Lecture Notes: part 1

Financial Econometrics
2020-2021

Francisco Blasques
and
Paolo Gorgi
2
Contents

1 Introduction 5
1.1 Models for the conditional mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.1 Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.2 AR(1) model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Properties of Financial Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Random walk of stock prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Volatility clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

I Observation Driven Models 13

2 Autoregressive Conditional Heteroskedasticity Models 15
2.1 The ARCH(1) model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The ARCH(q) model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Generalized ARCH Models 23

3.1 The GARCH(1,1) model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 The GARCH(p,q) model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Simulating GARCH models with R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4 Parameter Estimation 29
4.1 Deriving the likelihood function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Maximum Likelihood Estimator and Asymptotic properties . . . . . . . . . . . . . . . . . . . 32
4.3 Statistical Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 Numerical Optimization of the log-Likelihood Function . . . . . . . . . . . . . . . . . . . . . . 36
4.5 Estimating GARCH models with R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5 Financial Analysis of ARCH and GARCH Models 41

5.1 Estimation of the conditional volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Diagnostic tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.3 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.4 Value-at-Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.5 Forecasting conditional volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.6 Forecasting VaR and conditional density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.7 News Impact Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

6 Multivariate GARCH models 49

6.1 The VECH model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 The DVECH model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.3 The scalar DVECH model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.4 The BEKK model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3
 6.5 The CCC model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.6 The DCC model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.7 Other extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.8 Simulate from a bivariate DVECH(1,1) with R . . . . . . . . . . . . . . . . . . . . . . . . . . 57

7 Estimation of multivariate GARCH models 59

7.1 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.2 Estimating a bivariate scalar DVECH with R . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.3 Estimation of the sDVECH model with covariance targeting . . . . . . . . . . . . . . . . . . . 62
7.4 Estimating a sDVECH model with CT in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.5 Estimation of the CCC model equation by equation . . . . . . . . . . . . . . . . . . . . . . . 64
7.6 Estimating a CCC model equation by equation with R . . . . . . . . . . . . . . . . . . . . . . 64

8 Financial Analysis of Multivariate GARCH models 67

8.1 VaR portfolio prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
8.2 Dynamic portfolio optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.3 Dynamic portfolio optimization with R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.4 Out-of-sample evaluation of different portfolio strategies . . . . . . . . . . . . . . . . . . . . . 69

A Stock Return Properties: Empirical Evidence 71

4
Chapter 1

Introduction

From an econometric methodology perspective, this course is essentially devoted to the art of specifying
time-varying parameter models, and using them to conduct inference, probabilistic analysis, policy analysis
and forecasting. As we shall see, time-varying parameter models can be divided into two broad categories:
observation-driven models and parameter-driven models. Both classes of models are capable of describing
the temporal dynamics of time-series featuring time-varying conditional volatilities, time-varying tail prob-
abilities, time-varying regression coefficients, time-varying conditional moments of higher-order, and much
more!
While observation-driven models and parameter-driven models can be used to describe similar features
in the data, they actually approach the data in very different ways and require distinct statistical tools and
techniques.
Part 1 of these lecture notes is devoted to the study of observation-driven models for conditional volatility.
Part 2 is devoted to parameter-driven models for volatility and to other extensions. The focus is on the
practical implementation and analysis of these models. In the remainder of this introductory chapter we
shall use time-series of financial returns as a motivation for the use of time-varying parameter models. In
particular, financial return data clarifies the need to go beyond models of the conditional mean (like linear
regression and ARMA models) and make use of models that can describe time-variation in conditional
volatilities. Below, we provide first a quick recap of linear regression and ARMA models. Next, we show the
limitations of using these models for analyzing financial returns.

1.1 Models for the conditional mean

In this section, we shall revisit some basic models for the conditional mean: the linear regression model and
the autoregressive model.

1.1.1 Linear regression

In your introductory econometrics courses, you have most likely learned about regression. In particular, by
now, you should be familiar with the linear regression model

yt = α + βxt + εt

where yt is the dependent or endogenous variable, xt is the independent or explanatory variable, εt is the
error term or innovation, and α and β are the fixed unknown parameters typically called intercept and slope
respectively.
You should also remember that when the error term εt satisfies the assumption that E(εt |xt ) = 0, then
the linear regression model is a model of the conditional expectation of yt given xt . In other words, we have

5
 E(yt |xt ) = E(α + βxt + εt |xt )
= α + β E(xt |xt ) + E(εt |xt )
| {z } | {z }
=xt =0

= α + βxt .
The errors εt account for the fact that the relation between yt and xt holds only “on average”. The linear
regression model states essentially that, on average, the dependent variable yt is linearly related to the
explanatory variable xt .
The parameter β measures the expected change in yt given a unit change in xt , and the parameter α measures
the expected value of yt when xt = 0. If the parameters α and β were known, then the average relation
between yt and xt would also be known, and econometricians would not be needed! Fortunately however, α
and β are unknown, and hence, they must be estimated from the data!

1.1.2 AR(1) model

In your introductory econometrics courses you surely learned about time-series models. Most probably, you
studied linear time-series models like the autoregressive model of order 1, also called the AR(1) model. A
sequence {xt }t∈Z is said to follow an AR(1) process if

xt = φxt−1 + εt ∀t∈Z

where {εt }t∈Z is a white noise sequence with E(εt |xt−1 ) = 0. A white noise sequence is a sequence that is
serially uncorrelated, has mean zero E(t ) = 0, and finite unconditional variance Var(t ) = σ 2 .
Linear autoregressive models like the AR(1) are very useful in modeling the temporal dependence that we
usually observe in economic and financial time-series. In the AR(1) model, this dependence can be well
understood by noting that the conditional expectation of xt depends on the value of xt−1 . In particular, the
conditional expectation of xt given xt−1 is E(xt |xt−1 ) = φxt−1 .
You may also remember from your introductory courses that a time series {xt }t∈Z is weakly stationary if
its mean, variance and autocovariance function are invariant in time. Figure 1.1 shows a typical path of a
time-series generated by an AR(1) model with time-varying conditional distribution, but the multiple paths
reveal the time-invariant unconditional distribution.
Definition 1.1. (Weak Stationarity) A time-series {xt }t∈Z is said to be weakly stationary if the mean E(xt )
and the variance Var(xt ) are constant in t and the autocovariance function Cov(xt , xt−h ) is constant in t,
for each h.
You also learned that the linear Gaussian AR(1) model is weakly stationary as long as |φ| < 1. In other
words, the Gaussian AR(1) is stationary as long as it does not exhibit ‘too much’ temporal dependence.
Theorem 1.1. Let {xt }t∈Z be a time-series generated by the linear Gaussian AR(1) model

xt = φxt−1 + εt ∀t∈Z

with |φ| < 1 and innovations {εt }t∈Z that are white noise. Then {xt }t∈Z is weakly stationary.
This stationarity property of the time-series is important for understanding the properties of estimators
because it allows us to make use of laws of large numbers and central limit theorems.
As you have most likely learned in your introductory time series course, the AR(1) model presented
above can be generalized to the class of ARMA models. The AR(1) and ARMA models are useful for the
conditional mean in time series.

1.2 Properties of Financial Returns

Models for the conditional mean are useful in many empirical applications, but here we will argue that they
are not very useful for modeling financial returns. We will see that stock returns are basically unpredictable

6
Figure 1.1: Single path [above] shows time-varying conditional mean. Multiple paths [below] show invariance of the distribu-
tion (mean and variance are clearly constant over time).

in mean and therefore ARMA models are not of great use. Instead, the variance (or volatility) of stock
returns can be predicted. This justifies the need of models that go beyond the conditional mean and account
for time-variation in the conditional variance.

1.2.1 Random walk of stock prices

In this section we turn our attention to time-series of stock prices. In particular, we look at the stock prices of
the companies listed in the Standard and Poor’s top 100 companies in the US; commonly known as S&P100.
We denote with pt the price of a certain stock at time t.
We will argue that the time series {pt }t∈Z of each individual stock price seems to behave essentially like
a random walk. We shall say that a time series {pt }t∈Z follows a random walk if we have

pt = pt−1 + t ,

where {t }t∈Z is a white noise sequence with E(t |pt−1 ) = 0. The random walk dynamics imply that stock
prices are essentially impossible to forecast. In other words, the best forecast p̂t+1 for the price at time t + 1
conditional on the data until time t is simply given by the last observed price p̂t+1 = pt . This is easy to
show since
p̂t+1 = E(pt+1 |pt )
= E(pt + t+1 |pt )
= E(pt |pt ) + E(t+1 |pt )
= pt + 0 = pt .

Below we provide empirical evidence that stock prices essentially behave like random walks by studying stock
prices of several stocks. Naturally, if stock prices behave like random walks, then we should be able to find
evidence that they are unit-root non-stationary. Furthermore, we should also find that variation of stock
prices (referred as returns or log-returns) are not only stationary but white noise.
First we investigate whether stock prices are unit root non-stationary. Figure 1.2 plots the daily stock
prices of Apple and Intel over a period of 10 years, starting in 2006. The figure indeed suggests that stock

7
prices are non-stationary since their mean is not constant over time. The non-stationarity assumption can
also be formally tested using a Augmented Dickey Fuller (ADF) unit-root test. Table 1.1 reports the p-value

Apple Intel
150 40

30
100

20

50
10

0 0
2006 2008 2010 2012 2014 2016 2006 2008 2010 2012 2014 2016

Figure 1.2: Daily stock prices of Apple and Intel from August 2006 to August 2016
of the Augmented Dickey Fuller (ADF) unit-root test applied to the daily, weekly and monthly stock prices
of Apple and Intel. We can see that the test suggests that stock prices are indeed non-stationary. Table
1.2 below reports the fraction of times that the null hypothesis of a unit-root is rejected for the prices of all
stocks in the S&P100 index. There is overwhelming evidence of non-stationarity in stock prices. The results
for each stock in the S&P100 index are can be found in Tables A.1 and A.2 in Appendix A.

Table 1.1: P-values of ADF test for Apple and Intel stock prices

daily weekly monthly

Apple 0.239 0.188 0.230

Intel 0.313 0.356 0.115

Table 1.2: Fraction of H0 rejections for ADF test over all S&P100 stock prices

daily weekly monthly

0.00 0.00 0.00

We now focus our attention on returns (or log returns), i.e. price variations. Studying the properties of
returns (or log-returns) is of great interest because we are typically interested in how risky or remunerative
a certain investment is. Therefore we are actually more interested in the price variation more than the price
level itself. Furthermore, if stock prices are random walks then we should find that returns (or log-returns)
are white noise. In practice we focus on log-returns instead of returns.
Log-returns are defined as first differences of log-prices. In particular, log-returns {yt }t∈Z are
obtained as  
pt
yt = log(pt ) − log(pt−1 ) = log .
pt−1
We work with first differences of log-prices instead of prices because they have some appealing
properties. For instance, log-returns are a good approximation for returns rates
 
pt pt − pt−1
yt = log ≈ .
pt−1 pt−1

Therefore, if yt = 0.01 we can say that the price from time t − 1 to t increased of about 1%.
Throughout these notes we will always work with log-returns and sometimes for convenience we
shall refer to them as simply returns.

8
 Figure 1.3 plots the daily log-returns of Apple and Intel. The figure suggests that the log-returns of these
stocks are stationary. In particular, we can see that the mean seems to be constant over time. Table 1.3

Apple Intel
0.2 0.2

0.1 0.1

0 0

-0.1 -0.1

-0.2 -0.2
2006 2008 2010 2012 2014 2016 2006 2008 2010 2012 2014 2016

Figure 1.3: Daily log-returns of Apple and Intel from August 2006 to August 2016

reports the results of the ADF unit-root test applied to the daily, weekly and monthly log-returns of Apple
and Intel. Shaded values are significant at the 95% confidence level. We can see that the test suggests that
log-returns are stationary. Table 1.4 below reports the fraction of times the null hypothesis of a unit-root is
rejected for the log-returns of all stocks in the S&P100 index. The results further confirm the stationarity
of log-returns.

Table 1.3: P-values of ADF test for Apple and Intel log-returns

daily weekly monthly

Apple 0.001 0.001 0.001

Intel 0.001 0.001 0.001

Table 1.4: Fraction of H0 rejections for ADF test over all S&P100 stocks

daily weekly monthly

1.00 0.99 0.98

Having established the non-stationarity of stock prices and the stationarity of stock returns, we now
move further and investigate whether log-returns are white noise. If log-returns are white noise we should
find that they are uncorrelated. Figure 1.4 shows the estimated autocorrelation function for the daily stock
returns of Apple and Intel, ranging over 25 lags. The significance bounds (in red) reveal that there is little
evidence autocorrelation in the stock returns of Apple and Intel at the daily frequency. The evidence for
temporal dependence is even weaker at lower frequencies. Figure 1.5 plots the sample autocorrelation for
weekly returns. There is no evidence of autocorrelation of weekly log-returns.

Apple Intel

0.2 0.2

0.1 0.1

0 0

-0.1 -0.1

-0.2 -0.2

-0.3 -0.3
1 3 5 7 9 11 13 15 17 19 21 23 25 1 3 5 7 9 11 13 15 17 19 21 23 25

Figure 1.4: Sample ACF for daily log-returns of Apple and Intel

9
 Apple Intel

0.2 0.2

0.1 0.1

0 0

-0.1 -0.1

-0.2 -0.2

-0.3 -0.3
1 3 5 7 9 11 13 15 17 19 21 23 25 1 3 5 7 9 11 13 15 17 19 21 23 25

Figure 1.5: Sample ACF for weekly log-returns of Apple and Intel

Table 1.5 reports the estimated coefficient of an MA(1) model and the estimated coefficient of an AR(1)
model for Apple and Intel log-returns, measured at the daily, weekly, and monthly frequencies. Shaded
values are significant at the 5% confidence level. We can see that only the coefficients for daily log-returns
are significantly different from zero. Table 1.6 further confirms that the presence of autocorrelation in stock
returns depends to a large extent on the frequency at which returns are observed. In particular, this table
reports the frequency with which the MA and AR coefficients are found to be statistically significant at the
5% level, over all the stocks in the S&P100 index. The results for each individual stock in the S&P100 index
are reported in Tables A.3 and A.4 in Appendix A.

Table 1.5: Estimates of MA(1) and AR(1) coefficients for Apple and Intel log-returns

MA(1) AR(1) MA(1) AR(1) MA(1) AR(1)

daily daily weekly weekly monthly monthly

Apple -0.026 -0.026 0.040 0.037 0.035 0.036

Intel -0.044 -0.042 -0.040 -0.038 -0.038 -0.049

Table 1.6: Fraction of significant coefficients (5% level) over all the S&P100 log-returns

MA(1) AR(1) MA(1) AR(1) MA(1) AR(1)

daily daily weekly weekly monthly monthly

0.6337 0.6238 0.4356 0.4059 0.1584 0.1287

Overall, it seems fair to say that there is evidence of significant but weak autocorrelation in log-returns
at the daily frequency. So daily stock log-returns are not exactly white noise but the autocorrelation is
basically negligible and therefore we can essentially consider daily log-returns as white noise. At the weekly
and monthly frequencies there is not much evidence of autocorrelation in log-returns and therefore the white
noise assumption is well suited in these cases.

1.2.2 Volatility clustering

From the discussion in the previous paragraph, we can conclude that ARMA models, and more in general
models for the conditional mean, are not very useful to describe log-returns. Does this mean that log-returns
cannot be predicted? Well, the mean seems to be unpredictable but if we a closer look at Figure 1.3 we
can notice that there are time periods where the variability of the log-returns his higher and periods where

10
it is lower. For instance, we can see that around 2008 the volatility (variability) of log-returns is higher.
This changes of variability are well known as volatility clusters. In the following we see that past log-returns
can be useful to predict the volatility of future log-returns. Predicting volatility is of key importance since
volatility is one of the most important measures of financial risk.
Now instead of analyzing log-returns we consider squared log-returns, i.e. the square of log-returns. Given
that log-returns have a mean of approximately zero, squared log-returns offer a natural indicator of scale. As
such, the clusters of volatility may reveal themselves through autocorrelation in squared log-returns. Figure
1.6 plots the daily squared log-returns of Apple and Intel. The figure shows that squared log-returns tend

Apple Intel
0.04 0.02

0.03 0.015

0.02 0.01

0.01 0.005

0 0
2006 2008 2010 2012 2014 2016 2006 2008 2010 2012 2014 2016

Figure 1.6: Daily squared log-returns of Apple and Intel stocks.

Apple Intel

0.2 0.2

0.1 0.1

0 0

-0.1 -0.1

-0.2 -0.2

-0.3 -0.3
1 3 5 7 9 11 13 15 17 19 21 23 25 1 3 5 7 9 11 13 15 17 19 21 23 25

Figure 1.7: Sample ACF for the squared daily log-returns of Apple and Intel.

to be higher during the financial crisis in 2008. Figure 1.7 reports the autocorrelation function of the squared
log-returns. The figure provides strong evidence of autocorrelation in squared log-returns for both Apple and
Intel. The temporal dependence of squared log-returns is also made clear in Table 1.7. The table reports the
MA(1) and AR(1) coefficient estimates for the daily squared log-returns of Apple and Intel. Similar results
are obtained for other stocks in the S&P100 index as reported in tables A.7 and A.8 in Appendix A. These
results indicate the need of developing models that are able to describe the volatility clustering and capture
autocorrelation of financial squared log-returns.

Table 1.7: Estimates of MA(1) and AR(1) coefficients for daily squared log-returns

MA(1) AR(1)

Apple 0.184 0.178

Intel 0.189 0.183

Finally, we investigate other features of log-returns. Table 1.8 below suggests also that we will need to use
models capable of explaining non-Gaussian fat-tailed data. Large sample kurtosis (larger than 4) suggests

11
that the density of stock returns has tails that are fatter than those of a normal density. The Jarque-Bera
test statistic also suggests the rejection of the null hypothesis of Gaussian returns at a 5% significance level.
The evidence for the non-Gaussianity of stock returns is stronger at higher frequencies (e.g. for daily returns)
and weaker at lower frequencies (e.g. for weekly and monthly returns). Tables A.5 and A.6 provide similar
results for each of the stocks in the S&P100 index.

Table 1.8: Estimated moments and p-value of the Jarque-Bera test for Intel and Apple log-returns

Stock Mean Var Skew Kurt JB

Apple 0.006 0.036 -4.979 47.712 0.001

Intel -0.003 0.017 -1.873 10.394 0.001

The evidence for strong temporal dependence in squared stock returns will force us to go beyond models
of the conditional expectation. The evidence for fat-tailed data also requires us to consider alternative
models. In general, the linear-Gaussian regression models that were useful to us in the past are no longer up
to the task. The ARMA models, autoregressive distributed lag (ADL) models, or error correction models
(ECM) that you studied in your introductory time-series course, are not suitable to address the problems
at hand. In the coming chapters we shall explore uncharted territory! Together, we will design and study
models that can explain changes in volatilities, correlations, and other features in the data!

12
 Part I

Observation Driven Models

13
Chapter 2

Autoregressive Conditional
Heteroskedasticity Models

In this first part of the course, we study a class of models that are capable of describing time-series that
are uncorrelated over time but exhibit time-varying conditional volatility. These models are especially well
suited to describe the dynamics of financial returns.
In Chapter 2, we introduce the Autoregressive Conditional Heteroskedaticity (ARCH) model. As you
may recall from your introductory econometrics courses, the term heteroskedaticity refers to the variance
not being constant. In contrast, a time-series that is homoeskedastic is a time-series with constant fixed
variance. In this section, we will often talk about time-series with time-varying volatility rather than time-
varying variance since the variance may not always exist.1
In Chapter 3 we also look at a more general model called Generalized Autoregressive Conditional Het-
eroschedaticity (GARCH) model. Finally, in Chapter 6 we extend our models to the multivariate setting.
Multivariate models are capable of explaining not only the time-varying conditional volatilities, but also, the
time-varying conditional correlations of multiple financial returns.
These models, both univariate and multivariate, are often called time-varying parameter models. Fur-
thermore, in this chapter, all models belong to the class of observation-driven models. As we shall see, these
time-varying parameter models are said to be observation-driven since past observations are used to update
the values of the unobserved time-varying parameter.

2.1 The ARCH(1) model

Consider a sequence of financial returns {y1 , y2 , y3 , ...}. The Autoregressive Conditional Heteroschedaticity
(ARCH) model describes the dynamics of returns as

yt = σt ε t (2.1)

where σt is the conditional volatility at time t, and εt is an independent and identically distributed sequence
of shocks with mean zero and unit variance {εt }t∈Z ∼NID(0, 1).
In econometrics, equation (2.1) is called the observation-equation. This equation tells us how each observed
financial return yt is obtained from the unobserved conditional volatility σt and the unobserved shocks εt .
In addition to the observation equation stated above, we need an equation that tells us how the conditional
volatility σt evolves over time. In the case of the first-order ARCH model, labeled ARCH(1), the parameter
updating equation is given by

σt2 = ω + α1 yt−1
2
, ∀t∈Z (2.2)
1 Recall from your introductory probability courses that a random variable with fat tails (e.g. a Cauchy random variable, or

a student-t random variable with 2 degrees of freedom) may not have a finite variance.

15
where ω > 0 and α1 ≥ 0 are parameters that determine the behavior of the conditional volatility.
The idea behind the ARCH(1) updating equation is to capture time variation in the variance and, in this
way, describe the “volatility clustering” that is typically observed in stock returns. The squared observation
2 2
yt−1 can be seen as an estimate of the variance at time t − 1. When yt−1 is large, then σt2 also tends to be
2
large (for positive α1 ). Therefore, through the observation equation, yt is more likely to be large as well.
As a result, large (small) values of the variance at time t − 1 are likely to produce large (small) values of the
variance at time t. This exactly reflects the “volatility clustering” mentioned above. As we shall see in the
following, the variance is time varying only conditional on the past (for this reason ARCH models are said
to be “Conditional Heteroschedastic”). The marginal variance of the ARCH model is not time varying.

Definition 2.1. ARCH(1) Model: The ARCH(1) model is given by

yt = σt εt , σt2 = ω + α1 yt−1
2
, ∀t∈Z (2.3)

where ω > 0 and α1 ≥ 0 are parameters to be estimated, {εt }t∈Z is an exogenous N ID(0, 1) sequence.

Remark 2.1. The parameters ω and α1 are constrained to be non-negative to ensure that σt2 is positive.
Furthermore, as we shall see later, ω is strictly positive to guarantee that the unconditional variance is
non-zero.
Let us now analyze carefully the properties of this model. In particular, we are interested in verifying if the
ARCH(1) is capable of describing the main features of financial returns. Theorem 2.1 shows that, conditional
on past returns Y t−1 = {yt−1 , yt−2 , yt−3 , ...}, the distribution of yt is Gaussian with mean zero and variance
σt2 . As such, yt has a conditional variance σt2 that is time-varying.
Theorem 2.1. The conditional distribution of yt given the past Y t−1 is normal with mean E(yt |Y t−1 ) = 0
and variance Var(yt |Y t−1 ) = σt2 , namely yt |Y t−1 ∼ N (0, σt2 ).
Proof. The conditional mean is obtained as

E(yt |Y t−1 ) = E(σt εt |Y t−1 ) = σt E(εt |Y t−1 ) = σt E(εt ) = σt · 0 = 0,

where the second equality follows because σt is a constant conditional on Y t−1 and the third equality follows
from the independence of εt and the past Y t−1 . Similarly, the conditional variance is obtained as

Var(yt |Y t−1 ) = Var(σt t |Y t−1 ) = E(σt2 2t |Y t−1 ) = σt2 E(2t |Y t−1 ) = σt2 E(2t ) = σt2 · 1 = σt2 ,

where the third equality follows because σt2 is a constant conditional on Y t−1 and the fourth equality follows
from the independence of εt and the past Y t−1 . Finally, we have to show that the conditional distribution
of yt = σt εt given Y t−1 is normally distributed. This can be noted from the fact that, conditional on Y t−1 ,
the factor σt is a constant and εt ∼ N (0, 1). Therefore, since a Normal random variable multiplied by a
constant is normal as well, we can immediately conclude that yt |Y t−1 is normal. 
Theorem 2.1 is important because it tells us the conditional distribution of yt given the past. Therefore we
can use this result to calculate the probability of extreme events conditional on the recent behavior of the
market.
Figure 2.1 plots the time-series and the conditional volatility σt simulated from an ARCH(1) model
for different values of ω and α1 . It is clear that the ARCH(1) model is capable of generating clusters
of volatility. Furthermore, larger values of α1 are responsible for a stronger clustering behavior of the
conditional variance. Figure 2.2 shows that α1 determines the autocorrelation in squared returns yt2 but not
in the returns themselves.
Theorem 2.2 complements Figure 2.2 by showing that the returns yt generated by the ARCH(1) model are
indeed uncorrelated.

16
 y t with α 1 =0.4 y t with α 1 =0.8
4 4

2 2

0 0

-2 -2

-4 -4
0 500 1000 0 500 1000

σ t with α 1 =0.4 σ t with α 1 =0.8

4 4

3 3

2 2

1 1

0 0
0 500 1000 0 500 1000

Figure 2.1: Path of returns (black line) and volatility (red lines) simulated from an ARCH(1) model with
(ω, α1 ) = (0.1, 0.4) (left graphs) and (ω, α1 ) = (0.1, 0.8) (right graphs).

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4
1 3 5 7 9 11 13 15 1 3 5 7 9 11 13 15

Figure 2.2: Sample ACF for squared returns (left graph) and returns (right graph) obtained from a sample
of size T = 2000 observations simulated from an ARCH(1) model with (ω, α1 ) = (0.1, 0.4).

Theorem 2.2. The returns {yt }t∈Z generated by an ARCH(1) model have zero autocovariance at any lag.
Hence they are uncorrelated over time.
Proof. The autocovariance function for any l > 0 is given by
Cov(yt , yt−l ) = E(yt yt−l ) = E(E(yt yt−l |Y t−1 )) = E(yt−l E(yt |Y t−1 )) = E(yt−l · 0) = 0,
where the second equality follows by the law of total expectation and the third equality follows from the fact
that conditional on Y t−1 we have that yt−l is a constant for any l > 0. Given that all autocovariances are
zero, then the autocorrelation is also zero at any lag. 

17
While the ARCH(1) model is capable of generating a time-varying conditional variance, this does not mean
that the unconditional variance of returns {yt }t∈Z defined by the ARCH(1) changes over time. In fact, it
is possible to show that the unconditional distribution of {yt }t∈Z is invariant in time when α1 < 1. In
what follows we show that the unconditional mean, variance and autocovariances of returns generated by an
ARCH(1) model are all finite and time-invariant, so that {yt }t∈Z is weakly stationary.
We have already established in Theorem 2.2 that the unconditional autocovariances are all equal to zero
at any lag. Theorem 2.3 shows that also the unconditional mean is invariant in time and equal to zero.
Theorem 2.3. The returns {yt }t∈Z generated by an ARCH(1) model with α1 < 1 have unconditional mean
zero.
Proof. We know that the conditional mean E(yt |Y t−1 ) is equal to zero. Therefore we obtain that

E(yt ) = E(E(yt |Y t−1 )) = E(0) = 0,

by an application of the law of total expectation. 

Finally, we turn to the unconditional variance of the returns {yt }t∈Z generated by an ARCH(1) model. We
establish that the unconditional variance is finite and time-invariant in three steps. First, we show that
the squared observations yt2 of an ARCH(1) model follow an AR(1) model. This is known as the AR(1)
representation of the ARCH(1) model. This result is useful to show that when α1 < 1 the unconditional
variance of yt exists and also to find its value.
Theorem 2.4. Let {yt }t∈Z be generated by an ARCH(1) model. Then {yt2 }t∈Z follows an AR(1) model

yt2 = ω + α1 yt−1
2
+ ηt

where {ηt }t∈Z is a white noise sequence.

Proof. Define first a new error term ηt as ηt = yt2 − σt2 . It can be shown that {ηt }t∈Z is a white noise
sequence. Using this new definition, we go ahead and substitute σt2 for yt2 − ηt in the updating equation for
σt2 . This yields
yt2 − ηt = ω + αyt−1
2
.
which is naturally equivalent to
yt2 = ω + α1 yt−1
2
+ ηt .
Therefore we conclude that {yt2 }t∈Z follows an AR(1) process. 
Naturally you may ask: why is the AR representation useful? Well, the AR representation turns out to
be useful for model selection. It is common practice in empirical applications to obtain the autocorrelation
functions of the squared observations. The AR(1) representation tells us that the squared observation of an
ARCH(1) model should have an exponentially decreasing ACF and only the first lag of the PACF should be
different from zero.
Another advantage of the AR representation of squared returns is that you can use your knowledge of
time-series econometrics to obtain the unconditional variance of yt when α1 < 1. This result is shown in
Theorem 2.5.
Theorem 2.5. Let {yt }t∈Z be generated by an ARCH(1) model. If α1 < 1, then the unconditional variance
of yt is time-invariant and, in particular, given by Var(yt ) = ω/(1 − α1 ).
Proof. First, we note that
Var(yt ) = E(yt2 ).
Then, we can unfold the AR(1) representation of Theorem 2.4 and obtain
∞
X ∞
X ∞
X
yt2 = α1i ω + α1i ηt−i = ω/(1 − α1 ) + α1i ηt−i .
i=0 i=0 i=0

18
Therefore, since we know that E(ηt ) = 0 for any t we can conclude that
∞
X
E(yt2 ) = ω/(1 − α1 ) + α1i E(ηt−i ) = ω/(1 − α1 ).
i=0

Having established in Theorems 2.2, 2.3 and 2.5 that the returns {yt }t∈Z generated by an ARCH(1) model
have zero unconditional mean, fixed unconditional variance, and are uncorrelated at any lag, we are now in
a position to conclude that {yt }t∈Z is a weakly stationary white noise sequence2 .

Corollary 2.1. Let {yt }t∈Z be generated by an ARCH(1) model with α1 < 1. Then, {yt }t∈Z is a weakly
stationary white noise sequence.

Remark 2.2. The condition α1 < 1 is also sufficient for strict stationarity but not necessary.

It is important to note that the ARCH(1) model is not only capable of describing the clusters of volatility
observed in financial returns. It can also generate the fat tails (e.g. large kurtosis) observed in their uncondi-
tional sample distribution. Indeed, while the conditional distribution of yt given the past Y t−1 is Gaussian,
the unconditional distribution of yt is non-Gaussian. Figure 2.3 plots the unconditional distribution of re-
turns yt generated by an ARCH(1) model. The plots shows that the distribution of yt has fatter tails than
a Normal distribution.

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
0.01 0.01

0.005 0.005

0 0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

Figure 2.3: Unconditional density of data simulated from an ARCH model with (ω, α) = (0.1, 0.4) (left
graph) and (ω, α) = (0.1, 0.8) (right graph). The two bottom figures provide a ‘zoom in’ on the tails of each
density.

Curiously there is not a closed form analytic expression for the unconditional probability density function
of yt . Nonetheless, some properties of the unconditional distribution can be derived from higher order mo-
ments. Theorem 2.6 derives the Kurtosis is an indicator of how fat are the tails of a probability distribution,
namely how likely extreme observations are. The Kurtosis of the unconditional distribution of the ARCH
model is larger than 33 . This is coherent with what we typically observe in stock returns, namely a sample
Kurtosis larger than 3.
2 In introductory time-series courses you have focused on models of the conditional mean (e.g. ARMA) models. When

modeling the conditional mean, a white noise sequence is seen essentially as unstructured noise. In contrast, when focusing
on modeling higher-order moments (like the conditional variance), a white noise sequence may still contain structure to be
explored. The aim of ARCH models is to exploit this information to predict volatility.
3 The Kurtosis of the Normal distribution is equal to 3

19
Theorem 2.6. Let {yt }t∈Z be generated by an ARCH(1) model with α1 < √1 . Then the kurtosis of yt is
3
given by
E(yt4 ) 3(1 − α12 )
ku = 2 = .
E(yt ) 2 1 − 3α12
Proof. We will not discuss the details of this result. 

2.2 The ARCH(q) model

Very often, the conditional variance of stock returns shows strong persistence over time. This can be noted
from the autocorrelation functions of the squared observations. Consider, for example, the autocorrelation
function of the squared log-returns of the S&P 500 stock index in Figure 2.2. Clearly, an ARCH(1) model
will not be appropriate to model this type of dependence since the ACF does not decay exponentially.

0.4

0.2

-0.2

-0.4
1 3 5 7 9 11 13 15 17 19 21 23 25

Figure 2.4: ACF with 95% confidence intervals, obtained from daily squared log-returns of the S&P 500
stock index

Luckily, we can take this empirical evidence into account, by including more lags of yt2 in the updating
2
equation for the conditional variance σt2 . When the updating equation for σt2 depends not only on yt−1 , but
2
also, on yt−2 , then the resulting model is called an ARCH model of order 2, or ARCH(2). The ARCH(2)
model is given by

yt = σt t , {εt }t∈Z ∼ NID(0, 1) ,

σt2 2
= ω + α1 yt−1 2
+ α2 yt−2 , ∀t∈Z

where ω > 0, α1 ≥ 0 and α2 ≥ 0. In general, one may include as many lags of yt2 as desired for achieving a
correct description of the temporal dynamics of squared returns. When q lags are used, the resulting model
is called the ARCH model of order q, or ARCH(q) model.

Definition 2.2. ARCH(q) Model: The ARCH(q) model is given by

yt = σt t , {εt }t∈Z ∼ NID(0, 1) ,

q
X
σt2 = ω + 2
αi yt−i , ∀t∈Z
i=1

where ω > 0, α1 ≥ 0, . . . , αp ≥ 0 are strictly positive parameters.

In the same way as for the ARCH(1) model, it can be shown that the conditional distribution of the yt |Y t−1
generated by an ARCH(q) model is normal with mean zero and variance σt2 .

Lemma 2.1. The conditional distribution of yt given the past Y t−1 is normal with mean E(yt |Y t−1 ) = 0
and variance Var(yt |Y t−1 ) = σt2 , namely yt |Y t−1 ∼ N (0, σt2 ).

20
We have already seen that the ARCH(1) model can be re-written as an AR(1) model for the squared returns
yt2 . As we shall now see, a similar representation holds for the ARCH(q) model. In particular, the squared
returns of an ARCH(q) model follow an AR(q) process.
Theorem 2.7. Let {yt }t∈Z be generated by an ARCH(q) model. Then {yt2 }t∈Z follows an AR(q) model
q
X
yt2 =ω+ 2
αi yt−i + ηt
i=1

where {ηt }t∈Z is a white noise sequence.

Proof. The proof of this theorem is left as an exercise. 

The AR(q) representation for the squared returns of an ARCH(q) model, tells us that the ARCH(q) is capable
of generating arbitrary structures for the first q lags of the autocorrelation function. The exponential decay
of the ACF, with seasonal patterns, occurs only after q lags. For instance, for the S&P 500 series we could
choose q = 4.
The AR(q) representation for the squared returns of an ARCH(q) also allows us to easily establish
conditions for the weak stationarity of the sequence {yt2 }t∈Z . In particular, as you may recall from your
introductory time series courses,Pa necessary and sufficient condition for the weak stationarity of the {yt2 }t∈Z
q
generated by an AR(q) is that i=1 αi < 1.
Pq
Corollary 2.2. Let {yt }t∈Z be generated by an ARCH(q) model satisfying i=1 αi < 1. Then {yt2 }t∈Z is
weakly stationary.
Pq
Theorem 2.8. Let {yt }t∈Z be generated by the ARCH(q) model satisfying Pq i=1 αi < 1. Then, {yt }t∈Z is a
weakly stationary white noise sequence with E(yt ) = 0, Var(yt ) = ω/(1 − i=1 αi ) and Cov(yt , yt−l ) = 0 for
any l 6= 0.
Proof. The derivation of the mean and autocovariances follows the exact same argument as for the ARCH(1)
model. The derivations are identical because they involve only the observation equation (which is the same
for both models), and the fact that, conditional on Y t−1 , the conditional variance σt2 is given; which holds
for both ARCH(1) and ARCH(q) models.
The derivation of the unconditional variance uses again the AR representation of the ARCH model. First,
we make use of the observation equation to conclude that

Var(yt ) = E(yt2 ).

Next, we use the AR(q) representation of Theorem 2.7 to conclude that

q
X
E(yt2 ) = ω/(1 − αi ).
i=1

21
22
Chapter 3

Generalized ARCH Models

3.1 The GARCH(1,1) model

The specification of ARCH models with several lags is useful for describing the strong dynamics of squared
returns. However, in practice, this is not the most parsimonious model specification available. In particular,
when squared returns exhibit very strong dependence (or very long memory), then the use of an ARCH(q)
model would require a large order q and therefore many parameters to be estimated. One way of substantially
increasing the temporal dependence in squared returns requiring only a few additional parameters is to use
lags of σt2 in the updating equation. For example, when one lag of σt2 and one lag of yt2 is used in the
updating equation, we obtain the so-called generalized autoregressive conditional heteroeskedasticity model
of order (1,1), or GARCH(1,1).
Definition 3.1. GARCH(1,1) Model: The GARCH(1,1) model is given by

yt = σt t ,
σt2 = ω + β1 σt−1
2 2
+ α1 yt−1 , ∀t∈Z

where ω > 0, α1 ≥ 0, β1 ≥ 0 are parameters, {εt }t∈Z is an N ID(0, 1) sequence and εt is independent of the
past Y t−1 .
In the same way as discussed for the ARCH(1) model, it can be shown that yt |Y t−1 has a normal distribution
with mean zero and variance σt2 . The proof of Lemma 3.1 below is equal to that of Theorem 2.1.
Lemma 3.1. Let {yt }t∈Z be generated by a GARCH(1,1) model. The conditional distribution of yt given the
past Y t−1 is normal with mean E(yt |Y t−1 ) = 0 and variance Var(yt |Y t−1 ) = σt2 , namely yt |Y t−1 ∼ N (0, σt2 ).
Figure 3.1 shows that GARCH models are capable of modeling extreme changes in the conditional
volatility of returns, giving rise to strong clusters of volatility.
It is easy to show that the returns {yt }t∈Z generated by the GARCH(1,1) model are uncorrelated and have
mean zero, just as we did for the ARCH(1) and ARCH(q) models. In particular, the proofs are identical
because they involve only the observation equation (which is the same for both models), and the fact that,
conditional on Y t−1 , the conditional variance σt2 is given (which holds for both ARCH and GARCH models).
Lemma 3.2. The returns {yt }t∈Z generated by an GARCH(1,1) model have zero autocovariance at any lag
and are thus uncorrelated over time.
Lemma 3.3. The returns {yt }t∈Z generated by an GARCH(1,1) model have unconditional mean zero.
Once again, we can provide an ARMA representation for the squared returns of this model. In particular, the
squared observations generated by a GARCH(1,1) model can be shown to admit an autoregressive moving
average model of order (1,1), or ARMA(1,1), representation.

23
 simulated y t
5

-5

0 100 200 300 400 500 600 700 800 900 1000

simulated σ t
4

0
0 100 200 300 400 500 600 700 800 900 1000

Figure 3.1: Sample path of returns {yt }Tt=1 and volatility {σt }Tt=1 generated from a GARCH(1,1) model with
(ω, β1 , α1 ) = (0.1, 0.75, 0.2).

Lemma 3.4. Let {yt }t∈Z be generated by a GARCH(1,1) model. Then {yt2 }t∈Z admits an ARMA(1,1)
representation
yt2 = ω + (α1 + β1 )yt−1
2
+ ηt − β1 ηt−1
where {ηt }t∈Z is a white noise process.

Proof. Define the white noise sequence ηt = yt2 − σt2 . Then, plugging in σt2 = yt2 − ηt and σt−1
2 2
= yt−1 − ηt−1
in the updating equation for the GARCH(1,1), we obtain

yt2 = ω + (α1 + β1 )yt−1

2
+ ηt − β1 ηt−1 ,

which is indeed an ARMA(1,1) process. 

The ARMA(1,1) representation of the squared returns generated by a GARCH(1,1) model can be useful to
obtain the unconditional variance of yt . Theorem 3.1 tells us that the returns yt generated by a GARCH(1,1)
model with α1 + β1 < 1 have a time-invariant unconditional variance given by ω/(1 − β1 − α1 ).

Theorem 3.1. The returns {yt }t∈Z generated by an GARCH(1,1) model with α1 + β1 < 1 have a time-
invariant unconditional variance given by Var(yt ) = ω/(1 − β1 − α1 ).

Proof. We know that

Var(yt ) = E(yt2 ),

24
 therefore from the ARMA(1,1) representation of yt2 in Lemma 3.4 we immediately obtain

E(yt2 ) = ω/(1 − β1 − α1 ).

For a GARCH model with ω = 0.1, α1 = 0.2 and β1 = 0.5, Figure 3.2 shows that the variance of yt
increases dramatically as β1 increases from 0.5 to 0.75, and 1 − β1 − α1 approaches zero.

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

Figure 3.2: Unconditional sample density of yt generated by a GARCH(1,1) model with parameters
(ω, α1 , β1 ) = (0.1, 0.2, 0.5) [left figure] and (ω, α1 , β1 ) = (0.1, 0.2, 0.75) [right figure].

To better appreciate why the GARCH(1,1) model is able to capture high persistence in the variance, we
can rewrite the GARCH(1,1) model as an ARCH(∞) model with some constraints on the parameters. In
particular, unfolding the GARCH(1,1) updating equation, the conditional variance σt2 can be expressed as
∞
ω X
σt2 = + α1 β1i yt−1−i
2
.
(1 − β1 ) i=0

Theorem 3.2. Let {yt }t∈Z be a time-series generated by the GARCH(1,1) model and let the following
additional parameter restriction hold α1 +β1 < 1. Then, {yt }t∈Z is a weakly stationary White Noise sequence
with E(yt ) = 0, Var(yt ) = ω/(1 − β1 − α1 ) and Cov(yt , yt−l ) = 0 for any l 6= 0.

Figure 3.3 shows plots the ACF of yt2 and shows that the temporal dependence in squared returns is
greatly affected by the value of β1 .

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4
1 3 5 7 9 11 13 15 1 3 5 7 9 11 13 15

Figure 3.3: Sample ACF of squared returns yt2 generated by a GARCH(1,1) model with parameters
(ω, α1 , β1 ) = (0.1, 0.2, 0.5) [left figure] and (ω, α1 , β1 ) = (0.1, 0.2, 0.78) [right figure].

25
3.2 The GARCH(p,q) model
In order to describe additional temporal dynamics, the GARCH(1,1) model can be further extended to a
GARCH(p,q) with general orders p and q.
Definition 3.2. GARCH(p,q) Model: The GARCH(p,q) model is given by
p
X q
X
y t = σ t t , σt2 =ω+ 2
βi σt−i + 2
αi yt−i , ∀ t ∈ Z, (3.1)
i=1 i=1

where ω > 0, α1 ≥ 0, β1 ≥ 0 are parameters, {εt }t∈Z is an N ID(0, 1) sequence and εt is independent of the
past Y t−1 .
As for all previous models, it can be shown that yt |Y t−1 has a normal distribution with mean zero and
variance σt2 .
Lemma 3.5. Let {yt }t∈Z be generated by a GARCH(p,q) model. The conditional distribution of yt given the
past Y t−1 is normal with mean E(yt |Y t−1 ) = 0 and variance Var(yt |Y t−1 ) = σt2 , namely yt |Y t−1 ∼ N (0, σt2 ).
The proof that the unconditional mean and autocovariances of the returns {yt }t∈Z generated by the GARCH(p,q)
process are zero, is once again identical to that of the simple ARCH(1) model.
Lemma 3.6. The returns {yt }t∈Z generated by an GARCH(p,q) model have zero autocovariance at any lag,
and are thus uncorrelated over time.
Lemma 3.7. The returns {yt }t∈Z generated by an GARCH(p,q) model have unconditional mean zero.
An ARMA(q ∗ ,p) representation is also available for the squared returns of the GARCH(p,q) model, where
q ∗ = max{q, p}. Notice that the number p of MA lags in the ARMA representation of the GARCH(p, q)
is determined exclusively by the number of yt2 in the GARCH updating equation. In contrast, the number
of AR lags in the ARMA representation depends on the number of lags of both yt2 and σt2 featured in the
GARCH updating equation.
Lemma 3.8. Let {yt }t∈Z be generated by a GARCH(p,q) model. Then {yt2 }t∈Z admits an ARMA(max{q, p},p)
representation
Xq p
X p
X
yt2 = ω + 2
αi yt−i + 2
βi yt−i + ηt − βi ηt−i
i=1 i=1 i=1

where {ηt }t∈Z is a white noise process.

2
Proof. Define the white noise sequence ηt = yt2 − σt2 . Then, plugging in σt2 = yt2 − ηt and σt−i 2
= yt−i − ηt−i
in the updating equation for the GARCH(p,q), we obtain
q
X p
X p
X
yt2 =ω+ 2
αi yt−i + 2
βi yt−i + ηt − βi ηt−i ,
i=1 i=1 i=1

which is indeed an ARMA(max{q, p},p) process. 

Once again, the ARMA representation of the squared returns generated is useful to obtain the unconditional
variance of yt
Pq Pp
Theorem 3.3. The returns {yt }t∈Z generated by an GARCH(p,q), P model withP i=1 αi + i=1 βj < 1 have
q p
a time-invariant unconditional variance given by Var(yt ) = ω/(1 − i=1 αi − i=1 βj ).
We also obtain that a sequence {yt }t∈Z generated by a GARCH(p,q) is white noise
Pq Pp
Corollary 3.1. Let {yt }t∈Z be generated by a GARCH(p,q) model satisfying i=1 αi + i=1 βj < 1. Then
{yt }t∈Z is weakly stationary white noise sequence.

26
3.3 Simulating GARCH models with R
As we shall now see, it is very easy to simulate data from a GARCH(1,1) model using R. The code discussed
below can be found in the R file Simulate GARCH.R. First, we define the sample size n, and the parameter
values for ω, α1 and β1 .

n <- 1000
omega <- 0.1
alpha <- 0.2
beta <- 0.75

Next, we generate n Gaussian N(0, 1) random innovations, labeled epsilon, by calling the function
rnorm.

epsilon <- rnorm(n)

Then, we define 2 vectors of length n full of zeros to contain our simulated conditional variance {σt2 }nt=1
and simulated observations {yt }nt=1 .

sig2 <- rep(0,n)

x <- rep(0,n)

Finally, we simulate data from our GARCH(1,1) model using a for loop that runs from t = 2 to t = n.
Notice that before starting the for loop we must first set the initial value σ12 . A reasonable option is to set
it to the unconditional variance of yt , which is given by ω/(1 − α1 − β1 ).

sig2[1] <- omega/(1-alpha-beta)

x[1] <- sqrt(sig2[1]) * epsilon[1]

for(t in 2:n){

sig2[t] <- omega + alpha * x[t-1]^2 + beta * sig2[t-1]

x[t] <- sqrt(sig2[t]) * epsilon[t]

Stacking the entire code together we obtain the following R script for simulating a time series of length
n = 1000 from a GARCH(1,1) model.

n <- 1000
omega <- 0.1
alpha <- 0.2
beta <- 0.75

epsilon <- rnorm(n)

sig2 <- rep(0,n)

x <- rep(0,n)

sig2[1] <- omega/(1-alpha-beta)

x[1] <- sqrt(sig2[1]) * epsilon[1]

27
for(t in 2:n){
sig2[t] <- omega + alpha * x[t-1]^2 + beta * sig2[t-1]
x[t] <- sqrt(sig2[t]) * epsilon[t]
}

28
Chapter 4

Parameter Estimation

In practice, given an observed sequence of T stock returns {y1 , y2 , ..., yT }, generated by a GARCH(p, q) model
with θ0 = (ω, α1 , ..., αq , β1 , ..., βp ), we do not know what are the parameter values that correctly describe
the dynamics of the time-varying conditional volatility. The problem we face is that of trying to find the
values of the true parameter vector θ0 of the GARCH(p, q) model from which the data {y1 , y2 , ..., yT } was
generated. We thus need to find a way of estimating and conducting inference on the parameter vector θ0 .
A standard method used for the estimation of GARCH parameters is the method of Maximum Likelihood.
In this section, we shall see how to write down the Likelihood Function for any given GARCH(p, q) model,
and how to obtain the Maximum Likelihood Estimator (MLE) θ̂T of the unknown parameter vector θ0 .

4.1 Deriving the likelihood function

As you already know from your introductory probability courses, the joint density function f (x, y) of two
random variables x and y, can always be factorized into the product of a a conditional density f (x|y) and a
marginal density f (y),
f (x, y) = f (x|y) × f (y).
In your introductory time series courses you have also noted that this factorization can be very useful in
the context of maximum likelihood! In particular, the joint density function of the data {y1 , ..., yT } can be
factorized as the product of several conditional densities f (yt |yt−1 ) and a marginal density f (y1 ).
For concreteness, let p(y1 , . . . , yt ; θ) denote the joint probability density function of the vector of random
returns {y1 , . . . , yT } generated by a GARCH model. Note that the joint density of the data depends on
the vector of parameters θ = (ω, α1 , ..., αq , β1 , ..., βp ) since, as we verified in the previous chapter, these
parameters determine the distributional properties of the data. Recall from your introductory statistics
courses that, the likelihood function is exactly the same as the joint density function p(y1 , . . . , yt ; θ). The only
difference is that that the likelihood function takes the data {y1 , ..., yT } as given and analyses p(y1 , . . . , yt ; θ)
as being a function of the parameter vector θ, whereas the joint density function takes the parameter vector
θ as given (it is fixed) and analyses p(y1 , . . . , yt ; θ) as a function of the data {y1 , ..., yT }.
In any case, you can naturally factorize the joint density function as follows

p(y1 , . . . , yt ; θ) = p(y2 , . . . , yt |y1 ; θ) × p(y1 ; θ).

Furthermore, we can also factorize the joint density p(y2 , . . . , yt |y1 ; θ) as

p(y2 , . . . , yt |y1 ; θ) = p(y3 , . . . , yt |y1 ; θ) × p(y2 |y1 ; θ)

which implies that

p(y1 , . . . , yt ; θ) = p(y3 , . . . , yt |y2 , y1 ; θ) × p(y2 |y1 ; θ) × p(y1 ; θ).

29
Repeating this procedure, we obtain the desired factorization of the joint density function
p(y1 , . . . , yT ; θ) = p(y1 ; θ) × p(y2 |y1 ; θ)p(y3 |y2 , y1 ; θ) × · · · × p(yT |yT −1 , . . . , y1 ; θ)
T
Y
= p(y1 ; θ) p(yt |yt−1 , . . . , y1 ; θ), (4.1)
t=2

where p(y1 ; θ) denotes the marginal density of y1 and p(yt |yt−1 , . . . , y1 ; θ) denotes the conditional density of
yt given all the previous elements {yt−1 , . . . , y1 }.
Naturally, you may ask: why is this factorization useful? Well, the answer is that while the joint density
of the data p(y1 , . . . , yT ; θ) is very complicated and potentially even intractable, each of the conditional
densities p(yt |yt−1 , . . . , y1 ; θ) is perfectly simple. Indeed, as we have seen in Chapters 2 and 3, for any
GARCH model, the distribution of yt conditional on its past {yt−1 , yt−2 , ...}, is Gaussian with mean zero
and variance σt2 ,
yt |yt−1 , yt−2 , ... ∼ N (0, σt2 ).
The reason for this is that, conditional on the past data {yt−1 , yt−2 , ...}, the conditional variance σt2 is given
(i.e. it is known!). The probability density function of a normal random variable yt with mean zero and
variance σt2 is simply given by
yt2
 
1
p(yt |yt−1 , yt−2 , ...; θ) = p exp − 2 .
2πσt2 2σt
As you may also recall, we often work with the logarithm of the likelihood function, the so called log-
likelihood function. In the following, we discuss how to obtain the log-likelihood function for ARCH and
GARCH models.

Log-likelihood function of the ARCH(1) model

In the case {yt }Tt=1 is an observed sample of size T from an ARCH(1) model, we have that p(yt |yt−1 , . . . , y1 ; θ) =
p(yt |yt−1 ; θ) because yt depends only on the previous observation yt−1 . Furthermore, as discussed in the
previous section p(yt |yt−1 ; θ), is a Normal density with mean zero and variance σt2 = ω + α1 yt−1 2
. More
specifically,
yt2
 
1
p(yt |yt−1 ; θ) = q exp − 2 ) .
2π(ω + α y 2 ) 2(ω + α1 yt−1
1 t−1

As a result, for the ARCH(1) model, we can write the log-likelihood function as
T
X
log (p(y1 , . . . , yT ; θ)) = log (p(y1 ; θ)) + log (p(yt |yt−1 ; θ))
t=2
T 
yt2

1X 2


= log (p(y1 ; θ)) − log(2π) + log ω + α1 yt−1 + 2 . (4.2)
2 t=2 ω + α1 yt−1

In practice, the marginal density p(y1 ; θ) has an unknown functional form and therefore it is common practice
to use a conditional log-likelihood function where y1 is treated as given. Therefore the term log (p(y1 ; θ)) in
(4.2) can be simply dropped and we consider the following log-likelihood function
T
X
L(y1 , ..., yT , θ) = lt (θ), (4.3)
t=2

where each term lt (θ), which denotes the contribution to the likelihood of observation t, is given by
yt2
 
1 2


lt (θ) = − log ω + α1 yt−1 + 2 .
2 ω + α1 yt−1

30
Note that in the log-likelihood in (4.3) we also dropped the term log(2π) as it is an additive constant and
therefore irrelevant. Additive constants can be always dropped when writing log-likelihood functions because
the vector θ that maximizes L(y1 , ..., yT , θ) is the same that maximizes L(y1 , ..., yT , θ) + c, where c is any
given constant not depending on θ. Furthermore, derivatives of L(y1 , ..., yT , θ) are the same as derivatives of
L(y1 , ..., yT , θ) + c. This can be also noted from Figure 4.1. The first plot shows the log-likelihood function
in (4.3) for different values of α1 using Apple stock returns. The second plot instead shows the log-likelihood
with the additional term log(2π) as in (4.2). As you can see the likelihood function is exactly the same, it
is only shifted along the vertical axis. Therefore, maximizing this two functions leads to the same result.
Finally, the third plot of Figure 4.1 shows the likelihood function instead of the log-likelihood. This shows
that both function, though different, are always maximized at the same point. This is the reason why we
can use the log-likelihood instead of the likelihood.

log-likelihood

-515

-520
0.1 0.2 0.3 0.4 0.5 0.6 0.7
log-likelihood - constant
-745

-750

0.1 0.2 0.3 0.4 0.5 0.6 0.7

×10 -223 likelihood
1.5

0.5

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7

Figure 4.1: Likelihood functions for Apple daily log-returns in 2015 for different values of α1 with fixed
ω = 2. The red lines denote the values α1 that maximize the functions.

Likelihood function of the ARCH(q) model

Using a similar argument as discussed for the ARCH(1) model, it can be shown that the conditional log-
likelihood function of an ARCH(q) model is given by

T
X
L(y1 , ..., yT , θ) = lt (θ),
t=q

31
where lt (θ) is given by

yt2
 
1 2 2


lt (θ) = − log ω + α1 yt−1 + · · · + αq yt−q + 2 2 .
2 ω + α1 yt−1 + · · · + αq yt−q

Note that in this case the first q observations, y1 , . . . , yq , are treated as given constants and the sum starts
from q + 1. This because we need q past observations, {yt−q , . . . , yt−1 }, to obtain the conditional variance
σt2 and the first observation is available at time t = 1.

Likelihood function of the GARCH(1,1) model

When the sample of data is generated by a GARCH(1,1) model, then we can simply use the general formula
for the conditional density
y2
 
1
p(yt |yt−1 ; θ) = p exp − t2 .
2πσt2 2σt
where σt2 is naturally obtained as
σt2 = ω + α1 yt−1
2 2
+ β1 σt−1 .
As a result, for the GARCH(1,1) model, we can write the log-likelihood function as
T
X
log (p(y1 , . . . , yT ; θ)) = log (p(y1 ; θ)) + log (p(yt |yt−1 ; θ))
t=2
T 
y2

1X
= log (p(y1 ; θ)) − log(2π) + log σt2 + t2 . (4.4)
2 t=2 σt

As argued before, the constant terms and marginal density of y1 are typically ignored, giving rise to a
simplified log-likelihood function of the form
T T
yt2
 
X X 1
L(y1 , ..., yT , θ) = lt (θ) = − log σt2 + . (4.5)
t=2 t=2
2 σt2

Note that since we are using the recursion

σt2 = ω + α1 yt−1
2 2
+ β1 σt−1 .

at time t = 2 we have that σ22 = ω + α1 y12 + β1 σ12 . Indeed y1 is observed but what about σ12 ? In practice σ12
is fixed to be equal to a given constant (often the sample variance of the observations is used).

Likelihood function of the GARCH(p, q) model

For the GARCH(p, q) model the expression of the log-likelihood is the same as the one for the GARCH(1,1)
model given in (4.5). The only difference lies in the way the σt2 is obtained through the updating equation
GARCH(p, q) model.

4.2 Maximum Likelihood Estimator and Asymptotic properties

Once we have obtained the conditional log-likelihood function L(y1 , ..., yT , θ), the maximum likelihood es-
timator (MLE) θ̂T obtained from the sample of data {y1 , ..., yT }, is simply defined as the argument that
maximizes the log-likelihood function L(y1 , ..., yT , θ) over the parameter space Θ containing all possible
values of the parameters. In particular,

θ̂T = arg max L(y1 , ..., yT , θ).

θ∈Θ

32
 It is important to note that the log-likelihood function L(y1 , ..., yT , θ) is a random variable. Indeed,
for every new realization of the random sample {y1 , ..., yT }, we obtain a new log-likelihood function to be
maximized over θ. As such, the maximum likelihood estimator θ̂T is also a random variable. In particular,
for every new sample of data {y1 , ..., yT } there is a new value θ̂T that maximizes the log-likelihood. We call
this value the point estimate of the true parameter θ0 .
While the MLE θ̂T is a continuous random variable, and hence, it will be almost surely different from θ0 ,
it does have important properties. In particular, under standard regularity conditions, it can be shown that
θ̂T is consistent and asymptotically normal for θ0 .
Recall from introduction to econometrics that an estimator θ̂T is said to be a consistent estimator of the
true unknown θ0 if θ̂T converges in probability to θ0 as the sample √ size T diverges to infinity. Furthermore,
θ̂T is said to be asymptotically normal for θ0 if the random variable T (θ̂T − θ0 ) converges in distribution to
a multivariate normal random variable N (0, Ω), where 0 is a vector of zeros and Ω is a variance covariance
matrix called the asymptotic variance of θ̂T .
Lemma 4.1. Under appropriate regularity conditions θ̂T converges in probability to θ0 as the sample size
diverges
p
θ̂T → θ0 as T → ∞.
Lemma 4.2. Under appropriate regularity conditions θ̂T is asymptotically normal for θ0 ,
√ d
T (θ̂T − θ0 ) → N (0, Ω) as T →∞

where Ω = I(θ0 )−1 is the inverse Fisher Information matrix

! !
∂ 2 lt (θ) 1 1 ∂σt2 ∂σt2
I(θ0 ) = −E = E .
∂θ∂θ> 2 σt4 ∂θ ∂θ>

At this point, it is important to note that:

(a) the inverse Fisher Information matrix I(θ0 ) must be inverted for us to obtain the asymptotic variance
Ω of the MLE;
∂σt2
(b) the Fisher Information matrix depends on the derivative process ∂θ .

In practice, point (a) means that we must be careful in scaling the log-likelihood function since any
computer program may have trouble in finding the inverse of a matrix that is not properly scaled.
∂σ 2
Point (b) is important since it shows that the random sequence ∂θt plays a role in obtaining the asymp-
∂σ 2
totic variance of the maximum likelihood estimator. Note that { ∂θt } is a vector time-series determined by
its own updating equation. Take the GARCH(1,1) model as an example, then we have

∂σt2 h
∂σt2 ∂σt2 ∂σt2
i
= ∂ω ∂α ∂β
∂θ
where the elements of this vector are obtained through the following updating equations
2 2
∂σt2 ∂ω ∂αyt−1 ∂βσt−1
= + +
∂ω ∂ω ∂ω ∂ω
2
∂σt−1
= 1 + 0 + β ,
∂ω
2 2
∂σt2 ∂ω ∂αyt−1 ∂βσt−1
= + +
∂α ∂α ∂α ∂α
∂σ 2
2
= 0 + yt−1 + β t−1 ,
∂α

33
 2 2
∂σt2 ∂ω ∂αyt−1 ∂βσt−1
= + +
∂β ∂β ∂β ∂β
2
2 ∂σt−1
= 0 + 0 + σt−1 + β .
∂α
Taking all equations together we obtain the following lemma.
Lemma 4.3. The conditional volatility derivative process {∂σt2 /∂θ} of the GARCH(1,1) model satisfies the
following updating equation  
1 ∂σ 2
∂σt2 2  + β t−1 .
=  yt−1
∂θ 2 ∂θ
σt−1

4.3 Statistical Inference

The asymptotic normality result stated in Lemma 4.2 above shows that, for large T , the maximum likelihood
estimator θ̂T is a random variable that is approximately Gaussian, centered at the unknown θ0 , and with a
variance that vanishes to zero as T → ∞. Indeed, the asymptotic normality of the MLE tells us that
√   app  
T θ̂T − θ0 ∼ N 0 , I(θ0 )−1
app
where ∼ denotes an ‘approximate’ distribution. This means naturally, that
app
 1 
θ̂T − θ0 ∼ N 0 , I(θ0 )−1
T
and hence that
app
 1 
θ̂T ∼ N θ0 , I(θ0 )−1 .
T
Notice how the variance T1 I(θ0 )−1 os the MLE vanishes to zero as T → ∞. This means that the distribution
of the maximum likelihood estimator collapses to the true parameter θ0 and becomes becomes increasingly
accurate as T → ∞.
Figure 4.2 below shows the density of the MLE for the parameters ω, α1 and β1 in the context of a
GARCH(1,1) model. These densities were obtained through a Monte Carlo exercise. In particular, we first
simulate data {y1 , ..., yT } from a GARCH(1,1) model with parameter vector θ0 = (ω, α1 , β1 ) = (0.1, 0.2, 0.75),
and then use the simulated data to obtain a point estimate θ̂T pretending that we do not know θ0 . By re-
peating this procedure many times, we obtain a new point estimate θ̂T for each new time series that we
simulate from the GARCH(1,1) model. Figure 4.2 shows the density of the MLE that we obtained by sim-
ulating N = 1000 time series with sample size T = 500, T = 1000 and T = 5000. This figure shows that
the distribution of the MLE is indeed approximately normal in large samples, and furthermore, that it is
collapsing to the true parameter θ0 .
√
Figure 4.3 uses√the same Monte Carlo procedure to obtain the density of the quantity T (θ̂T − θ0 ). The
figure shows that T (θ̂T − θ0 ) is indeed approximately normally distributed.

In practice, the Fisher information matrix I(θ0 ) = −E(∂ 2 lt (θ)/∂θ∂θ> ) is unknown since it depends on the
true unknown parameter θ0 , and also, because the expectation E is unknown. We can however approximate
I(θ0 ) by its plug-in estimator
! T
∂ 2 lt (θ) 1 X ∂ 2 lt (θ̂T )
I(θ0 ) = −E ≈ − .
∂θ∂θ> T t=1 ∂θ∂θ>
PT
Notice that, in the expression above, we have replaced the expectation E with the sample average 1/T t=1
and we have substituted the unknown true parameter θ0 by the sample estimate θ̂T .

34
 30

25
T=500

25
T=1000
T=5000
25

true

20

20
20

15

15
ω
15

α
β
10

10
10

5
5
0

0
0.00 0.05 0.10 0.15 0.20 0.60 0.65 0.70 0.75 0.80 0.85 0.10 0.15 0.20 0.25 0.30

Figure 4.2: Distribution of the ML estimator θ̂T = (ω̂T , α̂T , β̂T ) for different sample sizes T . The red line
denotes the true value θ0 .
0.35

0.4
0.4

T=500
T=1000
0.30

T=5000
Normal

0.3
0.25
0.3

0.20

0.2
ω

α
β
0.2

0.15
0.10

0.1
0.1

0.05
0.00
0.0

0.0

−4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4

√
Figure 4.3: Distribution of T (θ̂T − θ0 ) the for different sample sizes T . The red line denotes the normal
density function.

We can now invert this estimate of the Fisher information matrix to obtain an estimate of the asymptotic
variance-covariance matrix of the MLE θ̂T

T
!−1
1 X ∂ 2 lt (θ̂T )
Ω̂ = − .
T t=1 ∂θ∂θ>

With Ω̂ we are now in a position to report standard errors for our point estimates, construct confidence
intervals for θ0 and produce p-values. All of these are done in the same way as you have done in introductory
statistics courses.
In particular, the standard error of the ith element of the vector θ, can be obtained by taking the square
root of the ith diagonal element of Ω̂ii ,
r
i 1
SE(θ̂T ) = Ω̂ii
T

where θ̂Ti denotes the ith element of the vector θ̂T and Ω̂ii the element in the ith row ad ith column of the
variance-covariance matrix Ω̂.
An approximate 95% confidence interval for the ith element of θ0 , labeled θ0i , can be obtained by taking

35
an interval around the point estimate θ̂Ti that spans ±1.96 standard errors
r r
h
i 1 i 1 i
θ̂T − 1.96 × Ω̂ii , θ̂T + 1.96 × Ω̂ii
T T

4.4 Numerical Optimization of the log-Likelihood Function

As you may recall, in introductory econometrics, we were always able to derive an expression for the maximum
likelihood estimator, by taking a derivative of the log-likelihood, setting it to zero, and solving for θ̂T . The
idea was, of course, that if θ̂T is the maximizer of the log-likelihood function L(y1 , ..., yT , θ), then it must
satisfy
T
X ∂lt (θ̂T )
= 0.
t=1
∂θ

In other words, the derivative of the likelihood must be zero at θ̂T . This is how we obtained the expression
for the MLE PT
yt xt
β̂ = Pt=1
T 2
t=1 xt
in the context of the simple linear Gaussian regression model

yt = βxt + t .

Similarly, you may recall that this was the way in which we obtained the expression for the MLE
PT
xt xt−1
ρ̂ = Pt=1
T 2
t=1 xt−1

in the context of the Gaussian AR(1) model

xt = ρxt−1 + t .

Unfortunately, in the context of a GARCH model it is not possible for us to obtain estimates of the
parameter vector in the same way. The problem is too complicated, and trying to set the derivative of the
log-likelihood to zero does not deliver a closed form expression for θ̂T .
Therefore, a naive way to proceed would be simply to evaluate the log-likelihood function for several
values of θ and pick the one that maximizes it. For instance, Table 4.1 reports the log-likelihood values of
a GARCH(1,1) model using Apple returns for different values of θ = (ω, α1 , β1 ). In this case we could pick
θ = (0.20, 0.07, 0.85) as it gives the maximum log-likelihood. This approach is naive and we can do much
better by using numerical Algorithms.

Table 4.1: Log-likelihood function of the GARCH(1,1) model for daily Apple log-returns (from 2010 to 2016)
evaluated at different values of θ.
parameter value θ = (ω, α1 , β1 ) log-lik value
(0.30, 0.10, 0.70) -2962.5
(0.20, 0.10, 0.70) -3090.0
(0.20, 0.07, 0.70) -3211.4
(0.20, 0.07, 0.80) -2941.6
(0.20, 0.07, 0.85) -2882.3

One simple algorithm to maximize the log-likelihood L(y1 , ..., yT , θ) is the Newton-Raphson algorithm.

36
Remark 4.1. Starting from an initial value θ(1) , the Newton-Raphson algorithm finds the maximum of the
log-likelihood function LT (θ) = L(y1 , ..., yT , θ) by updating the parameter as follows
 −1
θ(k+1) = θ(k) − ∇LT (θ(k) ) ∇2 LT (θk )

where ∇LT (θ(k) ) and ∇2 LT (θ(k) ) denote the Jacobian and Hessian matrix respectively
∂LT (θ(k) ) ∂ 2 LT (θ(k) )
∇LT (θ(k) ) = and ∇2 LT (θ(k) ) = .
∂θ ∂θ∂θ>
Lemma 4.4. Under appropriate regularity conditions, the Newthon-Raphson algorithm converges to the
MLE as the number of iterations k go to infinity; i.e. θ(k) → θ̂T as k → ∞.

4.5 Estimating GARCH models with R

We will now make use of R to evaluate and maximize the log-likelihood function of a GARCH(1,1) model.

Write the likelihood function

We first create an R function to evaluate the log-likelihood function. We call this function llik fun GARCH.
The function takes as input a vector of data labeled x and a parameter vector labeled par, and returns as
output the log-likelihood. You can find this function as an R file labeled llik fun GARCH.R.
The inputs and outputs of the functions are defined by starting the script with the following line
llik_fun_GARCH <- function(par,x){
First, we define n to be the number of observations in the vector of data x. Then, we define the parameter values
omega, alpha and beta from the input vector par. We consider some link functions to impose some restriction on
omega, alpha and beta, which are useful to avoid numerical problems. In particular, we consider an exponential link
function exp() to ensure ω > 0. In this way, par[1] will be allowed to take any value in the optimization (even
negative) but still ensuring ω > 0. Similarly, we impose 0 < α < 1 and 0 < β < 1 by using the logistic link function
logistic() = exp()/(1 + exp()).
n <- length(x)
omega <- exp(par[1])
alpha <- exp(par[2])/(1+exp(par[2]))
beta <- exp(par[3])/(1+exp(par[3]))
We then use the data x and the parameters to filter the conditional variance, which we label sig2 in the code.
In particular, we use a for loop to obtain the sequence {σt2 }n 2
t=1 recursively from t = 2 to t = n. We set σ1 equal to
the sample variance of the data var(x) to initialize the updating equation.
sig2 <- rep(0,n)
sig2[1] <- var(x)

for(t in 2:n){

sig2[t] <- omega + alpha*x[t-1]^2 + beta*sig2[t-1]

}
Finally, we use the filtered sequence {σt2 }n
t=1 to calculate the contribution to the log-likelihood function of each
observation yt . Below l is a vector that contains the log-likelihood values from t = 1 to n. Finally, we calculate the
average log-likelihood value llik as the average of all contributions and give it as output using return().

l <- -(1/2)*log(2*pi) - (1/2)*log(sig2) - (1/2)*x^2/sig2

llik <- mean(l)

return(llik)

37
 Stacking all the code together, we obtain the following script to evaluate the log-likelihood function of the GARCH
model.

llik_fun_GARCH <- function(par,x){

n <- length(x)
omega <- exp(par[1])
alpha <- exp(par[2])/(1+exp(par[2]))
beta <- exp(par[3])/(1+exp(par[3]))

sig2 <- rep(0,n)

sig2[1] <- var(x)

for(t in 2:n){
sig2[t] <- omega + alpha*x[t-1]^2 + beta*sig2[t-1]
}

l <- -(1/2)*log(2*pi) - (1/2)*log(sig2) - (1/2)*x^2/sig2

llik <- mean(l)

return(llik)
}

Optimizing the likelihood function

We are now in a position to optimize the log-likelihood function and obtain the maximum likelihood estimator
θ̂T . The code for estimating the parameters of the ARCH model through the MLE is available on the R file
Estimate ML GARCH.R . First, we load the likelihood function using source(). Note that the file “llik_fun_GARCH.R”
has to saved on the working directory. You can visualize the current working directory using the command getwd().
You can also change the working directory using setwd("C:/Users/R code"), where C:/Users/R code will be re-
placed with the path of the folder that you want to use as working directory.

source("llik_fun_GARCH.R")

Then, we load the series of observed data of interest to us. In this case, our data is saved in the file stock returns.txt,
and hence, we can load the data by writing

x <- scan("stock_returns.txt")

Next, we define the initial value for ω, α1 and β1 . These are the values that the numerical algorithm will use
to start the iteration towards the maximum of the log-likelihood. However, since the input par of the function
llik_fun_GARCH.R is transformed using link functions (see above), we set the initialization par_ini by using the
inverse of the link functions. The inverse of exp(p) is log(p) and the inverse of logistic(p) = exp(p)/(1 + exp(p)) is
log(p/(1 − p)).

a <- 0.2
b <- 0.6
omega <- var(x)*(1-a-b)

par_ini <- c(log(omega),log(a/(1-a)),log(b/(1-b)))

Finally, we obtain the point estimate θ̂T by optimizing the log-likelihood function using the R function optim().
Since the optim() function tries to find the minimum of any function, we must give it the negative of the log-likelihood
function

optim(fn=function(par) - llik_fun_GARCH(par,x))

In order for optim() to work properly, we also need to provide the initial values of the parameter and select the
numerical algorithm. We give it the starting value for the iterations par ini and select the algorithm BFGS. We run
the optimizer and store the results in est.

38
est <- optim(par=par_ini,fn=function(par)-llik_fun_GARCH(par,x), method = "BFGS")

Finally, we note that the output of the optim() function includes:

1. The point estimates θ̂T , which will be contained in est$par;
2. The value of the negative average log-likelihood evaluated at θ̂T , which will be contained in est$value;
3. An exit flag which will confirm if convergence was obtained or, in contrast, if the algorithm found an obstacle
and the optimization was aborted. When the variable est$convergence is zero, then the optimization has been
successful.

Standard errors and confidence intervals

The covariance matrix estimate Ω̂ can be obtained from the hessian matrix of the average negative log-likelihood
(second derivative with respect to the parameters). The function Hess_fun_GARCH() gives the average log-likelihood
but without transforming the parameter input with the link functions (see the R file Hess_fun_GARCH.R). This is
needed because we want the hessian with respect to the original parameters not through the link functions. We
obtain the hessian matrix using the R function optimHess() as follows:

hessian <- optimHess(par=theta_hat, fn=function(par)-Hess_fun_GARCH(par,x))

Next, we obtain the covariance matrix estimate Ω̂ by simply inverting the hessian matrix. We label the matrix
Ω̂ as SIGMA. This is done through the following code

SIGMA <- solve(hessian)

Finally, we can use the covariance matrix SIGMA to obtain a 95% confidence interval for β. Note that the estimate
of β is the third element of the vector theta hat and its corresponding variance is the element in position (3,3) of
the matrix SIGMA.

lb_beta <- theta_hat[3]-1.96*sqrt(SIGMA[3,3])/sqrt(length(x))

ub_beta <- theta_hat[3]+1.96*sqrt(SIGMA[3,3])/sqrt(length(x))

ci_beta <- c(lb_beta, ub_beta)

The resulting confidence interval is labeled ci beta. This code can be found in the file Estimate ML GARCH.R.

39
40
Chapter 5

Financial Analysis of ARCH and

GARCH Models

In this chapter, we turn to the econometric analysis of ARCH and GARCH models. In particular, we use estimated
ARCH and GARCH models to produce measures of risk that are useful for the economic and financial analysis of
data.

5.1 Estimation of the conditional volatility

Once the parameter vector θ0 has been estimated we usually are also interested in obtaining an estimate of sequence
of conditional variances {σt2 }Tt=1 . We denote this estimated sequence with {σ̂t2 }Tt=1 and we call it the filtered volatility.
How is {σ̂t2 }Tt=1 obtained? Well, simply through model updating equation and plugging in the estimated parameter
vector θ̂T . For example, for the GARCH(1,1) model our parameter estimate will be θ̂T = (ω̂, α̂1 , β̂1 ). Therefore,
{σ̂t2 }Tt=1 is obtained as
σ̂t2 = ω̂ + β̂1 σ̂t−1
2 2
+ α̂1 yt−1 , for t = 2, . . . , T.
The value σ̂12 can be set equal to the sample variance of the observations.
Deriving the estimated conditional volatility can be useful to understand how the risk of a certain financial asset
(or market) evolved over time.

Conditional volatility with R

Using R, we can obtain the estimated conditional variance of a GARCH(1,1) as follows. The R code for this chapter
is in the file analysis GARCH.R. First note that the maximum likelihood estimate of θ0 , as obtained in the previous
chapter, are given by omega_hat, alpha_hat and beta_hat. The time series is labeled x. We define a vector sigma2
that will contain our estimated volatility. We also initialize the conditional volatility using the sample variance.

n <- length(x)
sigma2 <- rep(0,n)
sigma2[1] <- var(x)

Now we are ready to obtain recursively the estimated conditional volatility using the GARCH(1,1) updating
equation and using theta hat as parameter values.

for(t in 2:n){

sigma2[t] <- omega_hat + alpha_hat*x[t-1]^2 + beta_hat*sigma2[t-1]

The estimated conditional volatility, together with the data series stock returns.txt, is plotted in Figure 5.1.

41
 Time series
10

-5

-10
0 100 200 300 400 500 600 700 800 900 1000

Estimated conditional variance

15

10

0
0 100 200 300 400 500 600 700 800 900 1000

Figure 5.1: Time series (first plot) and estimated filtered variance (second plot).

5.2 Diagnostic tests

One main focus of concern for the economic and financial evaluation of any econometric model is naturally whether
the model describes appropriately the dynamics of the data. In other words, it is important to test if the model
is correctly specified. If the model seems to be well specified, we can proceed with confidence and extract valuable
information from our model. In contrast, if there is strong evidence of model misspecification, then we should either
search for a better model or, at the very minimum, we should be very careful about the conclusions that we derive
from the model.
In the context of ARCH and GARCH models, we can use the residuals ut = yt /σ̂t to test for correct model
specification. The specification tests will fall into two main categories: (i) Homoscedasticity tests; and (ii) Normality
tests.
Homoscedasticity tests focus on the fact that the error term t is assumed to have fixed variance over time. Hence,
the residuals ut = yt /σ̂t ≈ t should also exhibit this feature, at least approximately. A simple way of testing the
homoscedasticity assumption, is to plot the autocorrelation function of the squared residuals {u2t }Tt=1 and analyzing
the Q-statistic associated to the ACF at each lag. In R we can obtain a vector, which we label u, that contains the
standardized residuals using the following code

u=x/sqrt(sigma2)

where x is the data vector and sigma2 contains the estimated conditional variance. An ACF for the squared
residuals can be obtained writing the command

acf(u^2,main="")

Figure 5.2 shows the squared residuals and the corresponding ACF obtained from the dataset stock returns.mat.

The normality tests focus on the fact that the error term t is assumed to be Gaussian. Hence, the residuals
ut = yt /σ̂t ≈ t should also be normal, at least approximately. The Jarque-Bera test and the qq-plot are two simple
ways of testing for normality of the residual term {ut }Tt=1 .

42
 squared residuals
15

10

0
0 100 200 300 400 500 600 700 800 900 1000

ACF residuals

0.2

-0.2

0 5 10 15 20 25 30

Figure 5.2: Squared residuals (first plot) and autocorrelation function of squared residuals (second plot).

The null hypothesis H0 of the Jarque-Bera test is that the residuals are normal against the alternative H1 that
the residuals are not normal. The Jarque-Bera test statistic is given by

T +1 2
µ̂3 + (µ̂4 − 3)2 ,


JB =
6

where µ̂3 denotes the sample skewness of the residuals {ut }Tt=1 and µ̂4 denotes the sample Kurtosis. Under the
null hypothesis H0 of normal residuals the Jarque-Bera statistics JB has a chi-square distribution with 2 degrees of
freedom χ22 . In R, a Jarque-Bera normality test for the residuals {ut }Tt=1 can be obtained using the command

jarque.bera.test(u)

The output gives several information including the p-value of the test.

5.3 Model Selection

Of course, in practice, it can happen that several competing models seem to be well specified, especially when they are
nested. In such cases, it is important to find ways of selecting the best model among the set of alternative competing
models.
Since we have learned how to estimate ARCH and GARCH models using the method of maximum likelihood, it
may seem natural to select the model that achieves the best log-likelihood for the given data set at hand. However, as
you may recall from your introductory econometrics courses, performing model selection by comparing log-likelihoods
leads to incorrect results. The reason for this is simple: nested models with a larger number of parameters can fit the
data better, hence they always achieve a higher log-likelihood in finite samples! It is important to highlight that a
higher sample log-likelihood may be achieved not because the model is better, but simply because it is able to overfit
the data. In other words, it can fit better the observed sample of data, but not the true dynamics.
One way of avoiding the problem of overfitting and incorrectly selecting larger models is to penalize the number
of parameters in the model. This is the main idea behind the majority of the so-called information criteria. Two

43
well known examples are the Akaike’s information criterion (AIC), and the Bayesian Information Criterion (BIC).
These information criteria take the value of the log-likelihood of the model log L(y1 , ..., yT ; θ̂T ), evaluated at the point
estimate θ̂T , and add a penalty for the number of parameters in the model k = dimension(θ),

AIC = 2k − 2 log L(y1 , ..., yT ; θ̂T ) ,

BIC = log(T )k − 2 log L(y1 , ..., yT ; θ̂T ) .

Note that both the AIC and BIC are based on a negative log-likelihood, so that the lower the value of the information
criterion, the better the model seems to be.
Comparing the AIC and BIC values of several models constitutes a reasonable basis for model selection.

5.4 Value-at-Risk
In finance, the Value-at-Risk (VaR) is a popular risk measure. Specifically, the daily α-VaR is the minimum amount
the investor stands to loose with probability α over a period of one day. For example, if a portfolio has a daily
10%-VaR of 1 million euros, this means that there is a 10% probability that the value of the portfolio will fall by
more than 1 million euros in one day.
The VaR can also stated in percentage loss. For example, if a portfolio has a daily 5%-VaR of 17%, then there is
a 5% probability that the value of the portfolio will fall by more than 17% of its value in one day. Mathematically,
given a portfolio value pt at time t, and a random percentual return yt = (pt − pt−1 )/pt−1 on the portfolio, the
5%-VaR in percentage loss is defined as the value c that satisfies

P (yt ≤ c) = 0.05.

Note that the VaR expressed in percentage loss can immediately be turned into the VaR in monetary loss by
multiplying c loss by the value of the stock at that time.
Definition 5.1. (Value-at-Risk - VaR) Let {pt }t∈Z be a random sequence of portfolio values, and yt = (pt −pt−1 )/pt−1
denote the return sequence (i.e. the percentage changes) on the portfolio. The estimated α-VaR in percentage loss
at time t is defined as the percentage value c that satisfies P (yt ≤ c) = α. Furthermore, the estimated α-VaR in
monetary loss at time t ∈ Z is given by c × pt .
Conditional volatility models like the GARCH, can easily provide us with an estimate of the time-varying VaR of
a stock at any time, conditional on past information. Indeed, since stock returns (typically expressed in percentage
changes or log-differences) are modeled to satisfy

yt = σt εt

the distribution of yt+1 conditional on the past and present Y t is easily tractable for any t. For example, when
the innovation sequence {εt } is N ID(0, 1), then yt+1 |Y t ∼ N (0, σt2 ). As a result, the VaR is obtained immediately
through application of the Gaussian quantile function Q (the inverse of the Gaussian cumulative distribution function
Φ(·))
Q(α) = inf{y ∈ R : α ≤ Φ(y)}.
In other words the α-VaR at time t + 1 is the value q that satisfies

P (yt+1 ≤ q|Y t ) = α,

that in the Gaussian case is

2
α-VaRt+1 = zα σt+1 ,
where zα is the quantile of level α of a standard Normal distribution. Obviously, the same reasoning applies if εt has
other distributions! Note that Monte Carlo simulations may be necessary for calculating a multiple step-ahead VaR.

Conditional VaR with R

The R file analysis GARCH.R also illustrates the calculation of the α-VaR for α = 0.1, 0.05, and 0.01, by applying
a GARCH filter to a time series of (percentage) stock returns. The first part of the code, as we have seen in the
previous section, provides the estimated conditional variance in the vector sigma2.
We obtain 3 vectors of length n, VaR10, VaR05 and VaR01, that contain the conditional VaR at each t for α = 0.1,
0.05, and 0.01. We use R quantile function for the normal qnorm() as follows:

44
VaR10 <- qnorm(0.1,0,sqrt(sigma2))
VaR05 <- qnorm(0.05,0,sqrt(sigma2))
VaR01 <- qnorm(0.01,0,sqrt(sigma2))

The output is shown in Figure 5.3 which plots the data (in black) and the respective conditional VaRs on the
bottom graph.

Time series
10

-5

-10
0 100 200 300 400 500 600 700 800 900 1000

VaR at 10%, 5% and 1% level

-2

-4

-6

-8

-10
0 100 200 300 400 500 600 700 800 900 1000

Figure 5.3: Conditional α-VaR for α = 0.1, α = 0.05, and α = 0.01, estimated from a GARCH(1,1) model.

5.5 Forecasting conditional volatility

Often we are interested in forecasting the risk of financial assets in the next time period. Here we will see in practice
how to obtain forecasts of the volatility of stock returns h-steps ahead.
In particular, assume we are at time T and we want to predict the variance of yT +h for h = {1, 2, . . . }. Indeed the
idea is to consider the variance of yT +h conditional on the past and present Y T as forecast. This variance forecast
σT2 (h) is given by
σT2 (h) = Var(yT2 +h |Y T ) = E(yT2 +h |Y T ) = E(σT2 +h |Y T ).
Note that it is immediate to see that when h = 1 we have σT2 (1) = σT2 +1 as the expectation of σT2 +1 conditional on
Y T is simply σT2 +1 . However, this is not the case for h > 1 and we have σT2 (h) 6= σT2 +h . We will now show how to
obtain σT2 (h) for ARCH and GARCH models.
For an ARCH(q) model σT2 (h) for h > 1 is given by
q
X
σT2 (h) = ω + αi σT2 (h − i),
i=1

where σT2 (h − i) = yT2 +h−i if h − i ≤ 0.

For the GARCH(1,1) model σT2 (h) for h > 1 is given by

σT2 (h) = ω + (α1 + β1 )σT2 (h − 1),

45
where the recursion is initialized at σT2 (1) = σT2 +1 .
In all cases it can be shown that σT2 (h) converges to the unconditional variance as h increases, provided that the
weak stationarity condition of the model is satisfied. For instance for the GARCH(1,1) model we have that σT2 (h)
converges to ω/(1 − β1 − α1 ). Therefore, in practical situations, when h is large we can approximate our forecast
σT2 (h) using the unconditional variance.
Finally, we note that σT2 (h) in practice cannot be directly obtained because the true parameter vector θ0 is
unknown. Therefore what we can do is to use the estimate σ̂T2 (h) obtained plugging in the maximum likelihood
estimate in place of θ0 .

5.6 Forecasting VaR and conditional density

Chapters 2 and 3 showed that ARCH and GARCH models provide a description of the conditional density of yt given
its past Y t−1 = {yt−1 , yt−2 , ...}. In particular, we have noted that, since σt2 is given when we condition on the past
Y t−1 , we obtain
yt |Y t−1 ∼ N (0, σt2 ) .
Now, given a sample of data {y1 , ..., yT }, one may naturally ask what is the conditional density of the future return
yT +1 . This is useful if, for example, we wish to calculate the VaR for the next trading day.
Luckily, since ARCH and GARCH models are observation-driven, the value of σT2 +1 is known, conditional on the
data {y1 , ..., yT }. As such, the conditional density of yT +1 is known conditional on y1 , ..., yT , and it is given by

yT +1 |y1 , ..., yT ∼ N (0, σT2 +1 ) .

Conditional densities for h-step ahead returns, yT +h , can also be obtained through simulations. Here, we shall
focus however on the one-step-ahead conditional density only. In particular, we note that the one-step-ahead VaR can
be obtained immediately through application of the Gaussian quantile function for a N (0, σT2 +1 ) random variable.

5.7 News Impact Curve

Finally, one interesting piece of information obtained immediately upon estimating the parameters of an ARCH or
GARCH model is the so-called news impact curve (NIC). Here, we shall the define the NIC as the updating function
2
that maps values of yt to values of σt+1 . In essence, the NIC fixes σt2 to some value σt2 = c, and looks at the GARCH
update as a function of yt .
Figure 5.4 below shows the NIC for a GARCH(1,1) model. The curve is obtained for ω = 0.1, β = 0.8, and three
different values of α = 0.05, 0.1, and 0.2. Moreover, it sets σt2 = 1. Hence, the plot corresponds to the following
function,
σt2 (yt ) = 0.9 + αyt2 for α = 0.05, 0.1, and 0.2.
The NIC shows how small absolute value of yt will lead to a decrease in the conditional volatility parameter (σt2 >
2
σt+1 ), and large values of yt will lead to an increase in the conditional volatility (σt2 < σt+1
2
).

46
 1.1

1.05
σt+1

0.95

0.9
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
yt

Figure 5.4: News impact curve for GARCH model.

47
48
Chapter 6

Multivariate GARCH models

In practical applications very often we deal with problems that involve multiple time series. For instance, we may
have several financial assets and be interested in choosing which assets to buy and which assets to sell. Assume
that we have n financial assets and the return of each asset i at time t is denoted by yi,t , for i = 1, . . . , n. We can
represent our data as a multivariate time series {yy t }Tt=1 , where y t = (y1,t , · · · , yn,t )> . In the previous chapter we
have seen univariate GARCH models, so models for a single element yi,t of our multivariate time series. A possible
approach could be to use n univariate models and deal with each component yi,t separately. This approach would
be appropriate if the series are independent as, in that case, there would not be any loss of information due to
the independence. However, empirical evidence suggests that different financial assets are not independent. More
specifically, it is well known that financial assets are typically positively correlated. For example this means that
when the price of Microsoft increases (decreases) the price of IBM tends to increase (decrease) as well. Figure 6.1
provides us the correlation matrix for all the stock returns in the S&P 100 index. As we can see, most stocks are
positively correlated with a correlation coefficient between 0.2 and 0.8. This further highlights the the importance of
using multivariate models that can take this dependence among different financial into account.

Figure 6.1: Sample correlation matrix of all daily stock log-returns in the S&P 100 index.

As we have discussed in the introduction of the course, past observations contain no information (or very little)
to predict the mean of financial assets. However, there is strong evidence of time variation in the variance. As you
have seen in previous econometric courses, the variance of a multivariate random variable is a matrix, which is often

49
 Rolling variance Microsoft Rolling variance IBM
10
12
10 8
8 6
6
4
4
2 2

0 1000 2000 3000 4000 0 1000 2000 3000 4000

Rolling covariance Rolling correlation

5
0.6
4
3
0.4
2
1
0.2
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Figure 6.2: Rolling covariance matrix and correlation estimated on a rolling window of length 250 using
daily log-returns of Microsoft and IBM.

called the covariance matrix. The covariance matrix contains the variance of each of the univariate variables in the
diagonal and the covariance between each pair of univariate variables outside the diagonal. At this point, the question
we want to ask is whether this covariance matrix is time varying and whether past observations contain information
to predict the covariance matrix. The answer is: yes, stock returns and more in general financial returns exhibit
time varying variances and covariances. To illustrate this fact in practical situations we can estimate the covariance
matrix on a rolling window. More specifically, given a multivariate sample {yy t }Tt=1 of length T we can fix a window
length h and for each t > h estimate the covariance matrix as follows
t
X
Σ̂t−h,t = y sy >
s .
s=t−h

In this way, we have sample estimates of the covariance matrix for T − h time windows and we can check whether
there is time variation. Figure 6.2 shows the covariance matrix and the correlation estimated on a rolling window
with h = 250 using daily log-returns of Microsoft and IBM. The figures suggest (in a qualitative way) that there is
indeed evidence of time variation. We can see that the estimated variance level changes dramatically over time. The
same happens for the covariance and the correlation. Note also that the correlation is always positive. This is also
coherent with the fact discussed before, namely, stock returns are positively correlated.
The time varying variance and covariance between stock returns has to be taken into account in order to properly
model the dynamics of multiple financial assets. Therefore we need appropriate multivariate GARCH models that
allows us to have time variation in the covariance matrix. In general, the idea is to consider multivariate models of
the form
1/2
y t = Σt εt,

where ε t = (ε1t , . . . , εnt )> is an n-dimensional error vector that is assumed to be i.i.d. with zero mean E(εεt ) = 0 n
and variance equal to the identity matrix Var(εεt ) = I n . More specifically, we will assume that ε t has a multivariate
normal distribution, namely ε t ∼ Nn (00n , I n ). The n × n-dimensional matrix Σ t depends only on the past Y t−1 =
{yy t−1 , y t−2 , . . . }. Similarly as for the univariate case, it can be easily shown that Σ t is the conditional covariance

50
matrix of y t given the past, i.e. Σ t = Cov(yy t |Y t−1 ) = E(yy ty >
t |Y
t−1
). Furthermore, the conditional distribution of y t
t−1
given Y is multivariate normal with zero mean and covariance matrix Σ t , namely y t |Y t−1 ∼ Nn (00n , Σ t ).
The conditional covariance matrix Σ t has the following form
 2 
σ1t σ12t ... σ1nt

2
.. .. 
.
 
 σ12t σ2t . 
Σt = 
 .
.
 . .. .. 
 . . . σ(n−1)nt 

2
σ1nt ... σ(n−1)nt σnt
2
where each diagonal element of the matrix σit denotes the conditional variance of yit , i = 1, . . . , n , and each other
element σijt denotes the conditional covariance between yit and yjt , i, j = 1, . . . , n, i 6= j.
In the bivariate case n = 2 the conditional covariance matrix is given by
" #
2
σ1t σ12t
Σt = 2
.
σ12t σ2t
Note the matrix Σ t is symmetric and also positive definite.
In the rest of this section, we will see several multivariate GARCH specifications. These GARCH specifications
differ on how the conditional covariance matrix Σ t is specified. Extending univariate GARCH models to a multivariate
setting leads to some issues such as the curse of dimensionality, i.e. too many parameters to be estimated. These
issues need to be taken into account when specifying multivariate GARCH models.

6.1 The VECH model

The most natural way to extend univariate GARCH models to the multivariate case is given by the VECH model.
The specification of the VECH model is based on the half vectorization operator vech(·).

The vech(·) operator stacks the lower triangular elements of a squared matrix into a vector. For instance
consider the 3 × 3 matrix A given by
 
α11 α12 α13
A = α21 α22 α23  ,
 

α31 α32 α33

then
α11
 
α21 
 
 
α31 
A) = 
vech(A α  .

 22 
 
α32 
α33
In general, for an n × n matrix A the operator vech(·) produces a vector of length ñ = n(n + 1)/2 that
contains all the lower triangular elements of the n × n matrix.

The idea is to represent the covariance matrix Σ t in vector form using the vech(·) operator. Note that we only need
to consider the lower triangular part of Σ t because Σ t is symmetric.
In the bivariate case, i.e. n = 2, the 2 × 2 covariance matrix Σ t of a VECH(1,1) model is specified as

2 2 2
         
σ1,t ω̃1 β̃11 β̃12 β̃12 σ1,t−1 α̃11 α̃12 α̃13 y1,t−1
Σt ) = σ12,t  = ω̃2  + β̃21
vech(Σ β̃22 β̃23  σ12,t−1  + α̃21 α̃22 α̃23  y1,t−1 y2,t−1  (6.1)
         
2 2 2
σ2,t ω̃3 β̃31 β̃32 β̃33 σ2,t−1 α̃31 α̃32 α̃33 y2,t−1

51
 2 2
The conditional covariance matrix Σ t depends on past squared observations y1,t−1 and y2,t−1 and on the product
y1,t−1 y2,t−1 . For the univariate GARCH we have discussed how past squared observations are a natural way to
2
update the time varying variance because a squared observation y1,t−1 can be seen as an estimate at time t − 1 of
2
the variance of y1,t−1 . A very similar idea applies for the covariance. In particular, the product y1,t−1 y2,t−1 can be
seen as estimate of the covariance between y1,t−1 and y2,t−1 . This justifies from an intuitive point of view the VECH
specification in (6.1).
We also note that the specification of the conditional covariance matrix in (6.1) is very general. For instance, the
conditional variance of y1t is given by
2 2 2 2 2
σ1t = β̃11 σ1,t−1 + β̃12 σ12,t−1 + β̃12 σ2,t−1 + α̃11 y1,t−1 + α̃12 y1,t−1 y2,t−1 + α̃13 y2,t−1 ,
2
therefore σ1t depends on all lagged vales of the conditional covariance matrix and also on lagged squared observations
2 2
y1,t−1 and y2,t−1 and on the product y1,t−1 y2,t−1 . However we can see that this specification requires a lot of
parameters. Even for this first order and bivariate case we already have 21 parameters to estimate!
In the general multivariate case, the conditional covariance matrix of the VECH(p,q) model with general order p
and q is given by
q p
X X
Σt ) = W̃ +
vech(Σ Ã i vech(yy t−iy >
t−i ) + Σt−i ),
B̃ i vech(Σ (6.2)
i=1 i=1

where W̃ is an ñ-dimensional vector of parameters and B̃ i and à i are square matrices of parameters of dimension
ñ × ñ, where ñ = n(n + 1)/2.
Although the conditional covariance matrix Σ t is time varying, the unconditional covariance matrix of a VECH(p,q)
is constant.

Remark 6.1. The unconditional covariance matrix Σ = Var (yy t ) of the VECH(p,q) model, when it exists, is given
by
q p
!−1
X X
Σ) = I ñ −
vech(Σ Ã i − B̃ i W̃ .
i=1 i=1

The VECH model has 2 main drawbacks that typically prevent its use in practical applications. The first is the
so called “curse of dimensionality”. The problem is that as the dimension n of the vector y t increases, the number
of parameters is of order O(n4 ). In particular, the number of parameters to estimate in the VECH(p,q) model is
ñ + (p + q)ñ2 . In practice, even when we have just a few financial assets to model the number of parameters becomes
huge. Just to have an idea of the problem, assume that we have n = 4 financial assets and we consider p = q = 1,
then the number of parameters to estimate is 210. If we have n = 10 stocks, then we have 6105 parameters. This is
completely infeasible in practical applications. The second problem of the VECH model is that it is not clear what
type of restrictions we should impose on the parameter matrices W̃ , B̃ i and à i to ensure that Σ t is a positive definite
matrix for any t.

6.2 The DVECH model

A possible solution to solve the “curse of dimensionality” problem of the VECH model is to impose that the matrices
B̃ i and à i are diagonal. This leads to the diagonal VECH model, which is often called the DVECH model. The
DVECH model is a special case of the VECH model.
For the Bivariate case, i.e. n = 2, the conditional covariance matrix of the DVECH(1,1) model is given by

2 2 2
         
σ1,t ω̃1 β̃11 0 0 σ1,t−1 α̃11 0 0 y1,t−1
σ21,t  = ω̃2  +  0 β̃22 0  σ21,t−1  +  0 α̃22 0  y1,t−1 y2,t−1  (6.3)
         
2 2 2
σ2,t ω̃3 0 0 β̃33 σ2,t−1 0 0 α̃33 y2,t−1

Compared to bivariate VECH(1,1) model, the DVECH specification in (6.3) reduces the number of parameters from
21 to 9. The DVECH allows us to attenuate the “curse of dimensionality” problem, however, this comes at a price.
The DVECH model does not allow for causality in variance. This can be noted from the expressions of the conditional

52
 2 2
variances σ1,t , σ2,t and the covariance σ12,t that are given by
2 2 2
σ1,t = ω11 + β11 σ1,t−1 + α11 y1,t−1 ,
2 2 2
σ2,t = ω22 + β22 σ2,t−1 + α22 y2,t−1 ,
2
σ21,t = ω21 + β21 σ21,t−1 + α21 y1,t−1 y2,t−1

2 2 2 2
The variances σ1,t and σ2,t depend only on their past values y1,t−1 and y2,t−1 respectively. Therefore, large values of
2 2
y1,t−1 does not have any effect on y2,t .
The DVECH model in (6.3) can be also rewritten in matrix form by relying on the Hadamard matrix product.

The Hadamard product is a matrix operation that given 2 matrices of the same dimension produces a
matrix where each element ij is the product of the elements ij of the two original matrices. For instance,
given two 3 × 3 matrices A = (αij ) and B = (βij ), we have that
     
α11 α12 α13 β11 β12 β13 α11 β11 α12 β12 α13 β13
A B = α21 α22 α23  β21 β22 β23  = α21 β21 α22 β22 α23 β23  .
     

α31 α32 α33 β31 β32 β33 α31 β31 α32 β32 α33 β33

The bivariate DVECH(1,1) conditional covariance matrix in (6.3) can be rewritten in matrix form as

" # " # " # " # " # " #

2 2 2
σ1,t σ21,t ω11 ω21 β11 β21 σ1,t−1 σ21,t−1 α11 α21 y1,t−1 y1,t−1 y2,t−1
2
= + 2
+ 2
σ21,t σ2,t ω21 ω22 β21 β22 σ21,t−1 σ2,t−1 α21 α22 y1,t−1 y2,t−1 y2,t−1

Note that all the the matrices in the above formulation are symmetric because the covariance matrix must be
symmetric.
Figure 6.3 shows the conditional variances, covariance and correlation generated from a bivariate DVECH model.

Conditional variance σ 21t Conditional variance σ 22t

4 6

3
4
2
2
1

0 0
0 500 1000 0 500 1000

Conditional covariance σ 12t Conditional correlation ρ 12t

2 0.6

0.4
1 0.2

0
0
-0.2

0 500 1000 0 500 1000

Figure 6.3: Conditional variances, covariance and correlation generated from a DVECH model.

53
 A general multivariate DVECH(1,1) model can thus be written as follows

Σt = W + A1 (yy t−1y >

t−1 ) + B 1 Σ t−1 , (6.4)

where W , B 1 and A 1 are symmetric n × n matrices containing the parameters to be estimated. A natural extension
to the VECH(1,1) model is the VECH(p,q) model. The conditional covariance matrix of a n-variate DVECH(p,q)
model is given by
q p
X X
Σt = W + Ai (yy t−iy >
t−i ) + Bi Σ t−i , (6.5)
i=1 i=1

where W , B i and A i are symmetric n × n matrices containing the parameters to be estimated.

The DVECH model is in fact a special case of the VECH model where the matrices of parameters are imposed
to be diagonal. This constraint attenuates the “curse of dimensionality”. However, a limitation of this model is that
it does not ensure that the conditional covariance Σ t be positive definite.

6.3 The scalar DVECH model

A very simple version of the DVECH model is the scalar DVECH (sDVECH) model. The sDVECH imposes that the
parameters in each of the matrices B i and A i are the same. The conditional covariance matrix of an n-dimensional
sDVECH(1,1) can be written as follows

Σ t = W + α1 y t−1y >
t−1 + β1 Σ t−1 , (6.6)

where W is a symmetric n × n matrix, and α1 and β1 are scalar parameters to be estimated. The name “scalar” is
due to the fact that the matrices B i and A i are replaced by scalars. The sDVECH(1,1) model in (6.6) is a special
case of the DVECH(1,1) model in (6.4) with matrices B 1 and A 1 given by
   
β1 ... β1 α1 ... α1
. .. ..   . .. .. 
 ..
B1 =  . . ,  ..
A1 =  . . .
 

β1 ... β1 α1 ... α1

The sDVECH model is widely used in practice when the number of stocks n is large. The sDVECH model strongly
attenuates the curse of dimensionality. The number of parameters of the sDVECH(1,1) in (6.6) is n(n + 1)/2 + 2.
The extension from the sDVECH(1,1) to the sDVECH(p,q) is straightforward. In particular, the conditional
covariance matrix of the sDVECH(p,q) is given by
q p
X X
Σt = W + αi y t−iy >
t−i + βi Σ t−i . (6.7)
i=1 i=1

where W is a symmetric n × n matrix, and {α1 , . . . , αq } and {β1 , . . . , βp } are scalar parameters to be estimated.

6.4 The BEKK model

In order to solve the issue of having a positive definite conditional covariance matrix Σ t , the BEKK model has been
proposed. Also the BEKK is a special case of the more general VECH model. In the bivariate case, the conditional
covariance matrix of the BEKK(1,1) model is given by

" # " #" # " #" #" #

2 2
σ1,t σ21,t ω11 0 ω11 ω21 β11 β12 σ1,t−1 σ21,t−1 β11 β21
2
= + 2
+ (6.8)
σ21,t σ2,t ω21 ω22 0 ω22 β21 β22 σ21,t−1 σ2,t−1 β12 β22
" #" #" #
2
α11 α12 y1,t−1 y1,t−1 y2,t−1 α11 α21
+ 2
α21 α22 y1,t−1 y2,t−1 y2,t−1 α12 α22

54
This specification allows us to easily obtain a positive definite covariance matrix Σ t , see Remark 6.2 for conditions
in a more general case. The updating equation for the conditional covariance matrix of a multivariate BEKK(1,1)
model is thus given by

Σ t = W W > + A 1 (yy t−1y > A>

t−1 )A
>
1 + B 1Σ t−1B 1 , (6.9)

The extension from the BEKK(1,1) to the BEKK(p,q) model is also trivial. In particular, the conditional covariance
matrix of a BEKK(p,q) model is updated as follows,
q p
X X
Σt = W W > + A i (yy t−iy > A>
t−i )A i + B iΣ t−iB >
i , (6.10)
i=1 i=1

where W is a lower triangular n × n matrix and A i and B i are n × n matrices. The conditions to ensure that Σ t is
positive definite are easy to be met for the BEKK model.
Remark 6.2. The BEKK(p,q) model in (6.10) has a positive definite conditional covariance matrix Σ t for any t ∈ N
if Σ 0 , . . . , Σ −p−1 are positive definite and W or any B i is a full rank matrix.
As already discussed, the advantage of the BEKK model is that we can ensure the covariance matrix to be positive
definite. One of the disadvantages of the BEKK formulation is that the parameters are difficult to interpret.

6.5 The CCC model

Another approach to deal with the curse of dimensionality is provided by the Constant Conditional Correlation (CCC)
model. As the name suggests, the peculiarity of this model is that the conditional correlation matrix is constant and
the time variation in the conditional covariance matrix Σ t is only provided by dynamic variances. This model, unlike
the DVECH and the BEKK, is not a special case of the very general VECH model.
The conditional covariance matrix of a bivariate CCC model is given by
" # " #" #" #
2
σ1t σ12t σ1t 0 1 ρ12 σ1t 0
Σt = 2
= , (6.11)
σ12t σ2t 0 σ2t ρ12 1 0 σ2t
2 2
where the conditional variances σ1,t and σ2,t are specified as
2 2 2
σ1,t = ω1 + β1 σ1,t−1 + α1 y1,t , (6.12)
2 2 2
σ2,t = ω2 + β2 σ2,t−1 + α2 y2,t . (6.13)

This model is called CCC because the conditional correlation matrix R is constant
" #
1 ρ12
R= .
ρ12 1
In particular, the conditional correlation between y1t and y2t is given by Corr(y1t , y2t |Y t−1 ) = ρ12 . This does not mean
that the conditional covariance is constant. In fact, the conditional covariance is time varying Cov(y1t , y2t |Y t−1 ) =
2 2
σ12t = σ1t σ2t ρ12 . Figure 6.4 below shows the conditional variances σ1,t and σ2,t , the conditional covariance σ12t
and the conditional correlation ρ12 generated from a CCC model. The pictures clearly illustrate how the conditional
covariance is time varying even though the conditional variance is constant.
The bivariate CCC model can be easily extended to a general multivariate case of order n. The conditional
covariance matrix of an n-variate CCC model is given by

Σ t = D t RD t , (6.14)

where R is an n × n correlation matrix and D t = diag(σ1,t , . . . , σn,t ) is an n × n diagonal matrix containing the
2
conditional standard deviation of each component. The variance of each component σi,t , i = 1, . . . , n, follows a
GARCH(1,1) dynamic, namely
2 2 2
σi,t = ωi + βi σi,t−1 + αi yi,t .
The CCC model is very appealing because it handles both the curse of dimensionality and the positive definiteness
of Σ t . However, the assumption that the conditional correlation matrix is constant can be very restrictive. In practical
applications, we often see evidence of changing conditional correlation.

55
 conditional variance σ 21t conditional variance σ 22t

0.2 0.2

0.15 0.15

0.1 0.1

0 100 200 300 400 500 0 100 200 300 400 500

conditional covariance σ 12t conditional correlation ρ 12

1

0.08

0.06 0.5

0.04
0
0 100 200 300 400 500 0 100 200 300 400 500
Figure 6.4: Conditional variances, covariance and correlation generated from a CCC model.

6.6 The DCC model

The assumption of constant conditional correlation is often too restrictive. For this reason, the Dynamic Conditional
Correlation (DCC) model has been introduced. The specification of the DCC is similar to the CCC the difference
lies on the fact that the conditional correlation matrix R varies over time.
The conditional covariance matrix of the bivariate DCC model is given by

" # " #" #" #

2
σ1t σ12t σ1t 0 1 ρ12t σ1t 0
Σt = 2
= , (6.15)
σ12t σ2t 0 σ2t ρ12t 1 0 σ2t
2 2
where the conditional variances σ1,t and σ2,t are specified as
2 2 2
σ1,t = ω1 + β1 σ1,t−1 + α1 y1,t ,
2 2 2
σ2,t = ω2 + β2 σ2,t−1 + α2 y2,t .
So far the specification is the same as the CCC model but we are now left with the specification of the dynamic
conditional correlation ρ12t . First, we need to define the standardized observations v1t = y1t /σ1t and v2t = y2t /σ2t .
The conditional variances of v1t and v2t are equal to 1, i.e. Var(v1t |Y t−1 ) = Var(v2t |Y t−1 ) = 1 and the conditional
covariance (correlation) between v1t and v2t is given by E(v1t v2t |Y t−1 ) = ρ12t . Note that v1t and v2t are different
from the errors ε1t and ε2t . Given the definition of v1t and v2t , we are now ready to specify the conditional correlation
as
q12t
ρ12t = √ √ , (6.16)
q11t q22t
where
2
q11t = ωq + βq q11t−1 + αq v1,t−1 ,
2
q22t = ωq + βq q22t−1 + αq v2,t−1 ,
q12t = ωq + βq q12t−1 + αq v1,t−1 v2,t−1 .

56
The formulation in (6.16) for the conditional correlation is needed to ensure that ρ12t is between −1 and 1. Note
that each equation for qijt has the same static parameters ωq , βq and αq .

6.7 Other extensions

All the multivariate GARCH models presented in this section can be extended to include a mean different from zero
or also a time varying conditional mean. For instance, including a constant mean the multivariate GARCH model
becomes
1/2
y t = µ + Σt εt, (6.17)

where µ = (µ1 , . . . , µn )> is an n-dimensional vector containing the means of each element yi,t . The specification of
Σ t can then be one of those discussed in the previous section.
Finally, we mention that there are several different specifications that we have not discussed in this course such
as the Factor GARCH model and Multivariate GAS models.

6.8 Simulate from a bivariate DVECH(1,1) with R

This section describes how to simulate from a bivariate DVECH model with p = q = 1 using R. The code can be
found on the R file generate DVECH.R. The first step is to set the sample size T , which is labeled n, and choose the
parameter values. For simplicity, we use the notation that the parameters ω11 , ω22 and ω12 are labeled w11, w22, w12.
The same is true for the βij and αij parameters where we use bij and aij, for i, j = 1, 2.

n <- 1000

w11 <- 0.1

w22 <- 0.2
w12 <- 0.02

b11 <- 0.7

b22 <- 0.7
b12 <- 0.7

a11 <- 0.2

a22 <- 0.2
a12 <- 0.15

The second step is to define the matrices x and VECHt that will contain the generated series {yy t }Tt=1 and the generated
conditional covariance matrix {vech(Σ Σt )}Tt=1 . The initial value vech(Σ
Σ1 ) of the conditional covariance matrix is set
to be equal to the unconditional covariance matrix.

x <- matrix(0,nrow = n, ncol = 2)

VECHt <- matrix(0,nrow=n,ncol=3)

VECHt[1,1] <- w11/(1-b11-a11)

VECHt[1,3] <- w22/(1-b22-a22)
VECHt[1,2] <- w12/(1-b12-a12)

Next, we generate the first observation y1 . The R function mvrnorm() is used to generate from a multivariate
normal distribution. This function mvrnorm() is part of the package MASS. Once we have Σ 1 , we can generate y1
from a bivariate normal with mean 0 and covariance matrix Σ 1 . The first argument of mvnrnd() is the number of
observations to generate (1 in our case), the second argument is the mean (a vector of zeros) and the third argument
is the covariance matrix. Therefore, we first need to get vech(Σ Σ1 ) in matrix form. For this reason we define a 2
matrix labeled SIGMAt where we put in matrix form vech(Σ Σ1 ).

57
SIGMAt <- cbind(c(VECHt[1,1],VECHt[1,2]),c(VECHt[1,2],VECHt[1,3]))
x[1,] <- mvrnorm(1,rep(0,2),SIGMAt)

Σt ) of the
Finally, we are ready to use a for loop to obtain the generated series. We iterate the equation of vech(Σ
bivariate DVECH model with the observation equation for t from 2 to T .

for(t in 2:n){
VECHt[t,1] <- w11 + b11*VECHt[t-1,1] + a11*x[t-1,1]^2
VECHt[t,3] <- w22 + b22*VECHt[t-1,3] + a22*x[t-1,2]^2
VECHt[t,2] <- w12 + b12*VECHt[t-1,2] + a12*x[t-1,1]*x[t-1,2]

SIGMAt <- cbind(c(VECHt[t,1],VECHt[t,2]),c(VECHt[t,2],VECHt[t,3]))

x[t,] <- mvrnorm(1,rep(0,2),SIGMAt)
}

Figure 6.5 shows a generated series from a bivariate DVECH model.

Series y 1t

-2

0 100 200 300 400 500 600 700 800 900 1000

Series y 2t
4

-2

-4
0 100 200 300 400 500 600 700 800 900 1000
Figure 6.5: Series generated from a bivariate DVECH model.

58
Chapter 7

Estimation of multivariate GARCH

models

7.1 Maximum likelihood estimation

Multivariate GARCH models can be estimated by maximum likelihood. The log-likelihood function is given by
T
1 X 
L(y1 , ..., yT , θ) = − Σt | + y >
log |Σ −1
t Σt y t .
2 t=1

The covariance matrix Σ t is obtained recursively using the the observed data and the updating equation of the specific
multivariate GARCH we are estimating. For instance, for a bivariate VECH model with p = q = 1 we can use the
following updating equations to obtain Σ t
2 2 2
σ1,t = ω11 + β11 σ1,t−1 + α11 y1,t−1 ,
2 2 2
σ2,t = ω22 + β22 σ2,t−1 + α22 y2,t−1 ,
2
σ21,t = ω21 + β21 σ21,t−1 + α21 y1,t−1 y2,t−1 ,
for t = 1, . . . , T . As for the univariate case, the updating equations need to be initialized. A practical way is to set
the initial condition Σ 1 equal to the sample covariance matrix.
Once the log-likelihood function is obtained, the estimation of a multivariate GARCH is equivalent to the esti-
mation of a univariate GARCH. In particular, the ML estimator θ̂T is the maximizer of the log-likelihood function
θ̂T = arg max L(y1 , ..., yT , θ).
θ∈Θ

Furthermore, in large samples, the following approximation for the distribution of θ̂T holds true
√   app  
T θ̂T − θ0 ∼ N 0 , I(θ0 )−1 .

See the section on the estimation of univariate GARCH models for more details and how to estimate the Fisher
information matrix I(θ0 ).

7.2 Estimating a bivariate scalar DVECH with R

In this section, we will see how to estimate a bivariate sDVECH model by maximum likelihood using R. More specifi-
cally, we are only going to discuss how to write the log-likelihood function because the optimization is then equivalent
to the univariate GARCH case (see Section 4.5 see how to optimize the likelihood). The R file estimation sDVECH.R
contains the code to optimize the log-likelihood function of the sDVECH model. The log-likelihood function is labeled
llik fun sDVECH(,) and it is contained in the file llik fun sDVECH.R. We need to create an R function that takes as
argument the observed time series, labeled x, and a parameter vector, labeled par, and gives as output the average
log-likelihood value.
The first line of code defines the name of the function and the input.

59
llik_fun_sDVECH <- function(par,x){

Then, the time series length is obtained and each parameter value is set equal to an element of the input parameter
vector par using some appropriate link functions. Furthermore, the matrix VECHt that will contain the conditional
covariance matrix is defined.

w11 <- exp(par[1])

w12 <- par[2]
w22 <- exp(par[3])
a <- exp(par[4])/(1+exp(par[4]))
b <- exp(par[5])/(1+exp(par[5]))

d <- dim(x)
n <- d[1]

VECHt <- matrix(0,nrow=n,ncol=3)

Now, the conditional variance is initialized using the sample covariance matrix. The average log-likelihood output
llik is defined and set to zero.

llik <- 0

C <- cov(x)
VECHt[1,] <- c(C[1,1],C[1,2],C[2,2])

Finally, a for loop allows us to obtain recursively the conditional covariance matrix using the updating equation of
the sDVECH model. Furthermore, the average log-likelihood is summed up at each iteration of the loop (see last line
of code in the for loop). Note that in R matrix product is obtained with the operator %*%. The average log-likelihood
is then returned as output of the function.

for(t in 2:n){

VECHt[t,1] <- w11+b*VECHt[t-1,1]+a*x[t-1,1]^2

VECHt[t,3] <- w22+b*VECHt[t-1,3]+a*x[t-1,2]^2
VECHt[t,2] <- w12+b*VECHt[t-1,2]+a*x[t-1,1]*x[t-1,2]

SIGMAt <- cbind(c(VECHt[t,1],VECHt[t,2]),c(VECHt[t,2],VECHt[t,3]))

llik <- llik-0.5*(log(det(SIGMAt))+x[t,]%*%solve(SIGMAt)%*%t(t(x[t,])))/n
}

return(llik)
}

The full code to obtain the log-likelihood function is given below.

llik_fun_sDVECH <- function(par,x){

w11 <- exp(par[1])

w12 <- par[2]
w22 <- exp(par[3])

60
 a <- exp(par[4])/(1+exp(par[4]))
b <- exp(par[5])/(1+exp(par[5]))

d <- dim(x)
n <- d[1]

VECHt <- matrix(0,nrow=n,ncol=3)

llik <- 0

C <- cov(x)
VECHt[1,] <- c(C[1,1],C[1,2],C[2,2])

for(t in 2:n){

VECHt[t,1] <- w11+b*VECHt[t-1,1]+a*x[t-1,1]^2

VECHt[t,3] <- w22+b*VECHt[t-1,3]+a*x[t-1,2]^2
VECHt[t,2] <- w12+b*VECHt[t-1,2]+a*x[t-1,1]*x[t-1,2]

SIGMAt <- cbind(c(VECHt[t,1],VECHt[t,2]),c(VECHt[t,2],VECHt[t,3]))

llik <- llik-0.5*(log(det(SIGMAt))+x[t,]%*%solve(SIGMAt)%*%t(t(x[t,])))/n
}

return(llik)
}

Figure 7.1 provides a plot of conditional volatility and conditional correlations between the stock returns of Google
and IBM.

conditional sd Google
6

0 500 1000 1500 2000 2500 3000

conditional sd IBM
4

0 500 1000 1500 2000 2500 3000

conditional correlation
0.8
0.6
0.4
0.2
0

0 500 1000 1500 2000 2500 3000

Figure 7.1: Conditional variances and correlations estimated from IBM and Google log-returns.

61
7.3 Estimation of the sDVECH model with covariance targeting
One of the main issues in estimating multivariate GARCH models is that the large number of parameters can cause
numerical problems when optimizing the log-likelihood function. Covariance targeting is a 2-steps method that
allows us to reduce the number of parameters in the likelihood optimization. In particular, in a first step, the
unconditional covariance Σ is estimated using the sample covariance of the log-returns. Then, in a second step,
the estimated unconditional covariance is plugged-in into the log-likelihood function and the remaining parameters
are estimated by optimizing the resulting log-likelihood. In the following, we focus on covariance targeting for the
sDVECH model, but we note that covariance targeting can applied to VECH models in general.
The first step to perform covariance targeting is to reparametrize the model in terms of the unconditional co-
variance. As discussed before, the conditional covariance matrix Σ t = Var(yy t |Y t−1 ) of the sDVECH(p,q) is given
by
q p
X X
Σt = W + αi y t−iy >
t−i + βi Σ t−i . (7.1)
i=1 i=1

Furthermore, it can be shown that the unconditional covariance matrix Σ = Var(yy t ) = E(yy ty >
t ) of the sDVECH(p,q)
model is given by
q p
!−1
X X
Σ = 1− αi − βi W.
i=1 i=1

Therefore, the matrix W in (7.1) can be expressed as

q p
!
X X
W = 1− αi − βi Σ .
i=1 i=1

PT
Σ = T −1
The unconditional variance Σ can be estimated using the observed data as follows: Σ̂ >
t=1 y ty t . Finally, we
Σ into expression (7.1) and obtain
can plug-in the expression of W with the estimated Σ̂
q p q p
!
X X X X
Σt = 1 − αi − Σ+
βi Σ̂ αi y t−iy >
t−i + βi Σ t−i . (7.2)
i=1 i=1 i=1 i=1

The updating equation in (7.2) is used to obtain the log-likelihood function. The parameters {α1 , . . . , αq } and
{β1 , . . . , βp } can be estimated by maximum likelihood as discussed in Section 7.1. In this way, we are left with
only p + q parameters to be estimated in the likelihood optimization. This is very appealing because the number of
parameters in the optimization does not depend on the number of stocks. This makes estimation much more reliable
and fast.
The covariance targeting approach is summarized by the following steps:
Σ = T −1 Tt=1 y ty > >
P
1. Estimate the unconditional covariance Σ from the log-returns Σ̂ t , where y t = (y1t , . . . , ynt ) .

2. Obtain the log-likelihood function using the updating equation given in (7.2).
3. Maximize the log-likelihood function to estimate the parameters {α1 , . . . , αq } and {β1 , . . . , βp }.

7.4 Estimating a sDVECH model with CT in R

In the following, we illustrate how to estimate a sDVECH(1,1) model by covariance targeting with R. We are going
to see how the likelihood function can be obtained by incorporating the estimated covariance Σ̂ Σ. The resulting
likelihood function can then be optimized with respect to the parameters β1 and α1 (see Section 4.5 on how to
optimize the likelihood). The R file CT estimation sDVECH.R contains the code to optimize the log-likelihood function
of the sDVECH model with covariance targeting. The log-likelihood function with covariance targeting is labeled
llik CT sDVECH(,) and it is contained in the file llik CT sDVECH.R. We need to create an R function that takes as
argument the observed time series, labeled x, and a parameter vector that contains β1 and α1 , labeled par, and gives
as output the average log-likelihood value.
The first line of code defines the name of the function and the input.

llik_CT_sDVECH <- function(par,x){

62
 Then, the time series length is obtained and each parameter value is set equal to an element of the input parameter
vector par using appropriate link functions. Note that the only parameters that enter into the likelihood are β1 and α1
since we are using the covariance targeting approach. Furthermore, the matrix VECHt that will contain the conditional
covariance matrix is defined and the average log-likelihood output llik is defined and set to zero.

a <- exp(par[1])/(1+exp(par[1]))
b <- exp(par[2])/(1+exp(par[2]))

d <- dim(x)
n <- d[1]

VECHt <- matrix(0,nrow=n,ncol=3)

llik <- 0

Now, the sample covariance of the observed data is obtained. Furthermore, the conditional covarince matrix is
initialized using the sample covariance matrix.

C <- cov(x)
VECHt[1,] <- c(C[1,1],C[1,2],C[2,2])

Finally, a for loop allows us to obtain recursively the conditional covariance matrix using the updating equation
of the sDVECH model. Note that W is replaced by Σ̂ Σ(1 − β1 − α1 ) as described before in the application of the
covariance targeting approach. Furthermore, the average log-likelihood is summed up at each iteration of the loop
(see last line of code in the for loop).

for(t in 2:n){

VECHt[t,1] <- C[1,1]*(1-a-b)+b*VECHt[t-1,1]+a*x[t-1,1]^2

VECHt[t,3] <- C[2,2]*(1-a-b)+b*VECHt[t-1,3]+a*x[t-1,2]^2
VECHt[t,2] <- C[1,2]*(1-a-b)+b*VECHt[t-1,2]+a*x[t-1,1]*x[t-1,2]

SIGMAt <- cbind(c(VECHt[t,1],VECHt[t,2]),c(VECHt[t,2],VECHt[t,3]))

llik <- llik-0.5*(log(det(SIGMAt))+x[t,]%*%solve(SIGMAt)%*%t(t(x[t,])))/n

}

return(llik)

The full code to obtain the log-likelihood function with covariance targeting is given below.

llik_CT_sDVECH <- function(par,x){

a <- exp(par[1])/(1+exp(par[1]))
b <- exp(par[2])/(1+exp(par[2]))

d <- dim(x)
n <- d[1]

VECHt <- matrix(0,nrow=n,ncol=3)

llik <- 0

C <- cov(x)
VECHt[1,] <- c(C[1,1],C[1,2],C[2,2])

63
 for(t in 2:n){

VECHt[t,1] <- C[1,1]*(1-a-b)+b*VECHt[t-1,1]+a*x[t-1,1]^2

VECHt[t,3] <- C[2,2]*(1-a-b)+b*VECHt[t-1,3]+a*x[t-1,2]^2
VECHt[t,2] <- C[1,2]*(1-a-b)+b*VECHt[t-1,2]+a*x[t-1,1]*x[t-1,2]

SIGMAt <- cbind(c(VECHt[t,1],VECHt[t,2]),c(VECHt[t,2],VECHt[t,3]))

llik <- llik-0.5*(log(det(SIGMAt))+x[t,]%*%solve(SIGMAt)%*%t(t(x[t,])))/n

}

return(llik)
}

7.5 Estimation of the CCC model equation by equation

The CCC model can be also estimated through equation by equation approach. This approach involves only the
estimation of univariate GARCH models. Then the constant correlation matrix R is estimated using the standardized
residuals obtained from the univariate GARCH. This method is appealing because there is no need to optimize the
log-likelihood function over the full parameter vector of the model. When the dimension of the parameter vector is
large, numerical optimization methods become time-consuming and we are also more likely to encounter numerical
problems.
The steps to estimate a CCC model are the following:
1. Estimate a univariate GARCH model for each series {yit }Tt=1 , i = 1, . . . , n.
2. Obtain the standardized errors from each of these series ε̂it = (yit − µ̂i )/σ̂it , i = 1, . . . , n.
R = T −1 Tt=1 ε̂εtε̂ε> εt = (ε̂1t , . . . , ε̂nt )> .
P
3. Estimate the correlation matrix from the residuals R̂ t , where ε̂

7.6 Estimating a CCC model equation by equation with R

In the following, we show how to estimate a bivariate CCC model using R. The code can be found on the R file
estimation CCC.R. The bivariate time series is contained in the matrix labeled x. As discussed in the previous section,
the first step is to estimate univariate GARCH(1,1) models for each of the series. The following lines of code do
that. In case you do not remember how to estimate a univariate GARCH model yous can go back to Chapter 4. The
parameter estimates for the first series are labeled omega hat1, alpha hat1 and beta hat1 whereas, the parameter
estimates for the second series are labeled omega hat2, alpha hat2 and beta hat2.

est1 <- optim(par=par_ini,fn=function(par)-llik_fun_GARCH(par,x[,1]), method = "BFGS")

est2 <- optim(par=par_ini,fn=function(par)-llik_fun_GARCH(par,x[,2]), method = "BFGS")

The second step is to obtain the standardized residuals. This is done through the following R code. In case you
do not remember how to obtain standardized residuals from univariate GARCH models, you can go back to Chapter
5.

n <- length(x[,1])
s1 <- rep(0,n)
s2 <- rep(0,n)

s1[1] <- var(x[,1])

s2[1] <- var(x[,2])

for(t in 2:n){
s1[t] <- omega_hat1 + alpha_hat1*x[t-1,1]^2 + beta_hat1*s1[t-1]

64
 s2[t] <- omega_hat2 + alpha_hat2*x[t-1,2]^2 + beta_hat2*s2[t-1]
}

e1 <- x[,1]/sqrt(s1)
e2 <- x[,2]/sqrt(s2)

Finally, the last step is to to calculate the correlation between the residuals of the first series and and the residuals
of the second series. This can be done using the R function cor().

r <- cor(e1,e2)

It is easy also to extend this equation by equation method to a larger dimension n instead of the bivariate case
with n = 2. What we need is just to estimate n univariate GARCH models, get the residuals and compute their
correlation matrix.

65
66
Chapter 8

Financial Analysis of Multivariate

GARCH models

8.1 VaR portfolio prediction

Assume we have a portfolio that contains n assets, where yi,t denotes the return of asset i at time t. The vector
y t = (y1,t , . . . , yn,t )> represents the vector of returns at time t. The proportion of our portfolio invested asset i at
time t is given by kit ∈ [0, 1]. The Pn quantity kit is called the weight of asset i in the portfolio. Indeed, the portfolio
>
weights have to sum to 1, i.e. i=1 kit = 1. The weights can be stacked into a vector k t = (k1t , . . . , knt ) . We
therefore obtain that the return of our portfolio yp,t at time t is given by
n
X
yp,t = ki,t yi,t = k >
t y t.
i=1

We consider a multivariate GARCH model for y t with conditional mean equal to µ t = (µ1t , . . . , µnt )> and covariance
matrix Σ t . Note that so far we have considered that the conditional mean is zero, namely µ t = 0 n . This because
stock returns shows no autocorrelation (or very little) and a sample mean close to zero. In general, as an alternative
to the zero mean assumption, we can either choose a static conditional mean µ t = µ or choose a dynamic specification
for µ t (AR model may be an option).
Therefore, as discussed in the previous sections, we have that y t |Y t−1 ∼ Nn (µ µt , Σ t ). We can now find the
conditional distribution of our portfolio return at time t, which is
yp,t |Y t−1 ∼ N (µp,t , σp,t
2
),
where the portfolio conditional mean µp,t is equal to
 
µp,t = E yp,t |Y t−1 = E k > t−1
= k> t−1

= k>


t y t |Y t E y t |Y t µt,

2
and the conditional variance σp,t is given by
 
2
= Var yp,t |Y t−1 = Var k > t−1
= k> t−1

kt = k>


σp,t t y t |Y t Var y t |Y t Σ tk t .

In case we only have 2 financial assets, i.e. n = 2, the conditional mean µp,t is given by
µp,t = k >
t µ t = k1,t µ1,t + k2,t µ2,t ,
2
and the conditional variance σp,t is given by
" 2 #" #
k1,t k2,t σ1,t σ12,t k1,t

2 > 2 2 2 2
σp,t = k t Σ tk t = 2
= k1,t σ1,t + k2,t σ2,t + 2k1,t k2,t σ12,t .
σ12,t σ2,t k2,t
Finally, we obtain that the conditional α-VaR of the portfolio at time t is given by
α-VaRt = µp,t + zα σp,t ,
where zα is the quantile of level α of a standard normal distribution.

67
8.2 Dynamic portfolio optimization
Another useful application of multivariate GARCH models is dynamic portfolio optimization. We have a portfolio of
assets and at each time period we need to choose which assets to sell and buy. Therefore, at time t we have to decide
the portfolio weights for time t + 1. The idea is to choose the weights in such a way to maximize our utility function,
i.e. high returns but low volatility. In the following we consider that our objective is to choose portfolio weights, k t ,
to maximize the so called Shape ratio, which is given by
µp,t
Sp,t = .
σp,t

The intuition behind maximizing the Shape ratio is that we want high portfolio returns µp,t and at the same time
low risk (volatility) σp,t . The optimization problem can be written as
n
k >µ t X
max p t , s.t. ki,t = 1, kit ≥ 0. (8.1)
kt k>t Σ tk t i=1

This problem in general can be solved using numerical routines in R.

In the simplest case where we only have 2 risky assets, i.e. n = 2, the optimization problem in (8.1) can be written
as
k1t µ1t + (1 − k1t )µ2t
max p , s.t. 0 ≤ k1t ≤ 1. (8.2)
2 2 2
k1t k1t σ1t + (1 − k1t )2 σ2t + 2k1t (1 − k1t )σ12t

For this special case there is a closed form solution for the optimal weights when the constraint k1t , k2t ≥ 0 is not
imposed. In some situations this is reasonable as we can have a so called short position on an asset, namely a negative
portfolio weight. The optimal weights are then given by
2
µ1t σ2t − µ2t σ12t
k1t = 2 2
,
µ1t σ2t + µ2t σ1t − (µ1t + µ2t )σ12t
k2t = 1 − k1t . (8.3)

8.3 Dynamic portfolio optimization with R

The R file portfolio CCC.R provides you the code to obtain optimal portfolio weights in terms of maximum Sharpe
Ratio. In this file a CCC model is used to obtain the estimate of the conditional covariance matrix Σt . Furthermore,
the conditional mean µ t is assumed to be constant, i.e. µ t = µ , and it is estimated using the sample mean of the
log-returns. In general, a different model can be also used and the conditional mean can be also time varying. In
the following, we will only discuss how to obtain the optimal weights for a given conditional expectation µ t and
conditional covariance matrix Σt .
The constant conditional means for the first series and the second series are labeled mu1 and mu2 respectively.
2 2
Instead, the vectors s1, s2 and s12 contain σ1t , σ2t and σ12t respectively, for each time t = 1, . . . , T .
First, we create a matrix labeled kt that will contain the optimal portfolio weights. Then we use a for loop to obtain
the portfolio weights at time each time t from t = 1, . . . , T . This is achieved through the function max SR portfolio()
which is contained in the R file max SR portfolio.R. The function requires 2 inputs. The first argument requires
the conditional mean vector µ t and the second argument requires conditional covariance matrix Σ t . Note that
max SR portfolio() gives the weights that maximize the Sharpe Ratio under the constraint that each weight is
positive.

kt = matrix(0,nrow=n,ncol=2)

mu1 <- mean(x[,1])

mu2 <- mean(x[,2])
mut <- cbind(mu1,mu2)

for(t in 1:n){
SIGMAt <- cbind(c(s1[t],s12[t]),c(s12[t],s2[t]))

68
 kt[t,] <- max_SR_portfolio(mut,SIGMAt)
}

Figure 8.1 shows the optimal portfolio weights that maximize the Sharpe Ratio. The series considered are monthly
log-returns of Microsoft and IBM.

Portfolio weight to Microsoft k1t

1

0.5

0
0 50 100 150 200 250 300

Portfolio weight to IBM k2t

1

0.5

0
0 50 100 150 200 250 300

Figure 8.1: Optimal portfolio weights obtained using Microsoft and IBM monthly log-returns

8.4 Out-of-sample evaluation of different portfolio strategies

There are different methods we can use to decide on which assets to invest. For instance, above we have given an
example of dynamic portfolio optimization based on the CCC model. However, we could use a different model and
this would lead to different portfolio weights. The question that arises now is: how can we decide which portfolio
strategy is best? The answer is simple: we can consider a sub-sample of the observed data and see how different
strategies perform in this sub-sample. In particular, we can take our observed dataset {yy t }, for t = 1, . . . , T , and split it
into two sub-samples: an in-sample dataset, for t = 1, . . . , T1 , and an out-of-sample dataset, for t = T1 + 1, . . . , T . We
can then use the in-sample dataset to estimate our models and the out-of-sample dataset to evaluate the performance
of our portfolio strategies. To evaluate the performance of our portfolios we can calculate their means and variances
and obtain their sharpe ratios. We can then choose the portfolio strategy that delivers the highest sharpe ratio.
The out-of-sample portfolio evaluation is as follows:
1. Estimate a multivariate GARCH model using the in-sample dataset, t = 1, . . . , T1 . The estimation can be
based on ML, covariance targeting or equation-by-equation, depending on the model we are using.
2. For the out-of-sample dataset, t = T1 + 1, . . . , T , obtain an estimate of the conditional mean µ t and the
conditional covariance matrix Σ t of the returns by using the multivariate GARCH model estimated in the
previous point.
3. For the out-of-sample dataset, t = T1 + 1, . . . , T , use the conditional covariance matrix and the conditional
mean to obtain the log-returns of the optimal portfolio (see Sections 8.2 and 8.3).

69
4. Use the log-returns of the optimal portfolio {yp,t }, for t = T1 + 1, . . . , T , to obtain an empirical estimate of the
sharpe ratio of the portfolio as
ȳp
Ŝp = ,
σ̂p
where
T T
1 X 1 X
ȳp = yp,t , σ̂p2 = (yp,t − ȳp )2 .
T − T1 t=T +1 T − T1 t=T +1
1 1

70
Appendix A

Stock Return Properties: Empirical

Evidence

71
Table A.1: p-values of ADF test for log returns of S&P100 stocks.

daily daily weekly weekly monthly monthly

Stock prices returns prices returns prices returns

AAPL 0.239 0.001 0.188 0.001 0.230 0.001

ABBV 0.615 0.001 0.643 0.001 0.651 0.001
ABT 0.372 0.001 0.606 0.001 0.619 0.001
ACN 0.308 0.001 0.369 0.001 0.379 0.001
AGN 0.184 0.001 0.206 0.001 0.266 0.001
AIG 0.607 0.001 0.573 0.001 0.677 0.001
ALL 0.411 0.001 0.509 0.001 0.484 0.001
AMGN 0.198 0.001 0.296 0.001 0.387 0.001
AMZN 0.054 0.001 0.051 0.001 0.069 0.001
AXP 0.486 0.001 0.431 0.001 0.500 0.001
BA 0.352 0.001 0.490 0.001 0.506 0.001
BAC 0.561 0.001 0.687 0.001 0.768 0.001
BIIB 0.191 0.001 0.177 0.001 0.266 0.001
BK 0.462 0.001 0.574 0.001 0.552 0.001
BLK 0.296 0.001 0.362 0.001 0.398 0.001
BMY 0.197 0.001 0.458 0.001 0.558 0.001
BRK-B 0.417 0.001 0.516 0.001 0.518 0.001
C 0.589 0.001 0.479 0.001 0.563 0.001
CAT 0.421 0.001 0.510 0.001 0.560 0.001
CELG 0.159 0.001 0.144 0.001 0.223 0.001
CL 0.355 0.001 0.554 0.001 0.607 0.001
CMCSA 0.198 0.001 0.336 0.001 0.357 0.001
COF 0.431 0.001 0.247 0.001 0.293 0.001
COP 0.525 0.001 0.583 0.001 0.584 0.001
COST 0.198 0.001 0.287 0.001 0.338 0.001
CSCO 0.344 0.001 0.262 0.001 0.251 0.001
CVS 0.294 0.001 0.386 0.001 0.369 0.001
CVX 0.454 0.001 0.564 0.001 0.588 0.001
DD 0.387 0.001 0.620 0.001 0.660 0.001
DHR 0.202 0.001 0.457 0.001 0.431 0.001
DIS 0.300 0.001 0.346 0.001 0.204 0.001
DOW 0.345 0.001 0.509 0.001 0.557 0.001
DUK 0.366 0.001 0.392 0.001 0.352 0.001
EMR 0.471 0.001 0.659 0.001 0.684 0.001
EXC 0.520 0.001 0.581 0.001 0.558 0.001
F 0.515 0.001 0.598 0.001 0.711 0.001
FB 0.494 0.001 0.512 0.001 0.526 0.001
FDX 0.382 0.001 0.419 0.001 0.457 0.001
FOX 0.459 0.001 0.482 0.001 0.543 0.001
FOXA 0.449 0.001 0.481 0.001 0.549 0.001
GD 0.263 0.001 0.429 0.001 0.462 0.001
GE 0.484 0.001 0.412 0.001 0.599 0.001
GILD 0.274 0.001 0.256 0.001 0.295 0.001
GM 0.691 0.001 0.703 0.001 0.731 0.001
GOOG 0.348 0.001 0.435 0.001 0.473 0.001
GOOGL 0.338 0.001 0.448 0.001 0.474 0.001
GS 0.510 0.001 0.529 0.001 0.542 0.001
HAL 0.357 0.001 0.431 0.001 0.406 0.001
HD 0.147 0.001 0.236 0.001 0.330 0.001
HON 0.253 0.001 0.390 0.001 0.400 0.001

72
Table A.2: (continued) p-values of ADF test for log returns of S&P100 stocks.

daily daily weekly weekly monthly monthly

Stock prices returns prices returns prices returns

IBM 0.488 0.001 0.495 0.001 0.604 0.001

INTC 0.313 0.001 0.356 0.001 0.115 0.001
JNJ 0.311 0.001 0.470 0.001 0.494 0.001
JPM 0.366 0.001 0.521 0.001 0.474 0.001
KHC 0.692 0.001 0.701 0.999 0.676 0.126
KMI 0.714 0.001 0.739 0.001 0.734 0.001
KO 0.473 0.001 0.613 0.001 0.644 0.001
LLY 0.358 0.001 0.498 0.001 0.562 0.001
LMT 0.141 0.001 0.215 0.001 0.241 0.001
LOW 0.159 0.001 0.277 0.001 0.312 0.001
MA 0.433 0.001 0.189 0.001 0.246 0.001
MCD 0.340 0.001 0.406 0.001 0.435 0.001
MDLZ 0.439 0.001 0.602 0.001 0.612 0.001
MDT 0.330 0.001 0.422 0.001 0.466 0.001
MET 0.522 0.001 0.548 0.001 0.617 0.001
MMM 0.274 0.001 0.432 0.001 0.473 0.001
MO 0.126 0.001 0.598 0.001 0.477 0.001
MON 0.435 0.001 0.492 0.001 0.494 0.001
MRK 0.437 0.001 0.562 0.001 0.514 0.001
MS 0.418 0.001 0.428 0.001 0.532 0.001
MSFT 0.249 0.001 0.425 0.001 0.692 0.001
NKE 0.191 0.001 0.542 0.001 0.639 0.001
ORCL 0.332 0.001 0.119 0.001 0.250 0.001
OXY 0.428 0.001 0.475 0.001 0.500 0.001
PCLN 0.188 0.001 0.201 0.001 0.271 0.001
PEP 0.378 0.001 0.477 0.001 0.494 0.001
PFE 0.413 0.001 0.520 0.001 0.470 0.001
PG 0.458 0.001 0.581 0.001 0.581 0.001
PM 0.551 0.001 0.609 0.001 0.608 0.001
PYPL 0.715 0.001 0.711 0.001 0.697 0.017
QCOM 0.346 0.001 0.011 0.001 0.153 0.001
RTN 0.175 0.001 0.306 0.001 0.295 0.001
SBUX 0.160 0.001 0.361 0.001 0.413 0.001
SLB 0.418 0.001 0.518 0.001 0.518 0.001
SO 0.375 0.001 0.536 0.001 0.560 0.001
SPG 0.175 0.001 0.281 0.001 0.321 0.001
T 0.375 0.001 0.631 0.001 0.609 0.001
TGT 0.420 0.001 0.438 0.001 0.497 0.001
TWX 0.195 0.001 0.098 0.001 0.447 0.001
TXN 0.158 0.001 0.141 0.001 0.103 0.001
UNH 0.136 0.001 0.191 0.001 0.221 0.001
UNP 0.295 0.001 0.477 0.001 0.542 0.001
UPS 0.507 0.001 0.586 0.001 0.603 0.001
USB 0.398 0.001 0.473 0.001 0.579 0.001
UTX 0.412 0.001 0.509 0.001 0.582 0.001
V 0.459 0.001 0.494 0.001 0.514 0.001
VZ 0.379 0.001 0.584 0.001 0.577 0.001
WBA 0.329 0.001 0.366 0.001 0.386 0.001
WFC 0.396 0.001 0.530 0.001 0.582 0.001
WMT 0.486 0.001 0.499 0.001 0.526 0.001
XOM 0.488 0.001 0.591 0.001 0.599 0.001
73
Table A.3: Estimated ACF(1), MA(1) and AR(1) coefficients for log returns of S&P100 stocks.

Daily Weekly Monthly

ACF(1) MA(1) AR(1) ACF(1) MA(1) AR(1) ACF(1) MA(1) AR(1)

AAPL -0.026 -0.026 -0.026 0.037 0.040 0.037 0.036 0.035 0.036
ABBV -0.031 -0.031 -0.031 -0.122 -0.171 -0.122 -0.061 -0.094 -0.061
ABT -0.010 -0.011 -0.011 -0.030 -0.034 -0.030 -0.059 -0.056 -0.060
ACN -0.026 -0.031 -0.026 -0.068 -0.072 -0.068 0.043 0.045 0.043
AGN -0.003 -0.003 -0.003 -0.022 -0.022 -0.023 0.037 0.036 0.039
AIG 0.137 0.125 0.137 -0.023 -0.023 -0.023 0.185 0.208 0.186
ALL -0.057 -0.062 -0.057 -0.076 -0.065 -0.076 -0.001 -0.001 -0.001
AMGN -0.043 -0.050 -0.043 -0.045 -0.048 -0.045 -0.042 -0.059 -0.042
AMZN 0.008 0.009 0.008 -0.005 -0.005 -0.005 -0.015 -0.017 -0.015
AXP -0.057 -0.062 -0.057 -0.009 -0.009 -0.009 0.091 0.138 0.091
BA 0.004 0.005 0.004 -0.066 -0.066 -0.066 -0.009 -0.008 -0.009
BAC -0.008 -0.008 -0.008 -0.036 -0.036 -0.036 0.111 0.112 0.111
BIIB 0.014 0.016 0.014 -0.054 -0.048 -0.054 0.001 0.001 0.001
BK -0.118 -0.135 -0.118 -0.072 -0.070 -0.101 -0.134 -0.148 -0.135
BLK -0.035 -0.038 -0.035 -0.129 -0.118 -0.129 -0.109 -0.092 -0.112
BMY -0.018 -0.020 -0.018 -0.025 -0.025 -0.025 -0.074 -0.090 -0.074
BRK-B 0.029 0.029 0.029 -0.028 -0.027 -0.028 -0.005 -0.005 -0.005
C 0.056 0.056 0.056 -0.040 -0.040 -0.040 0.082 0.096 0.083
CAT -0.005 -0.005 -0.005 -0.046 -0.046 -0.046 -0.033 -0.038 -0.033
CELG -0.001 -0.001 -0.001 0.005 0.005 0.005 -0.052 -0.058 -0.053
CL -0.009 -0.011 -0.009 -0.060 -0.061 -0.060 -0.154 -0.163 -0.154
CMCSA -0.063 -0.067 -0.063 -0.066 -0.073 -0.066 -0.073 -0.071 -0.073
COF -0.031 -0.031 -0.031 -0.069 -0.064 -0.069 0.113 0.116 0.113
COP -0.038 -0.042 -0.038 -0.063 -0.063 -0.063 -0.029 -0.032 -0.029
COST -0.015 -0.016 -0.015 -0.125 -0.132 -0.125 0.018 0.024 0.018
CSCO -0.044 -0.047 -0.044 -0.049 -0.050 -0.050 0.009 0.010 0.009
CVS -0.057 -0.060 -0.057 -0.054 -0.058 -0.054 -0.028 -0.025 -0.028
CVX -0.069 -0.069 -0.069 -0.038 -0.042 -0.039 -0.107 -0.109 -0.108
DD -0.011 -0.011 -0.011 -0.076 -0.070 -0.076 -0.021 -0.022 -0.021
DHR -0.013 -0.013 -0.013 -0.007 -0.006 -0.007 -0.049 -0.050 -0.049
DIS -0.028 -0.030 -0.028 -0.024 -0.024 -0.024 0.159 0.140 0.160
DOW -0.032 -0.032 -0.032 0.038 0.039 0.039 0.071 0.072 0.071
DUK -0.038 -0.038 -0.038 -0.020 -0.021 -0.020 -0.021 -0.022 -0.021
EMR -0.063 -0.073 -0.063 -0.064 -0.064 -0.064 -0.073 -0.081 -0.073
EXC -0.035 -0.037 -0.035 -0.033 -0.034 -0.033 0.114 0.132 0.114
F 0.012 0.012 0.012 -0.028 -0.028 -0.028 -0.087 -0.084 -0.087
FB 0.022 0.021 0.022 0.073 0.075 0.081 0.006 0.007 0.006
FDX -0.001 -0.001 -0.001 -0.014 -0.014 -0.014 -0.044 -0.042 -0.045
FOX 0.025 0.026 0.025 -0.009 -0.008 -0.009 -0.036 -0.037 -0.036
FOXA 0.006 0.006 0.006 -0.027 -0.027 -0.027 -0.059 -0.060 -0.060
GD -0.035 -0.035 -0.035 0.034 0.035 0.034 0.010 0.011 0.010
GE -0.011 -0.011 -0.011 -0.039 -0.039 -0.039 -0.034 -0.039 -0.034
GILD -0.019 -0.021 -0.019 -0.061 -0.068 -0.062 0.031 0.039 0.032
GM 0.026 0.025 0.026 -0.040 -0.041 -0.041 -0.034 -0.043 -0.035
GOOG 0.010 0.010 0.010 -0.020 -0.020 -0.020 0.098 0.117 0.104
GOOGL 0.010 0.010 0.010 0.001 0.001 0.001 0.098 0.112 0.104
GS -0.045 -0.049 -0.045 -0.152 -0.133 -0.152 0.083 0.098 0.083
HAL 0.019 0.021 0.019 -0.102 -0.095 -0.102 0.082 0.092 0.082
HD 0.009 0.010 0.009 -0.066 -0.068 -0.066 0.064 0.088 0.064
HON -0.001 -0.001 -0.001 0.006 0.006 0.006 0.065 0.099 0.066

74
Table A.4: (continued) Estimated ACF(1), MA(1) and AR(1) coefficients for log returns of S&P100 stocks.

Stock d.ACF(1) d.MA(1) d.AR(1) w.ACF(1) w.MA(1) w.AR(1) m.ACF(1) m.MA(1) m.AR(1)

IBM -0.027 -0.027 -0.027 -0.023 -0.023 -0.023 -0.206 -0.220 -0.208
INTC -0.042 -0.044 -0.042 -0.038 -0.040 -0.038 -0.048 -0.038 -0.049
JNJ 0.007 0.008 0.007 -0.098 -0.099 -0.098 -0.111 -0.124 -0.111
JPM -0.063 -0.067 -0.063 -0.088 -0.088 -0.088 -0.061 -0.057 -0.061
KHC 0.055 0.068 0.055 -0.063 -0.392 -0.065 -0.178 -0.366 -0.352
KMI 0.064 0.065 0.064 -0.053 -0.054 -0.054 0.164 0.202 0.165
KO 0.001 0.001 0.001 -0.021 -0.021 -0.021 -0.056 -0.063 -0.056
LLY -0.022 -0.025 -0.022 -0.040 -0.044 -0.040 -0.140 -0.165 -0.142
LMT -0.073 -0.073 -0.073 -0.012 -0.012 -0.012 0.115 0.143 0.115
LOW 0.019 0.021 0.019 -0.028 -0.029 -0.028 0.007 0.007 0.007
MA -0.058 -0.061 -0.058 0.016 0.017 0.017 -0.011 -0.011 -0.011
MCD -0.018 -0.018 -0.018 -0.042 -0.046 -0.042 0.036 0.033 0.036
MDLZ -0.055 -0.059 -0.055 -0.018 -0.018 -0.018 -0.104 -0.112 -0.104
MDT -0.002 -0.002 -0.002 -0.047 -0.047 -0.047 0.002 0.003 0.002
MET -0.073 -0.073 -0.073 -0.038 -0.042 -0.038 0.027 0.065 0.029
MMM -0.039 -0.039 -0.039 -0.063 -0.063 -0.063 -0.113 -0.131 -0.113
MO -0.042 -0.046 -0.042 -0.045 -0.045 -0.046 0.062 0.071 0.062
MON 0.003 0.003 0.003 -0.041 -0.049 -0.041 0.024 0.025 0.024
MRK 0.002 0.002 0.002 -0.042 -0.043 -0.042 -0.084 -0.069 -0.085
MS 0.010 0.011 0.010 -0.173 -0.178 -0.173 0.006 0.009 0.006
MSFT -0.033 -0.033 -0.033 -0.004 -0.004 -0.004 -0.143 -0.164 -0.144
NKE -0.010 -0.010 -0.010 -0.024 -0.024 -0.024 -0.069 -0.093 -0.070
ORCL -0.053 -0.059 -0.053 0.053 0.053 0.053 -0.056 -0.074 -0.057
OXY -0.051 -0.057 -0.051 -0.037 -0.037 -0.037 -0.022 -0.026 -0.022
PCLN 0.041 0.042 0.041 0.021 0.020 0.021 0.206 0.150 0.220
PEP -0.070 -0.078 -0.070 -0.058 -0.062 -0.060 -0.030 -0.039 -0.030
PFE -0.006 -0.007 -0.006 -0.049 -0.051 -0.049 -0.066 -0.064 -0.067
PG -0.040 -0.040 -0.040 -0.083 -0.083 -0.083 0.081 0.103 0.082
PM -0.036 -0.044 -0.036 -0.094 -0.101 -0.094 -0.059 -0.075 -0.059
PYPL 0.109 0.110 0.109 -0.141 -0.172 -0.177 -0.161 -0.259 -0.248
QCOM -0.031 -0.034 -0.031 -0.016 -0.016 -0.016 -0.113 -0.127 -0.113
RTN 0.028 0.029 0.028 -0.024 -0.024 -0.024 0.044 0.044 0.044
SBUX -0.073 -0.078 -0.073 -0.038 -0.043 -0.038 -0.044 -0.050 -0.044
SLB -0.031 -0.034 -0.031 -0.101 -0.105 -0.101 -0.015 -0.015 -0.015
SO -0.055 -0.055 -0.055 -0.127 -0.137 -0.127 -0.147 -0.248 -0.147
SPG -0.169 -0.173 -0.169 -0.062 -0.073 -0.062 0.003 0.003 0.003
T -0.026 -0.026 -0.026 -0.096 -0.103 -0.096 -0.017 -0.019 -0.017
TGT -0.054 -0.062 -0.054 -0.052 -0.053 -0.053 0.104 0.141 0.104
TWX 0.007 0.008 0.007 0.017 0.019 0.017 -0.214 -0.238 -0.216
TXN 0.005 0.006 0.005 -0.119 -0.114 -0.119 -0.005 -0.005 -0.005
UNH -0.006 -0.006 -0.006 -0.013 -0.014 -0.013 -0.003 -0.003 -0.003
UNP -0.016 -0.017 -0.016 -0.042 -0.043 -0.042 -0.033 -0.036 -0.033
UPS -0.039 -0.045 -0.040 -0.081 -0.081 -0.082 -0.040 -0.077 -0.040
USB -0.010 -0.010 -0.010 -0.117 -0.110 -0.117 -0.063 -0.065 -0.063
UTX -0.039 -0.039 -0.039 -0.016 -0.016 -0.016 -0.104 -0.119 -0.104
V -0.093 -0.106 -0.097 0.013 0.013 0.013 -0.031 -0.031 -0.033
VZ -0.066 -0.070 -0.066 -0.006 -0.006 -0.006 -0.086 -0.093 -0.086
WBA -0.049 -0.049 -0.049 -0.057 -0.053 -0.057 0.007 0.009 0.008
WFC -0.097 -0.098 -0.097 -0.125 -0.120 -0.125 -0.075 -0.079 -0.075
WMT -0.022 -0.025 -0.022 -0.059 -0.060 -0.059 -0.034 -0.040 -0.034
XOM -0.098 -0.118 -0.098 -0.069 -0.070 -0.069 -0.042 -0.042 -0.042

75
 Table A.5: Moments of Log Returns for S&P100 Stocks

Stock Mean Var Skew Kurt JB Mean Var Skew Kurt JB Mean Var Skew Kurt JB

AAPL 0.000 0.00 14.5 356 0.001 0.009 0.03 22.8 519 0.001 0.040 0.11 10.8 117 0.001
ABBV 0.000 0.00 8.6 110 0.001 0.002 0.00 5.1 38 0.001 0.004 0.00 1.3 4 0.006
ABT 0.000 0.00 10.9 165 0.001 0.002 0.00 22.4 509 0.001 0.006 0.00 10.6 115 0.001
ACN 0.000 0.00 14.1 294 0.001 0.001 0.00 6.7 72 0.001 0.004 0.00 2.0 7 0.001
AGN 0.000 0.00 13.3 297 0.001 0.001 0.00 10.9 160 0.001 0.004 0.00 3.3 19 0.001
AIG 0.002 0.00 28.8 1051 0.001 0.024 0.08 20.6 450 0.001 0.123 0.42 7.3 60 0.001
ALL 0.000 0.00 14.3 258 0.001 0.002 0.00 16.8 332 0.001 0.008 0.00 7.5 64 0.001
AMGN 0.000 0.00 9.6 131 0.001 0.001 0.00 6.6 64 0.001 0.005 0.00 5.1 38 0.001
AMZN 0.001 0.00 13.8 264 0.001 0.003 0.00 8.9 123 0.001 0.011 0.00 5.5 42 0.001
AXP 0.001 0.00 9.2 113 0.001 0.003 0.00 6.4 51 0.001 0.010 0.00 8.5 83 0.001
BA 0.000 0.00 9.5 160 0.001 0.002 0.00 7.5 90 0.001 0.006 0.00 4.6 30 0.001
BAC 0.001 0.00 10.3 134 0.001 0.006 0.00 9.5 102 0.001 0.023 0.00 5.9 44 0.001
BIIB 0.001 0.00 25.5 770 0.001 0.002 0.00 10.2 127 0.001 0.009 0.00 3.3 15 0.001
BK 0.001 0.00 15.9 377 0.001 0.002 0.00 6.3 56 0.001 0.005 0.00 3.8 23 0.001
BLK 0.001 0.00 9.0 115 0.001 0.002 0.00 7.3 78 0.001 0.008 0.00 5.5 40 0.001
BMY 0.000 0.00 6.4 60 0.001 0.001 0.00 5.5 44 0.001 0.004 0.00 2.1 7 0.001
BRK-B 0.000 0.00 20.5 627 0.001 0.029 0.41 22.8 520 0.001 0.120 1.67 10.8 118 0.001
C 0.002 0.00 17.4 408 0.001 0.018 0.05 20.9 459 0.001 0.067 0.20 10.3 109 0.001
CAT 0.000 0.00 8.8 119 0.001 0.002 0.00 5.2 39 0.001 0.010 0.00 5.3 35 0.001
CELG 0.000 0.00 10.9 167 0.001 0.003 0.00 21.5 481 0.001 0.011 0.00 8.6 85 0.001
CL 0.000 0.00 11.0 176 0.001 0.001 0.00 22.3 505 0.001 0.006 0.00 10.7 117 0.001
CMCSA 0.000 0.00 17.9 477 0.001 0.002 0.00 16.5 324 0.001 0.007 0.00 9.8 103 0.001
COF 0.001 0.00 9.5 123 0.001 0.005 0.00 12.4 196 0.001 0.014 0.00 8.1 77 0.001
COP 0.000 0.00 9.6 137 0.001 0.002 0.00 11.4 156 0.001 0.007 0.00 4.9 30 0.001
COST 0.000 0.00 11.9 209 0.001 0.001 0.00 6.5 58 0.001 0.003 0.00 2.8 13 0.001
CSCO 0.000 0.00 11.8 188 0.001 0.002 0.00 6.2 56 0.001 0.006 0.00 2.7 12 0.001
CVS 0.000 0.00 25.3 871 0.001 0.001 0.00 6.6 64 0.001 0.004 0.00 2.8 13 0.001
CVX 0.000 0.00 16.4 385 0.001 0.001 0.00 17.0 343 0.001 0.004 0.00 2.2 8 0.001
DD 0.000 0.00 7.7 82 0.001 0.002 0.00 4.4 27 0.001 0.007 0.00 3.0 12 0.001
DHR 0.000 0.00 8.8 118 0.001 0.002 0.00 22.0 496 0.001 0.008 0.00 10.6 115 0.001
DIS 0.000 0.00 10.3 156 0.001 0.001 0.00 11.6 189 0.001 0.004 0.00 3.1 14 0.001
DOW 0.001 0.00 11.4 212 0.001 0.003 0.00 4.6 28 0.001 0.014 0.00 6.7 54 0.001
DUK 0.000 0.00 16.5 379 0.001 0.003 0.00 20.6 443 0.001 0.014 0.01 10.1 106 0.001
EMR 0.000 0.00 12.5 231 0.001 0.002 0.00 22.4 509 0.001 0.009 0.00 10.3 110 0.001
EXC 0.000 0.00 14.2 292 0.001 0.001 0.00 10.2 149 0.001 0.004 0.00 3.1 13 0.001
F 0.001 0.00 12.9 208 0.001 0.005 0.00 14.5 230 0.001 0.023 0.01 7.0 53 0.001
FB 0.001 0.00 17.0 382 0.001 0.003 0.00 5.5 43 0.001 0.014 0.00 3.5 15 0.001
FDX 0.000 0.00 9.8 176 0.001 0.002 0.00 5.4 43 0.001 0.006 0.00 2.7 11 0.001
FOX 0.000 0.00 12.2 210 0.001 0.002 0.00 7.5 70 0.001 0.007 0.00 3.2 15 0.001
FOXA 0.001 0.00 10.6 157 0.001 0.002 0.00 7.1 65 0.001 0.008 0.00 4.6 29 0.001
GD 0.000 0.00 7.9 102 0.001 0.001 0.00 7.6 78 0.001 0.005 0.00 4.2 23 0.001
GE 0.000 0.00 10.2 157 0.001 0.002 0.00 9.5 128 0.001 0.007 0.00 4.8 30 0.001
GILD 0.000 0.00 11.0 163 0.001 0.003 0.00 15.8 255 0.001 0.014 0.00 7.8 65 0.001
GM 0.000 0.00 7.5 87 0.001 0.002 0.00 3.7 21 0.001 0.007 0.00 3.0 14 0.001
GOOG 0.000 0.00 13.1 243 0.001 0.003 0.00 21.9 494 0.001 0.012 0.00 10.2 109 0.001
GOOGL 0.000 0.00 13.2 247 0.001 0.003 0.00 21.9 491 0.001 0.012 0.00 10.0 106 0.001
GS 0.001 0.00 12.2 194 0.001 0.003 0.00 10.1 124 0.001 0.009 0.00 3.8 22 0.001
HAL 0.001 0.00 10.6 164 0.001 0.003 0.00 13.1 228 0.001 0.012 0.00 5.5 40 0.001
HD 0.000 0.00 9.3 141 0.001 0.002 0.00 8.3 97 0.001 0.004 0.00 2.5 10 0.001
HON 0.000 0.00 6.9 70 0.001 0.001 0.00 7.2 81 0.001 0.005 0.00 5.5 41 0.001

76
 Table A.6: (continued) Moments of Log Returns for S&P100 Stocks

Stock Mean Var Skew Kurt JB Mean Var Skew Kurt JB Mean Var Skew Kurt JB

IBM 0.000 0.00 9.0 134 0.001 0.001 0.00 6.0 49 0.001 0.003 0.00 5.4 41 0.001
INTC 0.000 0.00 8.2 102 0.001 0.001 0.00 5.1 37 0.001 0.005 0.00 3.2 16 0.001
JNJ 0.000 0.00 19.0 526 0.001 0.000 0.00 13.8 239 0.001 0.002 0.00 3.6 20 0.001
JPM 0.001 0.00 10.4 140 0.001 0.003 0.00 9.7 107 0.001 0.008 0.00 2.9 12 0.001
KHC 0.000 0.00 3.7 19 0.001 0.001 0.00 1.9 6 0.001 0.003 0.00 1.0 3 0.079
KMI 0.000 0.00 9.8 140 0.001 0.002 0.00 13.2 200 0.001 0.006 0.00 7.5 59 0.001
KO 0.000 0.00 18.3 481 0.001 0.002 0.00 22.3 505 0.001 0.007 0.00 10.7 117 0.001
LLY 0.000 0.00 13.8 286 0.001 0.001 0.00 17.1 345 0.001 0.003 0.00 6.1 44 0.001
LMT 0.000 0.00 9.3 118 0.001 0.001 0.00 10.3 127 0.001 0.004 0.00 5.8 41 0.001
LOW 0.000 0.00 9.7 155 0.001 0.002 0.00 6.7 64 0.001 0.006 0.00 2.1 8 0.001
MA 0.001 0.00 9.6 139 0.001 0.013 0.06 22.8 520 0.001 0.057 0.28 10.8 118 0.001
MCD 0.000 0.00 9.6 136 0.001 0.001 0.00 6.4 60 0.001 0.002 0.00 2.6 11 0.001
MDLZ 0.000 0.00 7.8 95 0.001 0.001 0.00 20.2 436 0.001 0.005 0.00 9.7 101 0.001
MDT 0.000 0.00 13.7 254 0.001 0.001 0.00 8.8 94 0.001 0.004 0.00 5.3 37 0.001
MET 0.001 0.00 11.1 163 0.001 0.004 0.00 8.6 84 0.001 0.011 0.00 6.4 48 0.001
MMM 0.000 0.00 7.6 81 0.001 0.001 0.00 7.7 91 0.001 0.003 0.00 3.1 13 0.001
MO 0.000 0.00 23.8 723 0.001 0.004 0.00 22.7 518 0.001 0.015 0.02 10.8 117 0.001
MON 0.000 0.00 11.4 181 0.001 0.002 0.00 6.8 66 0.001 0.007 0.00 2.5 9 0.001
MRK 0.000 0.00 12.7 239 0.001 0.001 0.00 6.8 62 0.001 0.004 0.00 3.6 20 0.001
MS 0.001 0.00 31.7 1275 0.001 0.007 0.00 15.7 275 0.001 0.014 0.00 6.9 62 0.001
MSFT 0.000 0.00 13.1 265 0.001 0.001 0.00 6.4 59 0.001 0.005 0.00 2.8 12 0.001
NKE 0.000 0.00 8.5 94 0.001 0.004 0.00 13.0 171 0.001 0.016 0.01 6.2 40 0.001
ORCL 0.000 0.00 8.4 96 0.001 0.001 0.00 4.4 28 0.001 0.005 0.00 3.1 16 0.001
OXY 0.001 0.00 11.8 190 0.001 0.003 0.00 20.2 436 0.001 0.006 0.00 3.8 21 0.001
PCLN 0.001 0.00 10.7 146 0.001 0.003 0.00 5.6 45 0.001 0.014 0.00 2.2 8 0.001
PEP 0.000 0.00 20.8 637 0.001 0.001 0.00 15.0 282 0.001 0.002 0.00 8.6 86 0.001
PFE 0.000 0.00 10.9 167 0.001 0.001 0.00 11.5 169 0.001 0.003 0.00 4.4 28 0.001
PG 0.000 0.00 11.4 200 0.001 0.001 0.00 14.5 274 0.001 0.002 0.00 2.9 13 0.001
PM 0.000 0.00 12.1 204 0.001 0.001 0.00 13.7 238 0.001 0.003 0.00 2.5 11 0.001
PYPL 0.000 0.00 4.0 25 0.001 0.002 0.00 3.2 13 0.001 0.004 0.00 1.7 4 0.011
QCOM 0.000 0.00 11.8 189 0.001 0.002 0.00 5.2 42 0.001 0.006 0.00 2.6 10 0.001
RTN 0.000 0.00 11.0 193 0.001 0.001 0.00 7.3 70 0.001 0.003 0.00 5.4 42 0.001
SBUX 0.000 0.00 9.5 150 0.001 0.003 0.00 19.9 428 0.001 0.010 0.00 8.7 85 0.001
SLB 0.001 0.00 13.0 238 0.001 0.002 0.00 6.8 71 0.001 0.009 0.00 5.6 42 0.001
SO 0.000 0.00 15.5 357 0.001 0.000 0.00 10.2 138 0.001 0.002 0.00 2.3 8 0.001
SPG 0.001 0.00 9.8 132 0.001 0.002 0.00 7.1 62 0.001 0.008 0.00 5.2 31 0.001
T 0.000 0.00 18.0 472 0.001 0.001 0.00 14.3 262 0.001 0.003 0.00 3.6 20 0.001
TGT 0.000 0.00 10.6 181 0.001 0.001 0.00 6.3 50 0.001 0.005 0.00 2.6 10 0.001
TWX 0.000 0.00 18.4 498 0.001 0.003 0.00 22.4 508 0.001 0.013 0.01 10.5 114 0.001
TXN 0.000 0.00 12.3 262 0.001 0.001 0.00 5.4 45 0.001 0.005 0.00 3.1 17 0.001
UNH 0.000 0.00 23.6 747 0.001 0.002 0.00 9.8 112 0.001 0.007 0.00 5.4 36 0.001
UNP 0.000 0.00 9.2 158 0.001 0.003 0.00 16.1 266 0.001 0.011 0.00 7.6 62 0.001
UPS 0.000 0.00 8.2 99 0.001 0.001 0.00 5.9 51 0.001 0.003 0.00 5.8 43 0.001
USB 0.001 0.00 9.9 126 0.001 0.003 0.00 12.4 177 0.001 0.006 0.00 9.7 101 0.001
UTX 0.000 0.00 12.0 226 0.001 0.001 0.00 5.4 43 0.001 0.003 0.00 2.1 8 0.001
V 0.000 0.00 9.6 123 0.001 0.006 0.01 20.8 434 0.001 0.025 0.04 9.9 98 0.001
VZ 0.000 0.00 15.2 370 0.001 0.001 0.00 9.2 109 0.001 0.003 0.00 2.1 7 0.001
WBA 0.000 0.00 14.9 297 0.001 0.001 0.00 6.1 52 0.001 0.006 0.00 2.4 9 0.001
WFC 0.001 0.00 11.9 180 0.001 0.005 0.00 13.3 215 0.001 0.009 0.00 5.5 35 0.001
WMT 0.000 0.00 12.6 216 0.001 0.001 0.00 8.6 109 0.001 0.002 0.00 4.1 24 0.001
XOM 0.000 0.00 14.8 285 0.001 0.001 0.00 15.3 300 0.001 0.002 0.00 2.0 7 0.001

77
Table A.7: Estimated ACF(1), MA(1) and AR(1) coefficients for squared log returns of S&P100 stocks.

Stock d.ACF(1) d.MA(1) d.AR(1) w.ACF(1) w.MA(1) w.AR(1) m.ACF(1) m.MA(1) m.AR(1)

AAPL 0.178 0.184 0.178 -0.003 -0.003 -0.003 -0.012 -0.012 -0.012
ABBV 0.174 0.180 0.175 0.240 0.256 0.241 0.250 0.551 0.266
ABT 0.129 0.131 0.129 -0.002 -0.002 -0.002 -0.021 -0.022 -0.021
ACN 0.070 0.070 0.070 0.277 0.301 0.277 -0.119 -0.130 -0.119
AGN 0.123 0.125 0.123 0.061 0.062 0.062 0.161 0.244 0.163
AIG 0.234 0.140 0.234 0.040 0.040 0.040 0.208 0.227 0.209
ALL 0.255 0.274 0.255 0.188 0.139 0.188 0.132 0.110 0.132
AMGN 0.161 0.165 0.161 0.011 0.011 0.011 0.004 0.004 0.005
AMZN 0.119 0.121 0.119 0.055 0.055 0.055 0.103 0.120 0.104
AXP 0.207 0.216 0.207 0.331 0.240 0.331 0.021 0.021 0.021
BA 0.211 0.220 0.211 0.110 0.111 0.110 -0.003 -0.003 -0.003
BAC 0.319 0.240 0.319 0.152 0.143 0.152 0.478 0.283 0.479
BIIB 0.010 0.010 0.010 0.020 0.020 0.020 0.014 0.013 0.015
BK 0.357 0.418 0.357 0.315 0.237 0.315 0.209 0.218 0.211
BLK 0.216 0.227 0.216 0.135 0.138 0.136 0.080 0.081 0.081
BMY 0.159 0.163 0.159 0.018 0.018 0.018 0.155 0.159 0.172
BRK-B 0.142 0.145 0.142 -0.002 -0.002 -0.002 -0.009 -0.009 -0.009
C 0.295 0.225 0.295 0.021 0.021 0.021 0.003 0.003 0.003
CAT 0.122 0.124 0.122 0.248 0.265 0.250 0.142 0.183 0.143
CELG 0.152 0.155 0.152 0.007 0.007 0.007 -0.043 -0.045 -0.043
CL 0.227 0.240 0.227 -0.003 -0.003 -0.003 -0.015 -0.015 -0.015
CMCSA 0.205 0.213 0.205 0.062 0.063 0.062 -0.029 -0.031 -0.029
COF 0.262 0.282 0.262 0.308 0.254 0.308 0.161 0.170 0.161
COP 0.367 0.436 0.367 0.105 0.106 0.105 0.039 0.046 0.039
COST 0.126 0.128 0.126 0.245 0.261 0.245 -0.045 -0.045 -0.045
CSCO 0.048 0.048 0.048 0.096 0.096 0.096 -0.109 -0.111 -0.110
CVS 0.140 0.143 0.140 0.112 0.113 0.112 0.106 0.113 0.107
CVX 0.285 0.312 0.285 0.116 0.117 0.116 -0.063 -0.068 -0.064
DD 0.209 0.218 0.209 0.238 0.253 0.238 0.250 0.362 0.251
DHR 0.220 0.232 0.220 -0.001 -0.001 -0.001 -0.011 -0.011 -0.011
DIS 0.190 0.196 0.190 0.268 0.290 0.268 0.157 0.161 0.158
DOW 0.189 0.196 0.189 0.215 0.168 0.216 0.142 0.080 0.142
DUK 0.255 0.274 0.255 -0.004 -0.004 -0.004 -0.012 -0.012 -0.012
EMR 0.169 0.174 0.169 -0.001 -0.001 -0.001 -0.022 -0.023 -0.022
EXC 0.198 0.206 0.198 0.365 0.304 0.365 0.042 0.042 0.042
F 0.269 0.292 0.269 0.134 0.117 0.134 0.059 0.061 0.060
FB 0.032 0.032 0.032 0.125 0.127 0.144 0.010 0.007 0.010
FDX 0.172 0.177 0.172 0.137 0.139 0.137 0.094 0.094 0.094
FOX 0.240 0.255 0.240 0.452 0.516 0.452 0.337 0.165 0.339
FOXA 0.277 0.302 0.277 0.473 0.519 0.473 0.212 0.110 0.213
GD 0.189 0.196 0.189 0.284 0.311 0.284 0.239 0.146 0.239
GE 0.280 0.305 0.280 0.295 0.225 0.295 0.473 0.424 0.475
GILD 0.101 0.102 0.101 -0.008 -0.008 -0.008 -0.033 -0.033 -0.033
GM 0.097 0.098 0.098 0.003 0.003 0.003 0.391 0.319 0.394
GOOG 0.068 0.068 0.068 -0.003 -0.003 -0.003 -0.024 -0.024 -0.024
GOOGL 0.065 0.066 0.065 0.001 0.001 0.001 -0.012 -0.012 -0.012
GS 0.245 0.262 0.245 0.379 0.405 0.379 0.393 0.357 0.396
HAL 0.218 0.229 0.218 0.268 0.256 0.268 0.278 0.315 0.279
HD 0.228 0.242 0.228 0.277 0.323 0.277 -0.079 -0.080 -0.079
HON 0.202 0.210 0.202 0.335 0.382 0.336 0.217 0.155 0.218

78
Table A.8: (continued) Estimated ACF(1), MA(1) and AR(1) coefficients for squared log returns of S&P100
stocks.

Stock d.ACF(1) d.MA(1) d.AR(1) w.ACF(1) w.MA(1) w.AR(1) m.ACF(1) m.MA(1) m.AR(1)

IBM 0.164 0.168 0.164 0.258 0.278 0.258 0.072 0.075 0.073
INTC 0.183 0.189 0.183 0.052 0.052 0.052 0.025 0.023 0.025
JNJ 0.212 0.221 0.212 0.382 0.461 0.382 -0.039 -0.049 -0.040
JPM 0.349 0.270 0.349 0.467 0.527 0.467 0.184 0.110 0.185
KHC 0.078 0.079 0.079 -0.051 -0.049 -0.051 -0.147 -0.151 -0.274
KMI 0.415 0.528 0.415 0.008 0.008 0.008 0.103 0.103 0.103
KO 0.287 0.314 0.287 -0.001 -0.001 -0.001 -0.013 -0.013 -0.013
LLY 0.260 0.280 0.260 0.191 0.198 0.191 0.112 0.114 0.113
LMT 0.250 0.268 0.250 0.186 0.192 0.186 0.072 0.072 0.072
LOW 0.131 0.134 0.131 0.321 0.361 0.321 0.058 0.058 0.059
MA 0.124 0.126 0.124 -0.002 -0.002 -0.002 -0.010 -0.010 -0.010
MCD 0.173 0.178 0.173 0.064 0.064 0.064 -0.008 -0.008 -0.008
MDLZ 0.104 0.105 0.104 -0.006 -0.006 -0.006 -0.037 -0.040 -0.037
MDT 0.096 0.097 0.096 0.184 0.190 0.184 0.294 0.368 0.295
MET 0.320 0.244 0.320 0.552 0.704 0.553 0.152 0.109 0.152
MMM 0.159 0.164 0.159 0.209 0.218 0.209 0.102 0.103 0.103
MO 0.142 0.145 0.142 -0.001 -0.001 -0.001 -0.004 -0.004 -0.004
MON 0.118 0.120 0.118 0.169 0.174 0.169 0.141 0.144 0.142
MRK 0.168 0.173 0.168 0.150 0.154 0.150 0.007 0.007 0.008
MS 0.248 0.158 0.248 0.425 0.578 0.425 0.167 0.172 0.167
MSFT 0.149 0.152 0.149 0.111 0.113 0.112 0.010 0.010 0.010
NKE 0.156 0.160 0.156 -0.002 -0.002 -0.002 -0.038 -0.041 -0.038
ORCL 0.158 0.162 0.158 0.084 0.084 0.084 0.094 0.094 0.094
OXY 0.230 0.244 0.230 0.020 0.020 0.020 0.189 0.188 0.190
PCLN 0.077 0.077 0.077 0.080 0.080 0.083 0.172 0.147 0.173
PEP 0.336 0.383 0.336 0.155 0.159 0.156 -0.049 -0.050 -0.049
PFE 0.179 0.184 0.179 0.132 0.135 0.133 0.385 0.342 0.387
PG 0.186 0.192 0.188 0.051 0.052 0.052 0.092 0.092 0.093
PM 0.125 0.127 0.125 0.106 0.107 0.106 0.026 0.027 0.026
PYPL 0.109 0.110 0.110 -0.107 -0.119 -0.143 0.475 0.386 0.538
QCOM 0.106 0.107 0.106 0.057 0.057 0.057 0.051 0.051 0.051
RTN 0.178 0.184 0.178 0.154 0.171 0.154 -0.079 -0.079 -0.079
SBUX 0.136 0.139 0.136 0.025 0.025 0.025 -0.017 -0.017 -0.017
SLB 0.183 0.189 0.183 0.216 0.227 0.216 0.154 0.141 0.155
SO 0.312 0.348 0.312 0.063 0.064 0.063 -0.014 -0.014 -0.014
SPG 0.347 0.402 0.347 0.527 0.531 0.527 0.264 0.218 0.265
T 0.139 0.141 0.139 0.307 0.342 0.307 0.016 0.016 0.016
TGT 0.198 0.206 0.198 0.499 0.443 0.500 0.219 0.186 0.220
TWX 0.174 0.179 0.174 0.003 0.003 0.003 0.047 0.048 0.047
TXN 0.101 0.102 0.101 0.139 0.142 0.139 0.009 0.008 0.009
UNH 0.152 0.156 0.152 0.298 0.269 0.298 0.004 0.002 0.004
UNP 0.125 0.127 0.125 -0.006 -0.006 -0.006 -0.008 -0.008 -0.008
UPS 0.160 0.164 0.160 0.191 0.198 0.192 0.003 0.003 0.003
USB 0.282 0.308 0.282 0.643 0.709 0.643 0.001 0.001 0.001
UTX 0.178 0.184 0.178 0.210 0.220 0.211 0.156 0.160 0.158
V 0.173 0.178 0.173 -0.003 -0.003 -0.003 -0.013 -0.013 -0.013
VZ 0.350 0.406 0.350 0.109 0.111 0.109 -0.063 -0.056 -0.063
WBA 0.098 0.099 0.098 0.110 0.114 0.110 -0.025 -0.025 -0.025
WFC 0.205 0.214 0.205 0.129 0.118 0.129 0.502 0.479 0.502
WMT 0.143 0.146 0.143 0.064 0.064 0.064 -0.114 -0.110 -0.115
XOM 0.323 0.364 0.323 0.145 0.148 0.145 -0.047 -0.065 -0.048
79