, ) = KerT N(d2 ) SeT N(d1 )
1
Where d1 =
502
(/)+(+2 2 )

and d2 = d1 T
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.
503
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.
504
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.
505
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.
506
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?
507
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.
508
The Volatility Smile for Foreign Currency Options

Implied
Volatility
Strike
Price
509
Implied Distribution for Foreign Currency Options

Both tails are heavier than the lognormal distribution.
It is also more peaked than the lognormal distribution.
510
The Volatility Smile for Equity Options

Implied
Volatility
Strike
Price
511
Implied Distribution for Equity Options

The left tail is heavier and the right tail is less heavy than the
lognormal distribution.
512
Ways of Characterizing the Volatility Smiles

Plot implied volatility against /0 (The volatility smile is then
more stable).
Plot implied volatility against /0 (Traders usually define an
option as at-the-money when equals the forward price, 0 ,
not when it equals the spot price 0 ).
Plot implied volatility against delta of the option (This

approach allows the volatility smile to be applied to some nonstandard 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 50delta options).
513
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 stock prices
because of the impact of leverage. This is consistent with the skew
that is observed in practice.
514
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
515
Example of a Volatility Surface

0
516
0.90
0.95
1.00
1.05
1.10
1 month
14.2
13.0
12.0
13.1
14.5
3 months
14.0
13.0
12.0
13.1
14.2
6 months
14.1
13.3
12.5
13.4
14.2
1 year
14.7
14.0
13.5
14.0
14.8
2 years
15.0
14.4
14.0
14.5
15.1
5 years
14.8
14.6
14.4
14.7
15.0
Session #11 (part b): Time Varying Volatility

Models
517
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
518
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.
519
ARCH(1)
= +
= where ~ 0,1
2
2 = + 1
1 2 = 1 2 2 = 2
so
2 is the conditional variance.
520
ARCH(1)
2
1
= 2 is the unconditional variance.
2
2 = + 1
2
= + 1
2
2
= + 1
1
2
= + 1
2
2
2 = 2 = 1
= 2
521
ARCH(1)
Fat tail?
4 =
=
4
2
2
1 4 4
=3
522
1 4
2
1 2 2
1 2 2
4
2
34
ARCH(1)
The last step follows because:
2 = 4 2
so
4
2
2

Yes fat tail!
523
GARCH(1,1)
= +
= where ~ 0,1
2
2
2 = + 1
+ 1
2
2
2 = + 1
+ 1
524
GARCH(1,1)
2 = + 2 + 2
4 =
525
1
1
3 1++ 1
1232 2
>3
EGARCH
= + where ~ 0,1
=
ln
=+
1
1
The first component
1
1
1
2
2
+ ln 1
= 1
is the absolute value of a normally distribution variable 1

minus its expectation.
526
EGARCH
The second component is
= 1
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.
527
EGARCH
More formally if 1 < 0 then the log conditional variance rises
by + .
If 1 > 0 then the log conditional variance rises by
This produces the asymmetric volatility effect volatility is
higher during down market and lower during up market.
528
Stochastic Volatility (SV)

There is a variety of SV models.
A popular one follows the dynamics

= + where 0,1
ln = 0 + 1 ln 1 + where ~ 0,1
cov , = 0
Notice, that unlike ARCH, GARCH, and EGARCH, here the
volatility itself has a stochastic component.
529
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.
530
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.
531
Implied Volatility (IV)

The B&S call Option price is given by
C(S, K, , r, T, ) = SeT N(d1 ) KerT N(d2 )
The put Option price is
P(S, K, , r, T, ) = KerT N(d2 ) SeT N(d1 )
1
Where d1 =
532
(/)+(+2 2 )

and d2 = d1 T
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.
533
Session #11 (part c): Stock Return

Predictability, Model Selection, and Model
Combination
534
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 ~ = 0, 2

+1 | =
+1 | = 2
where is the continuously compounded return and is the
set of information available at time .
535
Return Predictability
Also note that the variance of a two-period return + +1 is
equal to
+ +1 + 2 cov , +1 = 2 2
That is, variance grows linearly with the investment horizon,

while volatility grows in the rate square root.
536
Variance Ratio Tests

However, is it really the case?
Perhaps stock returns are auto-correlated, or

cov , +1 0
then:
+ +1
+ +1 + 2 , +1
2 =
=
2
2
=1
537
,+1
+
+1
=1+

Test: 0 : = 0
1 : 0
The test statistic is
2 1 0,1
538

More generally,
+ +1 + + +
=
=1+2
+ 1
=1
+1
0 : = 1 no auto correlation
1 :
539
Variance Ratios
Test statistic:
1 0,
+1
e.g
4 1
=2
2 1 0,1
=3
3 1
540
20
0,
9
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
541
Predictive Variables
The aggregate book-to-market ratio.
The market volatility.

The market illiquidity
542
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 = + 1 1 + 2 2 + + + +1
543
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.
544
Long Horizon Predictive Regressions

+1,+ = + + +1,+
The dependent variable is the sum of log excess return over the
investment horizon, which is 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.
545
Newey-West Correction
Rewriting the long horizon regression
+1,+ = + +1,+
= 1,
= ,
546
Newey-West Correction
The estimation error of the regression coefficient is represented
by
= 1
where is the Newey-West given by serially correlated
adjusted estimator
=
=
547

=+1

Tradeoff:
Higher better coverage of dependence.
But we loose degrees of freedom.
Feasible solution:
548
1
3

E.g
K=1:
= 1, = 0, = 1
=
=1
Here we have no serial correlation.
549

E.g
K=2:
= 2, = 1, = 0, = 1, = 2
1 1
=
2
=1
1
+1 +
1
=
550
2
=1
1
=2
2 +
=1
1 1
+
2
+1
=1
In the Presence of Heteroskedasticity

1 1
=

=
=
551
=1

=1
+1

=1
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
1 : = + 1 +
2 :
= +
Which one dominates? If 1 then there is predictability

otherwise, there is no.
552
Out of Sample Predictability

One way to test predictability is to compute the out of sample
2 :
2
=1 ,1
2
= 1
2
=1
Where ,1 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.
553
Model Selection
When variable are potential candidates for predicting stock
returns there are 2 linear combinations of predictive models.
In the extreme, the model that drops all predictors is the nopredictability 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.
554
Model Selection Criteria

= 2 2 ln
where is the maximized value of the likelihood function.
= ln 2 ln
2
=1 1
1
1
= 1
Bayesian posterior probability
555
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.
556
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.
557

Financial Econometrics

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Financial Econometrics

Uploaded by

Copyright:

Available Formats

Financial Econometrics

Prof. Doron Avramov

The course targets advanced master level and PhD level

Syllabus: Grade Components

Syllabus: Topics to be Covered - #1

Syllabus: Topics to be Covered - #2

Forming mean variance efficient portfolio, the Global

Error Volatility Portfolio.

Syllabus: Topics to be Covered - #3

The Delta method

Principal component analysis.

Syllabus: Topics to be Covered - #4

Option pricing: testing the validity of the B&S formula

Syllabus: Topics to be Covered - #5

Normal, Bivariate normal, and multivariate normal densities

Plotting the Efficient Frontier

However, how could you

Mean, volatility, and correlation estimates are noisy as reflected

Sample Mean and Volatility

First pass: the mean and volatility are estimated as

Mean-Variance vs. Down Side Risk

Why do we need to resort to down side risk?

Down Side Risk

Moreover, there is a major weakness embedded in the mean

Drawback in the Mean-Variance Setup

Stock A dominates stock B in every state of nature.

The Normal Distribution

The normal distribution is symmetric.

Dispersion around the Mean

When is small, the distribution is concentrated and vice

Probability Distribution Function

The Cumulative Density Function (CDF), which is the integral

Probability Integral Transform

Normally Distributed Return

Confidence Intervals for

= < 0.4 = 34.4%

Kurtosis the fourth moment is three.

Mesokurtic - A term used in a statistical context where the

The Essence of Skewness and Kurtosis

Bivariate Normal Distribution

The marginal densities of x and y are

When the correlation between and is positive and >

meaning that the realization of does not say anything about .

Conditional Standard Deviation

When goodness of fit is higher the conditional standard

mean? A few rules:

The Chi-Squared Distribution

If 1~(0,1), 2~ (0,1) & 1 2 then:

More about the Chi-Squared

where , is beta function and 1 is the number of

The Students t Distribution

Rewriting the system in a matrix form

Vectors and Matrices: some Rules

The identity matrix satisfies

The inverse of the matrix is 1 which satisfies:

Solving Large Scale Linear Equations

Linear Independence and Norm

A Positive Definite Matrix

is called positive definite if

for any nonzero vector V.

The Trace of a Matrix

This is the vectorization operator

The VECH Operator

but takes only the upper triangular