Escolar Documentos
Profissional Documentos
Cultura Documentos
A Concise Course
Francis X. Diebold
University of Pennsylvania
Edition 2016
Version 2016.02.15
Time Series Econometrics
Time Series Econometrics
A Concise Course
Francis X. Diebold
Copyright
c 2013-2016, by Francis X. Diebold.
All rights reserved.
This work is freely available for your use, but be warned: it is highly preliminary, significantly
incomplete, and rapidly evolving. It is licensed under the Creative Commons Attribution-
NonCommercial-NoDerivatives 4.0 International License. (Briefly: I retain copyright, but
you can use, copy and distribute non-commercially, so long as you give me attribution and do
not modify. To view a copy of the license, visit http://creativecommons.org/licenses/by-nc-
nd/4.0/.) In return I ask that you please cite the book whenever appropriate, as: Diebold,
F.X. (2016), Time Series Econometrics, Department of Economics, University of Pennsyl-
vania, http://www.ssc.upenn.edu/ fdiebold/Textbooks.html.
To Marc Nerlove,
who taught me time series,
Acknowledgments xviii
Preface xxii
Chapter 1. Introduction 1
Chapter 4. Markovian Structure, Linear Gaussian State Space, and Optimal (Kalman) Filtering 48
Chapter 6. Simulation for Economic Theory, Econometric Theory, Estimation, Inference, and
Optimization 92
Chapter 7. Bayesian Time Series Posterior Analysis by Markov Chain Monte Carlo 115
Appendices 183
Acknowledgments xviii
Preface xxii
Chapter 1. Introduction 1
1.1 Economic Time Series and Their Analysis 1
1.2 A Practical Toolkit 1
1.2.1 Software (and a Tiny bit of Hardware) 1
1.2.2 Data 2
1.2.3 Markup 3
1.2.4 Version Control 3
1.3 Exercises, Problems and Complements 3
1.4 Notes 4
Chapter 4. Markovian Structure, Linear Gaussian State Space, and Optimal (Kalman) Filtering 48
4.1 Markovian Structure 48
4.1.1 The Homogeneous Discrete-State Discrete-Time Markov Process 48
4.1.2 Multi-Step Transitions: Chapman-Kolmogorov 48
4.1.3 Lots of Definitions (and a Key Theorem) 49
4.1.4 A Simple Two-State Example 50
4.1.5 Constructing Markov Processes with Useful Steady-State Distributions 51
4.1.6 Variations and Extensions: Regime-Switching and More 52
4.1.7 Continuous-State Markov Processes 53
4.2 State Space Representations 54
4.2.1 The Basic Framework 54
4.2.2 ARMA Models 56
4.2.3 Linear Regression with Time-Varying Parameters and More 61
4.2.4 Dynamic Factor Models and Cointegration 63
4.2.5 Unobserved-Components Models 64
4.3 The Kalman Filter and Smoother 65
4.3.1 Statement(s) of the Kalman Filter 66
4.3.2 Derivation of the Kalman Filter 67
4.3.3 Calculating P0 70
4.3.4 Predicting yt 70
4.3.5 Steady State and the Innovations Representation 71
4.3.6 Kalman Smoothing 73
4.4 Exercises, Problems and Complements 73
4.5 Notes 79
Chapter 6. Simulation for Economic Theory, Econometric Theory, Estimation, Inference, and
Optimization 92
6.1 Generating U(0,1) Deviates 92
6.2 The Basics: c.d.f. Inversion, Box-Mueller, Simple Accept-Reject 94
6.2.1 Inverse c.d.f. 94
6.2.2 Box-Muller 95
6.2.3 Simple Accept-Reject 95
6.3 Simulating Exact and Approximate Realizations of Time Series Processes 97
6.4 more 97
6.5 Economic Theory by Simulation: Calibration 97
6.6 Econometric Theory by Simulation: Monte Carlo and Variance Reduction 97
6.6.1 Experimental Design 98
6.6.2 Simulation 99
6.6.3 Variance Reduction: Importance Sampling, Antithetics, Control Variates
and Common Random Numbers 100
6.6.4 Response Surfaces 105
6.7 Estimation by Simulation: GMM, SMM and Indirect Inference 106
6.7.1 GMM 106
6.7.2 Simulated Method of Moments (SMM) 106
6.7.3 Indirect Inference 107
6.8 Inference by Simulation: Bootstrap 108
6.8.1 i.i.d. Environments 108
6.8.2 Time-Series Environments 110
6.9 Optimization by Simulation 112
6.9.1 Local 112
6.9.2 Global 112
6.9.3 Is a Local Optimum Global? 113
6.10 Interval and Density Forecasting by Simulation 114
6.11 Exercises, Problems and Complements 114
6.12 Notes 114
Chapter 7. Bayesian Time Series Posterior Analysis by Markov Chain Monte Carlo 115
7.1 Bayesian Basics 115
7.2 Comparative Aspects of Bayesian and Frequentist Paradigms 115
7.3 Markov Chain Monte Carlo 117
7.3.1 Metropolis-Hastings Independence Chain 117
7.3.2 Metropolis-Hastings Random Walk Chain 117
7.3.3 More 117
7.3.4 Gibbs and Metropolis-Within-Gibbs 118
7.4 Conjugate Bayesian Analysis of Linear Regression 119
7.5 Gibbs for Sampling Marginal Posteriors 121
7.6 General State Space: Carter-Kohn Multi-Move Gibbs 121
7.7 Exercises, Problems and Complements 124
7.8 Notes 124
Appendices 183
Francis X. Diebold is Paul F. and Warren S. Miller Professor of Economics, and Professor
of Finance and Statistics, at the University of Pennsylvania, as well as Faculty Research As-
sociate at the National Bureau of Economic Research in Cambridge, Mass. He has published
widely in econometrics, forecasting, finance and macroeconomics, and he has served on the
editorial boards of numerous scholarly journals. He is an elected Fellow of the Econometric
Society, the American Statistical Association, and the International Institute of Forecasters;
the recipient of Sloan, Guggenheim, and Humboldt fellowships; and past President of the
Society for Financial Econometrics. Diebold lectures actively, worldwide, and has received
several prizes for outstanding teaching. He has held visiting appointments in Economics and
Finance at Princeton University, Cambridge University, the University of Chicago, the Lon-
don School of Economics, Johns Hopkins University, and New York University. His research
and teaching are firmly rooted in applications; he has served as an economist under Paul
Volcker and Alan Greenspan at the Board of Governors of the Federal Reserve System in
Washington DC, an Executive Director at Morgan Stanley Investment Management, Co-
Director of the Wharton Financial Institutions Center, and Chairman of the Federal Reserve
Systems Model Validation Council. All his degrees are from the University of Pennsylvania;
he received his B.S. from the Wharton School in 1981 and his economics Ph.D. in in 1986.
He is married with three children and lives in suburban Philadelphia.
About the Cover
The colorful graphic is by Peter Mills and was obtained from Wikimedia Commons. As
noted there, it represents the basins of attraction of the Gaspard-Rice scattering system
projected onto a double impact parameter (whatever that means). I used it mainly because
I like it, but also because its reminiscent of a trending time series.
Hyperlinks to internal items (table of contents, index, footnotes, etc.) appear in red.
Hyperlinks to bibliographic references appear in green.
All media (images, audio, video, ...) were either produced by me or obtained from the
public domain repository at Wikimedia Commons.
List of Figures
Time Series Econometrics (TSE ) provides a modern and concise masters or Ph.D.-level
course in econometric time series. It can be covered realistically in one semester, and I
have used it successfully for many years with first-year Ph.D. students at the University of
Pennsylvania.
The elephant in the room is of course Hamiltons Time Series Analysis. TSE complements
Hamilton in three key ways.
First, TSE is concise rather than exhaustive. (Nevertheless it maintains good breadth
of coverage, treating everything from the classic early framework of Wold, Wiener, and
Kolmogorov, through to cutting-edge Bayesian MCMC analysis of non-linear non-Gaussian
state space models with the particle filter.) Hamiltons book can be used for more extensive
background reading for those topics that overlap.
Second and crucially, however, many of the topics in TSE and Hamilton do not overlap,
as TSE treats a variety of more recently-emphasized ideas. It stresses Markovian structure
throughout, from linear state space, to MCMC, to optimization, to non-linear state space
and particle filtering. Simulation and Bayes feature prominently, as do nonparametrics,
realized volatility, and more.
Finally, TSE is generally e-aware, with numerous hyperlinks to internal items, web sites,
code, research papers, books, databases, blogs, etc.)
Francis X. Diebold
Philadelphia
Introduction
Any series of observations ordered along a single dimension, such as time, may be thought
of as a time series. The emphasis in time series analysis is the study of dependence among
the observations at different points in time.1
Many economic and financial variables, such as prices, sales, stocks, GDP and its com-
ponents, stock returns, interest rates and foreign exchange rates, are observed over time;
in addition to being interested in the interrelationships among such variables, we are also
concerned with relationships among the current and past values of one or more of them,
that is, relationships over time.
At its broadest level, time series analysis provides the language for of stochastic dy-
namics. Hence its the language of even pure dynamic economic theory, quite apart from
empirical analysis. It is, however, a great workhorse of empirical analysis, in pre-theory
mode (non-structurally getting the facts straight before theorizing, always a good idea),
in post-theory mode (structural estimation and inference), and in forecasting (whether
non-structural of structural).
Empirically, the analysis of economic time series is central to a wide range of applica-
tions, including business cycle measurement, financial risk management, policy analysis,
and forecasting. Special features of interest in economic time series include trends and non-
stationarity, seasonality, cycles and persistence, predictability (or lack thereof), structural
change, and nonlinearities such as volatility fluctuations and regime switching.
1 Indeed what distinguishes time series analysis from general multivariate analysis is precisely the temporal
Eviews, however, can sometimes be something of a black box. Hence youll also want
to have available slightly lower-level (mid-level) environments in which you can quickly
program, evaluate and apply new tools and techniques. R is one very powerful and popular
such environment, with special strengths in modern statistical methods and graphical data
analysis.2 R is available for free as part of a massive and highly-successful open-source
project. RStudio provides a fine R working environment, and, like R, its free. A good R
tutorial, first given on Coursera and then moved to YouTube, is here. R-bloggers is a massive
blog with all sorts of information about all things R.
If you need real speed, such as for large simulations, you will likely need a low-level
environment like Fortran or C++. And in the limit (and on the hardware side), if you
need blazing-fast parallel computing for massive simulations etc., graphics cards (graphical
processing units, or GPUs) provide stunning gains, as documented for example in Aldrich
et al. (2011). Actually the real limit is quantum computing, but were not there yet.
For a compendium of econometric and statistical software, see the software links site,
maintained by Marius Ooms at the Econometrics Journal.
1.2.2 Data
Here we mention just a few key must-know sites. Resources for Economists, maintained
by the American Economic Association, is a fine portal to almost anything of interest to
economists. It contains hundreds of links to data sources, journals, professional organiza-
tions, and so on. FRED (Federal Reserve Economic Data) is a tremendously convenient
source for economic data. The National Bureau of Economic Research site has data on U.S.
business cycles, and the Real-Time Data Research Center at the Federal Reserve Bank of
1.2.3 Markup
Markup languages effectively provide typesetting or word processing. HTML is the most
well-known example. Research papers and books are typically written in LaTeX. MiCTeX
is a good and popular flavor of LaTeX, and TeXworks is a good editor designed for LaTeX.
knitr is an R package, but its worth mentioning separately, as it powerfully integrates R
and LaTeX./footnoteYou can access everything in RStudio.
Another markup language worth mentioning is Sphinx, which runs under Python. The
Stachurchski-Sargent e-book Quantitative Economics, which features Python prominently,
is written in Sphinx.
Git and GitHub are useful for open/collaborative development and version control. For
my sorts of small-group projects I find that Dropbox or equivalent keeps me adequately
synchronized, but for serious large-scale development, use of git or equivalent appears crucial.
Consider the following point/counterpoint items. In each case, which do you think
would be more useful for analysis of economic time series? Why?
Continuous / discrete
linear / nonlinear
deterministic / stochastic
univariate / multivariate
time domain / frequency domain
conditional mean / conditional variance
trend / seasonal / cycle / noise
ordered in time / ordered in space
stock / flow
stationary / nonstationary
aggregate / disaggregate
Gaussian / non-Gaussian
1.4 NOTES
The study of time series of, for example, astronomical observations predates recorded
history. Early writers on economic subjects occasionally made explicit reference to
astronomy as the source of their ideas. For example, Cournot stressed that, as in as-
tronomy, it is necessary to recognize secular variation that is independent of periodic
variation. Similarly, Jevons made clear his approach to the study of short-term fluc-
tuations used the methods of astronomy and meteorology. During the 19th century
interest in, and analysis of, social and economic time series evolved into a new field of
study independent of developments in astronomy and meteorology. Time-series anal-
ysis then flourished. Nerlove et al. (1979) provides a brief history of the fields early
development.
For references old and new, see the library of useful books in Appendix B.
Chapter Two
Strict Stationarity
Weak Stationarity
Eyt = , t
Autocovariance Function
(a) symmetric
( ) = ( ),
X
g(z) = ( ) z
=
Autocorrelation Function
( )
( ) =
(0)
iid
Independent (strong) white noise: t (0, 2 )
iid
Gaussian white noise: t N (0, 2 )
E(t ) = 0
var(t ) = 2
Conditional Moment Structure of Strong White Noise
E(t |t1 ) = 0
where
t1 = t1 , t2 , ...
THE WOLD REPRESENTATION 7
(
2 , = 0
( ) =
0, 1
(
1, = 0
( ) =
0, 1
where:
b0 = 1
X
b2i <
i=0
X
yt = B(L)t = bi ti
i=0
t W N (0, 2 )
b0 = 1
X
b2i <
i=0
!
X X X
E(yt ) = E bi ti = bi Eti = bi 0 = 0
i=0 i=0 i=0
!
X X X
var(yt ) = var bi ti = b2i var(ti ) = 2 b2i
i=0 i=0 i=0
(t1 = t1 , t2 , ...)
X
= 0 + b1 t1 + b2 t2 + ... = bi ti
i=1
Autocovariance Structure
THE WOLD REPRESENTATION 9
" ! !#
X X
( ) = E bi ti bh t h
i= h=
X
= 2 bi bi
i=
(where bi 0 if i < 0)
(Obvious truncation)
Unconditional moment structure, conditional moment structure, autocovariance func-
tions, stationarity and invertibility conditions
2.5.1 Extraction
2.5.2 Prediction
yt = t + b1 t1 + ...
Prediction Error
h1
X
eT +h,T = yT +h yT +h,T = bi T +hi
i=0
E(eT +h,T ) = 0
h1
X
var(eT +h,T ) = 2 b2i
i=0
History:
{yt }Tt=1
Immediately,
yT +1,T = yT
yT +2,T = yT +1,T = 2 yT
..
.
yT +h,T = yT +h1,T = h yT
2.6 MULTIVARIATE
E(y1t ) = 1 t
E(y2t ) = 2 t
!
y1t 1
y1 y2 (t, ) = E (y1,t 1 , y2,t 2 )
y2t 2
THE WOLD REPRESENTATION 11
!
11 ( ) 12 ( )
=
21 ( ) 22 ( )
= 0, 1, 2, ...
12 ( ) 6= 12 ( )
12 ( ) = 21 ( )
y1 y2 ( ) = 0y1 y2 ( ), = 0, 1, 2, ...
X
Gy1 y2 (z) = y1 y2 ( ) z
=
Cross Correlations
Ry1 y2 ( ) = Dy1
1 y2
y1 y2 ( ) Dy1
1 y2
, = 0, 1, , 2, ...
!
1 0
D =
0 2
! ! !
y1t B11 (L) B12 (L) 1t
=
y2t B21 (L) B22 (L) 2t
yt = B(L)t = (I + B1 L + B2 L2 + ...)t
(
if t = s
E(t 0s ) =
0 otherwise
X
k Bi k2 <
i=0
12 CHAPTER 2
Autocovariance Structure
X
0
y1 y2 ( ) = Bi Bi
i=
(where Bi 0 if i < 0)
Gy (z) = B(z) B 0 (z 1 )
Wiener-Kolmogorov Prediction
yt = t + B1 t1 + B2 t2 + ...
h1
X
T +h,T = yT +h yT +h,T = Bi T +hi
i=0
E[T +h,T ] = 0
h1
X
E[T +h,T 0T +h,T ] = Bi Bi0
i=0
(L)yt = t
t W N (0, )
where:
(L) = I 1 L ... p Lp
THE WOLD REPRESENTATION 13
! ! ! !
y1t 11 12 y1t1 1t
= +
y2t 21 22 y2t1 2t
! ! !!
1t 0 12 12
WN ,
2t 0 12 22
(L)yt = t
yt = 1 (L)t = (L)t
where:
(L) = I + 1 L + 2 L2 + ...
14 CHAPTER 2
! ! ! !
1 0 11 12 y1t 1t
L =
0 1 21 22 y2t 2t
! ! ! !
1 1
y1t 1t 11 12 1t1
= + 1 1
+ ...
y2t 2t 21 22 2t1
(I 1 L ... p Lp )yt = t
t W N (0, )
Problem:
generally not diagonal, so how to shock j alone?
(I 1 L ... p Lp )yt = t
t W N (0, )
Problem:
generally not diagonal, which makes things tricky, as the variance of a sum of
innovations is therefore not the sum of the variances.
THE WOLD REPRESENTATION 15
Original:
Equivalently:
yt = (I + 1 L + 2 L2 + ...) P vt
= (P + 1 P L + 2 P L2 + ...) vt
Note how the the contemporaneous IRF and VD for h = 1 are driven by the Cholesky
choice of P .
Other choices are possible.
vt W N (0, I)
16 CHAPTER 2
0
yt = C0 vt + C1 vt1 + C2 vt2 + ... (Q : What is C12 ?)
! ! ! ! !
y1t c011 c012 v1t c111 c112 v1t1
= + + ...
y2t c021 c022 v2t c121 c122 v2t1
vt W N (0, I)
! ! ! ! !
1t+2,t c011 c012 v1t+2 c111 c112 v1t+1
= +
2t+2,t c021 c022 v2t+2 c121 c122 v2t+1
(c012 )2 + (c112 )2
V D12 (2) =
(c011 )2 + (c012 )2 + (c111 )2 + (c112 )2
Graphic: IRF Matrix for 4-Variable U.S. Macro VAR
840
700
820
800 650
780 600
760 550
1959 1965 1971 1977 1983 1989 1995 2001 1959 1965 1971 1977 1983 1989 1995 2001
Orthogonalizing/Identifying VARs
Figure 1.1: VARMore Generally
evidence on US data
Structural VARs
Structure:
A0 yt = A1 yt1 + ... + Ap ytp + vt , vt (0, D)
where D is diagonal.
Reduced form:
yt = A1
0 A1 yt1 + ... + A1 1
0 Ap ytp + A0 vt
= 1 yt1 + ... + p ytp + et ,
where et = A1
0 vt .
IRF:
yt = (I + 1 L + 2 L2 + ...) et
= (I + 1 L + 2 L2 + ...) A1
0 vt
= (A1
0 + 1 A1 1 2
0 L + 2 A0 L + ...) vt
18 CHAPTER 2
T | |
1 X
( ) = xt xt+| | , = 0, 1, ..., (T 1)
T t=1
T | |
1 X
( ) = xt xt+| | , = 0, 1 , ..., (T 1)
T | | t=1
d
) N (0, )
T (
asycov(
( ), ( + v)) = 0
PT 2
t=1 et
M SE =
T
THE WOLD REPRESENTATION 19
PT 2
2 t=1 et
R = 1 PT
t=1 (yt y)2
M SE
= 1 1
PT
T t=1 (yt y)2
Still bad:
PT 2
2 t=1 et
s =
T k
PT !
2
2 T t=1 et
s =
T k T
PT 2
2 = 1 P t=1 et / T k
R T
t=1 (yt yt )2 / T 1
s2
= 1 PT
t=1 (yt yt )2 / T 1
Good:
PT !
2
t=1 et
SIC = T ( T )
k
More generally,
2lnL KlnT
SIC = +
T T
m
X
QBP = T 2 ( ) 2 (m)
=1
m
X 1
QLB = T (T + 2) 2 ( )
=1
T
20 CHAPTER 2
1. Ergodicity.
We shall say (loosely speaking) that a time series is ergodic if consistent inference
regarding its stochastic structure can be made on the basis of one realization. While
ergodicity is a deep mathematical property of the distribution function characteriz-
ing the time series in question, its meaning for a stationary time series is essentially
independence of observations far enough apart in time.
Ergodicity refers to consistent moment estimability based only on a single realization,
as opposed to stationarity, which is concerned with the timeconstancy of the prob-
ability structure of a stochastic process. It is therefore nonsensical to pose questions
regarding the ergodicity of nonstationary processes. We stress that ergodicity cannot
be checked, even with a (doubly) infinitely sample path. The intuition is simple:
regardless of whether or not a timeseries is ergodic, sample moments converge to a
random variable. If the series is ergodic, that random variable is in fact a (degenerate)
constant. It is immediately clear, then, that even with an infinitely large sample one
cannot tell whether or not sample moments converge to a constant (fixed in repeated
realizations) or just one particular realization of a random variable (which will change
from realization to realization). To check ergodicity, one must have available an entire
ensemble, which is never the case in practice.
Due to the impossibility of empirically checking ergodicity in observed time series,
attention has focused on the study of specific parameterizations for which ergodic-
ity can be theoretically established. For example, the important LRCSSP, discussed
below, is always ergodic. More generally, we seek sufficient conditions under which
laws of large numbers (LLN) can be shown to hold. For a time series of independent,
identically distributed random variables, Kolmogorovs LLN holds. For dependent,
identically (unconditionally) distributed time series, sufficient conditions for the LLN
are well known. Much recent research examines conditions sufficient for the LLN in
more general situations, such as dependent time series with heterogeneous innovations.
The resulting theories of mixing, martingale difference, and nearepoch dependent se-
quences are discussed in White (1984), Gallant and White (198*), and White (199*),
among many others.
3. Predicting AR processes.
Show the following.
yt = (1 + 2 )yt1 1 2 yt2 + t ,
where ytj = ytj , for j = 0, 1, ..., at time t. Thus for pure autoregressions, the
MMSE prediction is a linear combination of only the p most recently observed
values.
4. Predicting M A process.
If yt is M A(1),
yt = t t1 ,
yt yt1 = t t1 ,
22 CHAPTER 2
6. Prediction-error dynamics.
Consider the general linear process with strong white noise innovations. Show that
both the conditional (with respect to the information set t = {t , t1 , ...}) and
unconditional moments of the Wiener-Kolmogorov h-step-ahead prediction error are
identical.
(a) Obtain the usual quarterly expenditure-side U.S. GDPE from FRB St. Louis,
1960.1-present.
(b) Leaving out the 12 most recent quarters of data, perform a full correlogram
analysis for GDPE logarithmic growth.
(c) Again leaving out the 12 most recent quarters of data, specify, estimate and de-
fend appropriate AR(p) and ARM A(p, q) models for GDPE logarithmic growth.
(d) Using your preferred AR(p) and ARM A(p, q) models for GDPE logarithmic
growth, generate a 12-quarter-ahead linear least-squares path forecast for the
hold-out sample. How do your AR(p) and ARM A(p, q) forecasts compare to
the realized values? Which appears more accurate?
(e) Obtain ADNSS GDPplus logarithmic growth from FRB Philadelphia, read about
it, and repeat everything above.
(f) Contrast the results for GDPE logarithmic growth and GDPplus logarithmic
growth.
(a) Obtain monthly U.S. housing starts and completions data from FRED at FRB
St. Louis, seasonally-adjusted, 1960.1-present. Your two series should be of equal
length.
THE WOLD REPRESENTATION 23
(b) Using only observations {1, ..., T 4}, perform a full correlogram analysis of starts
and completions. Discuss in detail.
(c) Using only observations {1, ..., T 4}, specify and estimate appropriate univari-
ate ARM A(p, q) models for starts and completions, as well as an appropriate
V AR(p). Discuss in detail.
(d) Characterize the Granger-causal structure of your estimated V AR(p). Discuss in
detail.
(e) Characterize the impulse-response structure of your estimated V AR(p) using all
possible Cholesky orderings. Discuss in detail.
(f) Using your preferred ARM A(p, q) models and V AR(p) model, specified and es-
timated using only observations {1, ..., T 4}, generate linear least-squares path
forecasts for the four quarters of hold out data, {T 3, T 2, T 1, T }. How
do your forecasts compare to the realized values? Discuss in detail.
under the usual assumptions. Suppose further that B11 (L) = B21 (L) = 0 and 1t = 2t = t
(with variance 2 ). Discuss the nature of this system. Why might it be useful in eco-
nomics?
2.9 NOTES
Spectral Analysis
Spectral Analysis
X
yt = B(L)t = bi ti
i=0
= 2 B(z)B(z 1 )
Spectrum
Evaluate g(z) on the unit circle, z = ei :
X
g(ei ) = ( ) ei , < <
=
= 2 B(ei ) B(ei )
= 2 | B(ei ) |2
Spectrum
Trigonometric form:
X
g() = ( )ei
=
X
( ) ei + ei
= (0) +
=1
X
= (0) + 2 ( ) cos( )
=1
1
f () = g()
2
1 X
f () = ( )ei ( < < )
2 =
1 1X
= (0) + ( ) cos( )
2 =1
2
B ei B ei
=
2
2
| B ei |2
=
2
SPECTRAL ANALYSIS 27
1. symmetric around = 0
2. real-valued
3. 2-periodic
4. nonnegative
X
g() = ( )ei
=
Z
1
( ) = g()ei d
2
Z
1
( ) = g()ei d
2
Z
= f ()ei d
Hence
Z
(0) = f ()d
T 1
X | |
x ) 0,
T ( 1 ( )
T
=(T 1)
x ) N (0, gx (0))
T (
28 CHAPTER 3
yt = t
t W N (0, 2 )
2
B ei B ei
f () =
2
2
f () =
2
AR(1) Spectral Density
yt = yt1 + t
t W N (0, 2 )
2
f () = B(ei )B(ei )
2
2 1
=
2 (1 ei )(1 ei )
2 1
=
2 1 2 cos() + 2
How does shape depend on ? Where are the peaks?
ARMA(1, 1) Spectral Density
(1 L)yt = (1 L)t
2 1 2 cos() + 2
f () =
2 1 2 cos() + 2
Rational spectral density
Internal peaks? What will it take?
SPECTRAL ANALYSIS 29
3.4 MULTIVARIATE
1 X
= yx ( ) ei , < <
2 =
(Complex-valued)
Co-Spectrum and Quadrature Spectrum
1 X
Cyx () = yx ( ) cos( )
2 =
1 X
Qyx () = yx ( ) sin( )
2 =
Cross Spectrum
fyx () = gayx ()exp(i phyx ()) (generic cross spectrum)
30 CHAPTER 3
1
2
gayx () = [Cyx () + Q2yx ()] 2 (gain)
Qyx ()
phyx () = arctan Cyx () (phase)
ph()
(Phase shift in time units is )
|fyx ()|2
cohyx () = (coherence)
fxx ()fyy ()
Squared correlation decomposed by frequency
Useful Spectral Results for Filter Design and Analysis
fyx ()
B(ei ) =
fxx ()
yt = .5xt1 + t
SPECTRAL ANALYSIS 31
t W N (0, 1)
xt = .9xt1 + t
t W N (0, 1)
Correlation Structure
Autocorrelation and cross-correlation functions are straightforward:
y ( ) = .9| |
x ( ) .9| |
yx ( ) .9| 1|
1
xt = t
1 .9L
1 1 1
= fxx () = i
2 1 .9e 1 .9ei
1 1
=
2 1 2(.9) cos() + (.9)2
1
=
11.37 11.30 cos()
Shape?
Spectral Density of y
yt = 0.5Lxt + t
1
= fyy () =| 0.5ei |2 fxx () +
2
1
= 0.25fxx () +
2
32 CHAPTER 3
0.25 1
= +
11.37 11.30 cos() 2
Shape?
Cross Spectrum
B(L) = .5L
B(ei ) = 0.5ei
fyx () = B(ei )fxx ()
= 0.5ei fxx ()
= (0.5fxx ()) ei
0.5
gyx () = 0.5fxx () = 11.3711.30 cos()
P hyx () =
(In time units, P hyx () = 1, so y leads x by -1)
Coherence
| fyx () |2 2
.25fxx () .25fxx ()
Cohyx () = = =
fxx ()fyy () fxx ()fyy () fyy ()
1 1
.25 2 12(.9) cos()+.92
= 1 1 1
.25 2 12(.9) cos()+.92 + 2
1
=
8.24 + 7.20 cos()
Shape?
yt = xt xt1
= B(ei ) = 1 ei
2
i 1 X ij sin(5/2)
= B1 (e )= e =
5 j=2 5sin(/2)
zt = yt+5 yt5
B2 (ei ) = |2sin(5)|
B1 (ei )B2 (ei ) = sin(5/2) |2sin(5)|
5sin(/2)
(
fx () on [a, b] [b, a]
fy () =
0 otherwise,
where
X
yt = B(L)xt = bj tj
j=
How best to make this filter feasible in practice? What does that mean? Simple
truncation?
Phase shift?
3.6.1 Univariate
T 2 r T
! r T
!
2 X 2 X it 2 X it
I() = yt eit = yt e yt e
T T t=1 T t=1
t=1
2j
Usually examine frequencies j = T , j = 0, 1, 2, ..., T2
Sample Spectral Density
T 1
1 X
f() = ( )ei
2
=(T 1)
2
T
1 X
f() = yt e it
2T
t=1
T
! T
!
1 X it 1 X it
= yt e yt e
2T t=1 2T t=1
1
= I()
4
Properties of the Sample Spectral Density
2j
(Throughout we use j , j = T , j = 0, 1, ..., T2 )
Hence inconsistent
T 1 T 1
1 X 1 2 X
f() = ( )ei = (0) + ( ) cos( )
2 2 2 =1
=(T 1)
T 1
1 X
f () = ( )ei
( )
2
=(T 1)
3.6.2 Multivariate
2. HAC Estimation
4. Sample spectrum.
Generate samples of Gaussian white noise of sizes 32, 64, 128, 256, 512, 1024 and
2056, and for each compute and graph the sample spectral density function at the
usual frequencies. What do your graphs illustrate?
Solution: Assume normality, and then take draws from the process by using a noram
random number generator in conjunction with the Cholesky factorization of the data
covariance matrix. This procedure can be used to estimate the sampling distribution
of the autocorrelations, taken one at a time. One will surely want to downweight
the long-lag autocorrelations before doing the Cholesky factorization, and let this
downweighting adapt to sample size. Assessing sampling uncertainty for the entire
autocorrelation function (e.g., finding a 95% confidence tunnel) appears harder,
due to the correlation between sample autocorrelations, but can perhaps be done
numerically. It appears very difficult to dispense with the normality assumption.
9. Sample coherence.
If a sample coherence is completed directly from the sample spectral density matrix
(without smoothing), it will be 1, by definition. Thus, it is important that the sam-
ple spectrum and cross-spectrum be smoothed prior to construction of a coherence
estimator.
Solution:
2
|fyx ()|
coh() =
fx () fy ()
In unsmoothed sample spectral density analogs,
[yt eit xt e+it ][yt e+it xt eit ]
c
oh() = [xt eit xt e+it ][yt eit yt e+it ]
1.
10. De-meaning.
Consider two forms of a covariance stationary time series: raw and de-meaned.
Contrast their sample spectral density functions at ordinates 2j/T, j = 0, 1, ...,
40 CHAPTER 3
T/2. What do you conclude? Now contrast their sample spectral density functions at
ordinates that are not multiples of 2j/T. Discuss.
Solution: Within the set 2j/T, j = 0, 1, ..., T/2, only the sample spectral density at
frequency 0 is affected by de-meaning. However, de-meaning does affect the sample
spectral density function at all frequencies in [0, ] outside the set 2j/T, j = 0, 1,
..., T/2. See Priestley (1980, p. 417). This result is important for the properties of
time- versus frequency-domain estimators of fractionally-integrated models. Note in
particular that
1 X ij t 2
I(j ) | yt e |
T
so that
1 X 2
I(0) | yt | T y2 ,
T
which approaches infinity with sample size so long as the mean is nonzero. Thus it
makes little sense to use I(0) in estimation, regardless of whether the data have been
demeaned.
the variance of the sample mean of such a time series. If you are very ambitious, you
might want to explore in a Monte Carlo experiment the sampling properties of your
estimator of the standard error vs. the standard estimator of the standard error, for
various population models (e.g., AR(1) for various values of ) and sample sizes. If
you are not feeling so ambitious, at least conjecture upon the outcome of such an
experiment.
15. Coherence.
a. Write out the formula for the coherence between two time series x and y.
b. What is the coherence between the filtered series, (1 - b1 L) xt and (1 - b2 L) yt ?
(Assume that b1 6= b2 .)
c. What happens if b1 = b2 ? Discuss.
Fy () = B(ei ) B (ei )
Solution:
(a) G2 = 1 - e-i2 is monotonically increasing on [0, ]. This is an example of a high
pass filter.
(b) G2 = 1 + e-i2 is monotonically decreasing on [0, ]. This is an example of a low
pass filter.
(c) G2 = (1 - .5 e-12i)2 has peaks at the fundamental seasonal frequency and its
harmonics, as expected. Note that it corresponds to a seasonal autoregression.
(d) G2 = (1 - .5 e-12i)2 has troughs at the fundamental seasonal frequency and its
harmonics, as expected, because it is the inverse of the seasonal filter in (c) above.
42 CHAPTER 3
Thus, the seasonal process associated with the filter in (c) above would be appropri-
ately seasonally adjusted by the present filter, which is its inverse.
18. Filtering
(a) Consider the linear filter B(L) = 1 + L. Suppose that yt = B(L) xt,
where xt WN(0, 2). Compute fy().
(b) Given that the spectral density of white noise is 2/2, discuss how the filtering
theorem may be used to determine the spectrum of any LRCSSP by viewing it as a
linear filter of white noise.
Solution:
(a) fy() = 1 + e-i2 fx()
= 2/2 (1 + e-i)(1 + ei)
= 2/2 (1 + 2 + 2 cos ),
which is immediately recognized as the sdf of an MA(1) process.
(b) All of the LRCSSPs that we have studied are obtained by applying linear filters
to white noise. Thus, the filtering theorem gives their sdfs as
f() = 2/2 B(e-i)2
= 2/2 B(e-i) B(ei)
= 2/2 B(z) B(z-1),
evaluated on |z| = 1, which matches our earlier result.
Solution: The series must be deterministic, because one could design a filter such that
the filtered series has zero spectrum everywhere.
20. Period.
Period is 2/ and is expressed in time/cycle. 1/P, cycles/time. In engineering, time
is often measured in seconds, and 1/P is Hz.
(b) Compute and plot the spectral density f(), for various values of . Does it have
any internal peaks on (0, )? Discuss.
(c) The lowest-frequency internal peak occurs at the so-called fundamental seasonal
frequency. What is it? What is the corresponding period?
(d) The higher-frequency spectral peaks occur at the harmonics of the fundamental
seasonal frequency. What are they? What are the corresponding periods?
Solution:
(a) Monthly, because of the 12-period lag.
(b)
2
f () = (1 + 2 2 cos(12))
2
The sdf has peaks at = 0, /6, 2/6, ..., 5/6, and .
(c) The fundamental frequency is /6, which corresponds to a period of 12 months.
(d) The harmonic frequencies are 2/6, ..., 5/6, and , corresponding to periods of
6 months, 4 months, 3 months, 12/5 months and 2 months, respectively.
(d) The higher-frequency spectral peaks occur at the harmonics of the fundamental
seasonal frequency. What are they? What are the corresponding periods?
Solution: (a) Quarterly, because of the 4-period lag.
(b)
2
f () = (1 + 2 2 cos(4))
2
The sdf has peaks at = 0, /2 and .
(c) The fundamental frequency is /2, which corresponds to a period of 4 quarters.
(d) The only harmonic is , corresponding to a period of 2 quarters.
Let
T
1X
x
= xt .
T t=1
Then
1
PT PT
var(
x) = T2 s=1 t=1 (t s)
1
PT 1 | |
= T =(T 1) (1 T )( ),
Solution:
a. By the law of iterated expectations, we have
0.2
(0) = = 1
1 0.8
for =1,2,. . ..
Therefore
1 X 1 1
f () = ( )ei = (0) =
2 = 2 2
E(x4t ) = E[E(x4t |xt1 , xt2 )] = E[3(0.2 + 0.8x2t1 )2 ] = 3[0.04 + 0.32E(x2t1 ) + 0.64E(x4t1 )].
Because
x2t
x2 ( ) = 0.8x2 ( 1)
2 2
f () = [(1 + 0.8e12i )(1 + 0.8e12i )] = (1 + 1.6cos12 + 0.64)
2 2
SPECTRAL ANALYSIS 47
GLS = (X 0 1 X)1 X 0 1 Y
But asymptotically,
= P 0 DP
so
1 = P 0 D1 P
Thus asymptotically
GLS = (X 0 P 0 D1 P X)1 X 0 P 0 D1 P Y
3.8 NOTES
Harmonic analysis is one of the earliest methods of analyzing time series thought to exhibit
some form of periodicity. In this type of analysis, the time series, or some simple trans-
formation of it, is assumed to be the result of the superposition of sine and cosine waves
of different frequencies. However, since summing a finite number of such strictly periodic
functions always results in a perfectly periodic series, which is seldom observed in practice,
one usually allows for an additive stochastic component, sometimes called noise. Thus, an
observer must confront the problem of searching for hidden periodicities in the data, that
is, the unknown frequencies and amplitudes of sinusoidal fluctuations hidden amidst noise.
An early method for this purpose is periodogram analysis, initially used to analyse sunspot
data, and later to analyse economic time series.
Spectral analysis is a modernized version of periodogram analysis modified to take account
of the stochastic nature of the entire time series, not just the noise component. If it is assumed
that economic time series are fully stochastic, it follows that the older periodogram technique
is inappropriate and that considerable difficulties in the interpretation of the periodograms
of economic series may be encountered.
These notes draw in part on Diebold, Kilian and Nerlove, New Palgrave, ***.
Chapter Four
(Kalman) Filtering
{Xt }, t = 0, 1, 2, . . .
[time (t + 1)]
p11 p12
[time t] p21 p22
P
P
pij 0, j=1 pij = 1
(m)
Let P (m) pij .
STATE SPACE AND THE KALMAN FILTER 49
Chapman-Kolmogorov theorem:
Corollary: P (m) = P m
Two states i and j communicate (or are in the same class) if each is accessible from the
other. We write i j.
A Markov process is irreducible if there exists only one class (i.e., all states communicate).
(n)
State i has period d if pii = 0 n such that n/d 6 Z, and d is the greatest integer with
that property. (That is, a return to state i can only occur in multiples of d steps.) A state
with period 1 is called an aperiodic state.
A Markov process all of whose states are aperiodic is called an aperiodic Markov process.
Still more definitions....
The first-transition probability is the probability that, starting in i, the first transition to
j occurs after n transitions:
(n)
fij = P rob(Xn = j, Xk 6= j, k = 1, ..., (n 1)|X0 = i)
P (n)
Denote the eventual transition probability from i to j by fij (= n=1 fij ).
P = .
Then either:
50 CHAPTER 4
(1) All states are transient or all states are null recurrent
(n)
pij 0 as n i, j. No stationary distribution.
or
(n)
pij j as n i, j. {j , j = 1, 2, 3, ...} is the unique stationary distribution.
is any row of limn P n .
We will verify many of our claims, and we will calculate the steady-state distribution.
pij 0 i, j
2
X 2
X
p1j = 1, p2j = 1
j=1 j=1
Clearly, 1 2, so P is irreducible.
STATE SPACE AND THE KALMAN FILTER 51
4.1.4.4 Periodicity
State 1: d(1) = 2
State 2: d(2) = 2
(1) (n)
f12 = 1, f12 = 0 n > 1 f12 = 1
(1) (n)
f21 = 1, f21 = 0 n > 1 f21 = 1
4.1.4.6 Recurrence
Moreover,
(n)
X
11 = nf11 = 2 < (and similarly 22 = 2 < )
n=1
!
0 1
(.5, .5) = (.5, .5).
1 0
Note that in this example we can not get the stationary probabilities by taking limn P n .
Why?
In section 4.1.4 we considered an example of the form, for a given Markov process, character-
ize its properties. Interestingly, many important tools arise from the reverse consideration,
For a given set of properties, find a Markov process with those properties.
52 CHAPTER 4
st P
yt = cst + st yt1 + t
st Pt
STATE SPACE AND THE KALMAN FILTER 53
yt = cst + st yt1 + t
We call semi-Markov a process with transitions governed by P , such that the state durations
(times between transitions) are themselves random variables. The process is not Markov,
because conditioning not only the current state but also time-to-date in state may be useful
for predicting the future, but there is an embedded Markov process.
Key result: The stationary distribution depends only on P and the expected state dura-
tions. Other aspects of the duration distribution are irrelevant.
Theorem: If {Xt } is a stationary Markov process with transition probabilities pij and sta-
tionary probabilities i , then the reversed process is also Markov with transition probabilities
j
pij = pji .
i
In general, pij 6= pij . In the special situation pij = pij (so that i pij = j pji ), we say that
the process is time-reversible.
t = T t1 + Rt
yt = Zt + t
t N, t N
54 CHAPTER 4
t = Q(t1 , t )
yt = G(t , t )
t D , t D
Still Markovian!
Transition Equation
t = T t1 + R t
mx1 mxm mx1 mxg gx1
t = 1, 2, ..., T
Measurement Equation
yt = Z t + wt + t
1x1 1xm mx1 1xL Lx1 1x1
t = 1, 2, ..., T
(Important) Details
!
t
WN 0, diag( Q , |{z}
h )
t |{z}
gg 11
E(0 t 0 ) = 0mxg
E(0 t ) = 0mx1
STATE SPACE AND THE KALMAN FILTER 55
t = T t1 + R t
mx1 mxm mx1 mxg gx1
yt = Z t + wt + t
1x1 1xm mx1 1xL Lx1 1x1
!
t
WN 0, diag( Q , |{z}
h )
t |{z}
gg 11
t = T t1 + R t
mx1 mxm mx1 mxg gx1
yt = Z t + wt + t
1x1 1xm mx1 1xL Lx1 1x1
t = T B 1 B t1 + R t
mx1 mxm mxm mxm mx1 mxg gx1
yt = Z B 1 B t + wt + t
1x1 1xm mxm mxm mx1 mxL Lx1 1x1
(B t ) = (B T B 1 ) (B t1 ) + (B R) t
mx1 mxm mx1 mxg gx1
yt = (Z B 1 ) (B t ) + wt + t
1x1 1xm mx1 mxL Lx1 1x1
yt = yt1 + t
t W N (0, 2 )
t = t1 + t
yt = t
(T = , R = 1, Z = 1, = 0, Q = 2 , h = 0)
MA(1)
yt = (L)t
t W N (0, 2 )
where
(L) = 1 + 1 L
yt = t + t1
t W N (0, 2 )
STATE SPACE AND THE KALMAN FILTER 57
! ! ! !
1t 0 1 1,t1 1
= + t
2t 0 0 2,t1
yt = (1, 0) t = 1t
!
yt
t =
t
MA(q)
yt = (L)t
t W N (0, 2 )
where
(L) = 1 + 1 L + ... + q Lq
yt = t + 1 t1 + ... + q tq
t W N N (0, 2 )
1t 0 1,t1 1
2t 0 Iq 2,t1 1
..
=
..
.. + . t
.
. . . .
q+1,t 0 00 q+1,t1 q
yt = (1, 0, ..., 0) t = 1t
q tq + . . . + 1 t1 + t yt
..
..
t
.
=
.
q t1 + q1 t
q t1 + q1 t
q t q t
AR(p)
(L)yt = t
t W N (0, 2 )
where
(L) = (1 1 L 2 L2 ... p Lp )
t W N (0, 2 )
1t 1 1,t1 1
2t 2 Ip1 2,t1 0
t = . = + . t
. ..
.. .. ..
.
pt p 00 p,t1 0
yt = (1, 0, ..., 0) t = 1t
(L)yt = (L)t
t W N (0, 2 )
where
(L) = (1 1 L 2 L2 ... p Lp )
(L) = 1 + 1 L + ... + q Lq
t W N (0, 2 )
yt = (1, 0, ..., 0) t
ARMA(p,q) in State Space Form Recursive substitution from the bottom up yields:
1t 1 1,t1 + p 1,tp + t + 1 t1 + . . . + q tq
. .
. .
.
=
.
m1,t m1 1,t1 + m,t1 + m2 t
mt m 1,t1 + m1 t
yt
.
.
=
.
m1 yt1 + m yt2 + m1 t1 + m2 t
m yt1 + m1 t
t = T t1 + R t
mx1 mxm mx1 mxg gx1
60 CHAPTER 4
yt = Z t + Wt + t
N x1 N xm mx1 N xL Lx1 N x1
!
t
WN H )
0, diag( Q , |{z}
t |{z}
gg N N
N -Variable V AR(p)
t W N (0, )
yt = (IN , 0N , ..., 0N ) t
N x1 N xN p N px1
Multivariate ARMA(p,q)
+ t + 1 t1 + ... + q tq
N xN N xN
t W N (0, )
STATE SPACE AND THE KALMAN FILTER 61
Multivariate ARMA(p,q)
1 I
2 IN (m1) 1
t = .
.
t1 +
.. t
. .
N mx1
m 0N xN (m1) m1
yt = (I, 0, ..., 0) t = 1t
where m = max(p, q + 1)
yt = 0 xt + t
t = t1
Measurement:
yt = x0t t + t
(T = I, R = 0, Zt = x0t , = 0, H = 2 )
Note the time-varying system matrix.
Linear Regression with ARMA(p,q) Disturbances
yt = xt + ut
1 1
2 Im1 1
t = .
.
t1 +
.. t
. .
m 00 m1
62 CHAPTER 4
where m = max(p, q + 1)
t = t1 + t
Measurement:
yt = x0t t + t
(T = , R = I, Q = cov(t ), Zt = x0t , = 0, H = 2 )
Gradual evolution of tastes, technologies and institutions
Lucas critique
Stationary or non-stationary
t W N (0, I)
Reduced form:
yt = 1 1 1
0 1 yt1 + ... + 0 p ytp + 0 P t
1 1
1t 0 1 1,t1 0 P
1 1t
2t 0 2 I 2,t1 0
.
. = .. .. + .. .
.
.
. . . .
N t
pt 1
0 p 00 p,t1 0
yt = (IN , 0N , ..., 0N ) t
N x1 N xN p N px1
STATE SPACE AND THE KALMAN FILTER 63
y1t 1 1 1t
. . . .
. = . + . Ft + .
. . . .
yN t N N N t
Ft = Ft1 + t
(L) Ft = (L) t
Dynamic Factor Model Single ARMA(p,q) Factor State vector for F is state vector for
system:
1 1
2 Im1 1
t = .
.
t1 +
.. t
. .
m 00 m1
Dynamic Factor Model Single ARMA(p,q) factor System measurement equation is then:
y1t 1 1 1t
.
. = .. + .. (1, 0, ..., 0) t + ..
. . . .
yN t N N N t
1 1 0 ... 0 1t
. .
t + ..
= . .
. + . .
N N 0 ... 0 N t
! ! !
y1t 1 1t
= t +
y2t 2 2t
t = t1 + t
Common trend t
y1t y2t 1t 2t
Note that 1 2 = 1 2
That is, I(1) I(1) = I(0)
CI(1,0)
xt = xt1 + t
yt = xt + t
! ! !
t 2 0
WN 0,
t 0 2
(t = xt , T = , R = 1, Z = 1, = 0, Q = 2 , H = 2 )
Cycle + Seasonal + Noise
yt = ct + st + t
ct = ct1 + ct
st = st4 + st
ct = c,t1 + ct
STATE SPACE AND THE KALMAN FILTER 65
0 1
0 I3
s,t1 + 0 st
st =
0 0
00 0
Cycle + Seasonal + Noise Stacking transition equations gives the grand transition equa-
tion:
0 1 0 0 0 1 0
0 0 10 0 0 0
! ! !
st s,t1 st
= 0 0 0 1 0 +
0 0
ct c,t1 ct
0 0 0 0 0 0
0 0 0 0 0 1
Finally, the measurement equation is:
!
st
yt = (1, 0, 0, 0, 1) + t
ct
t = T t1 + R t
mx1 mxm mx1 mxg gx1
yt = Z t + Wt + t
N x1 N xm mx1 N xL Lx1 N x1
!
t
WN 0, diag( Q , |{z}
H )
t |{z}
gg N N
E(0 t0 ) = 0mxg
E(0 0t ) = 0mxN
66 CHAPTER 4
a0 = E(0 )
P0 = E(0 a0 ) (0 a0 )0
at/t1 = T at1
Pt/t1 = T Pt1 T 0 + R Q R0
(where Ft = Z Pt/t1 Z 0 + H)
t = 1, ..., T
State-Space in Density Form (Assuming Normality)
t |t1 N (T t1 , RQR0 )
yt |t N (Zt , H)
! !
x xx xy
N , = (x , y )0 =
y yx yy
= x|y N x|y , x|y
x|y = x + xy 1
yy (y y )
x|y = xx xy 1
yy yx
a0 = E0 (0 ) = E (0 )
P0 = var0 ( 0 ) = E [(0 a0 ) (0 a0 )0 ]
1 = T 0 + R1
68 CHAPTER 4
= a1/0 = E0 (1 ) = T E0 (0 ) + RE0 (1 )
= T a0
(subst. a1/0 ) = E0 (1 T a0 ) (1 T a0 )0
(subst. 1 ) = E0 (T (0 a0 ) + R1 ) (T (0 a0 ) + R1 )0
= T P0 T 0 + RQR0
(using E(0 t0 ) = 0 t)
1 |(0 y1 )
or
1 |1
E0 (1 ) = a1/0
E0 (y1 ) = Za1/0 + W1
STATE SPACE AND THE KALMAN FILTER 69
var0 (1 ) = E0 (1 a1/0 ) (1 a1/0 ) = P1/0
var0 (y1 ) = E0 (y1 Za1/0 W1 ) (y1 Za1/0 W1 )0
= E0 (Z(1 a1/0 ) + 1 ) (Z(1 a1/0 ) + 1 )0
= Z P1/0 Z 0 + H (using )
= P1/0 Z 0 (using )
! ! !!
1
0 N a1/0 P1/0 P1/0 Z 0
,
y1 Za1/0 + W1 ZP1/0 ZP1/0 Z 0 + H
(F1 = Z P1/0 Z 0 + H)
4.3.3 Calculating P0
Treatment of Initial Covariance Matrix: P0 = (0) (Covariance stationary case: All eigen-
values of T inside |z| = 1)
t = T t1 + Rt
= P0 = T P0 T 0 + RQR0
= (T T )vec(P0 ) + vec(RQR0 )
4.3.4 Predicting yt
Point prediction:
yt/t1 = Zat/t1 + Wt
Prediction error:
vt = yt (Zat/t1 + Wt )
Density Prediction of yt
yt |t1 N (yt/t1 , Ft )
or equivalently
vt | t1 N (0, Ft )
= Et1 [Z (t at/t1 ) + t ] = 0
STATE SPACE AND THE KALMAN FILTER 71
= ZPt/t1 Z 0 + H Ft
Kt = Pt/t1 Z 0 Ft1
at+1/t = T at/t1 + T Kt vt
t = T t1 + Rt
yt = Zt + t
E(t t0 ) = Q
E(t 0t ) = H
(Nothing new)
= Z at|t1 + vt (measurement)
Note that one-shock state space representation
has time-varying system matrices:
Covariance matrix of vt is Ft
t
at+1|t = T at|t1 + T Kv
yt = Z at|t1 + vt
where
= P Z 0 F 1
K
E(vt vt0 ) = F = Z P Z 0 + H
Prediction yt+1/t is now the projection of yt+1 on infinite past, and one-step prediction
errors vt are now the Wold-Wiener-Kolmogorov innovations
Remarks on the Steady State
2. Because the recursions for Pt|t1 and Kt dont depend on the data, but only on P0 ,
we can calculate arbitrarily close approximations to P and K
by letting the filter run
Then:
where
1
Jt = Pt T 0 Pt+1,t
(e) The expected number of returns to a recurrent state is infinite, and the expected
number of returns to a transient state is finite. That is,
X
n
State j is recurrent Pjj = ,
n=1
X
n
State j is transient Pjj < .
n=1
Solution:
STATE SPACE AND THE KALMAN FILTER 75
Pij 0 i, j
2
X 2
X
P1j = 1, P2j = 1
j=1 j=1
! ! ! !
3 .9 .1 .9 .1 .9 .1 .804 .196
P = =
.3 .7 .3 .7 .3 .7 .588 .412
State 1: d(1) = 1
State 2: d(2) = 1
(1)
f12 = .1
(2)
f12 = .9 .1 = .09
(3)
f12 = .92 .1 = .081
(4)
f12 = .93 .1 = .0729
(1)
f12 = .3
(2)
f21 = .7 .3 = .21
(3)
f21 = .72 .3 = .147
(4)
f21 = .73 .3 = .1029
Eventual:
X (n) .1
f12 = f12 = =1
n=1
1 .9
76 CHAPTER 4
X (n) .3
f21 = f21 = =1
n=1
1 .7
(g) Recurrence:
Because f12 = f21 = 1, both states 1 and 2 are recurrent. In addition,
P (n)
11 = n f11 <
Pn=1
(n)
22 = n=1 n f22 <
States 1 and 2 are therefore positive recurrent and (given their aperiodicity es-
tablished earlier) ergodic.
(h) Stationary distribution
We can iterate on the P matrix to see that:
!
.75 .25
lim P n =
n .75 .25
Hence 1 = 0.75 and 2 = 0.25.
Alternatively, in the two-state case, we can solve analytically for the stationary
probabilities as follows.
!
P11 P12
1 2 = 1 2
P21 P22
1 P11 + (1 1 )P21 = 1
P21
1 =
1 P11 + P21
1 P11
2 =
1 P11 + P21
Thus,
!
n 1 P21 1 P11
lim P =
n (1 P11 + P21 ) P21 1 P11
yt = yt1 + t + t1
t W N (0, 2 )
! ! ! !
1t 1 1,t1 1
= + t
2t 0 0 2,t1
yt = (1, 0) t = 1t
! !
yt1 + t1 + t yt
t = =
t t
7. Identification in UCMs.
Discuss the identifying assumption that UC innovations are orthogonal at all leads and
lags. What convenient mathematical properties does it entail for the observed sum of
the unobserved components? In what ways is it restrictive?
Solution: Orthogonality of component innovations implies that the spectrum of the
observed time series is simply the sum of the component spectra. Moreover, the or-
thogonality facilitates identification. The assumption is rather restrictive, however, in
that it entails no interaction between cyclical and secular economic fluctuations.
78 CHAPTER 4
xt = yt + ut
yt = yt1 + vt .
yt = yt1 + t + t1
and provide expressions for 2 and in terms of the underlying parameters , v2 and
u2 .
Solution:
Box and Jenkins (1976) and Nerlove et al. (1979) show the ARMA result and give the
formula for . That leaves 2 . We will compute var(x) first from the UCM and then
from the ARMA(1,1) reduced form, and equate them.
From the UCM:
v2
var(x) = + u2
1 2
From the reduced form:
(1 + 2 2)
var(x) = 2
1 2
Equating yields
v2 + u2 (1 2 )
2 =
1 + 2 2
extract the seasonal and substractboth methods yield the same answer;
ii) y
s , the estimated seasonal, is less variable than ys , the true seasonal, and yn , the
estimated nonseasonal, is less variable than yn , the true nonseasonal. It is paradoxical
that, by (ii), both estimates are less variable than their true counterparts, yet, by (i),
they still add up to the same observed series as their true counterparts. The paradox is
explained by the fact that, unlike their true counterparts, the estimates ys and yn are
correlated (so the variance of their sum can be more than the sum of their variances).
4.5 NOTES
Chapter Five
and Inference
yt N (, ())
Example: AR(1)
(yt ) = (yt1 ) + t
2
ij () = |ij|
1 2
T /2 1/2 1 0 1
L(y; ) = (2) |()| exp (y ) ()(y )
2
1 1
lnL(y; ) = const ln|()| (y )0 1 () (y )
2 2
T xT matrix () can be very hard to calculate (we need analytic formulas for the auto-
covariances) and invert (numerical instabilities and inaccuracies; slow even if possible)
Prediction-error decomposition and the Kalman filter:
Schweppes prediction-error likelihood decomposition is:
T
Y
L(y1 , . . . , yT ; ) = Lt (yt |yt1 , . . . , y1 ; )
t=1
or:
T
X
ln L(y1 , . . . , yT ; ) = ln Lt (yt |yt1 , . . . , y1 ; )
t=1
MAXIMUM LIKELIHOOD 81
Prediction-error decomposition
In the univariate Gaussian case, the Schweppe decomposition is
T T
T 1X 1 X (yt t )2
ln L = ln 2 ln t2
2 2 t=1 2 t=1 t2
T T
T 1X 1 X vt2
= ln 2 ln Ft
2 2 t=1 2 t=1 Ft
T T
NT 1X 1 X 0 1
= ln 2 ln |Ft | v F vt
2 2 t=1 2 t=1 t t
1. Specify (0)
4. Compute (m+1)
Convergence
Convergence Criteria
k s(m) k small
k (m) (m1) k small
Convergence Rates
k (m+1) k
p such that limm p
k (m) k
= O(1)
Method of Steepest Decent
Use D(m) = I, t(m) = 1, m.
Properties:
1
2 lnL 2 lnL
|
12 (m)
. . . 1 k | (m)
.
D(m) = H 1(m) = .
.
2 lnL 2 lnL
k 1 | (m) . . . k 2 | (m)
MAXIMUM LIKELIHOOD 83
An interesting duality...
Line search: First determine direction, then step
Trust region: First determine step, then direction
Approximate the function locally in a trust region
containing all admissible steps, and then determine direction
Classic example: Levenberg-Marquardt
84 CHAPTER 5
Related R packages:
trust (trust region optimization)
minpack.lm (R interface to Levenberg-Marquardt in MINPACK)
aT,T = aT
PT,T = PT
Smooth:
1
where Jt = Pt T 0 Pt+1,t
Initialize:
Then:
0 0
P(t1,t2),T = Pt1 Jt2 + Jt1 (P(t,t1),T T Pt1 )Jt2
t = T t1 + t
yt = Zt + t
MAXIMUM LIKELIHOOD 85
lnL(y; )
lnL y, {t }Tt=0 ;
Construct lnL(m) (y; ) E lnL y, {t }Tt=0 ;
2. M Step:
(m+1) = argmax lnL(m) (y; }
5.3.1 Not-Quite-Right EM
(But it Captures and Conveys the Intuition)
1. E Step:
Approximate a complete data situation by replacing
{t }Tt=0 with at,T from the Kalman smoother
2. M Step:
Estimate parameters by running regressions:
at,T at1,T
yt at,T
3. If convergence criterion not met, go to 1
Complete-Data Likelihood:
T
Y T
Y
f (y, 0 , {t }Tt=1 ) = fa0 ,P0 (0 ) fT,Q (t |t1 ) fZ,H (yt |t )
t=1 t=1
5.3.2.2 E Step
Construct: lnL(m) (y; ) E lnL y, {t }Tt=0 ;
h i 1 1 h i
E ln L(y, {t }T
t=0 ; ) = const ln |P0 | E (0 a0 )0 P01 (0 a0 )
2 2
T
T 1X
E (t T t1 )0 Q1 (t T t1 )
ln |Q|
2 2 t=1
T
T 1X
E (yt Zt )0 H 1 (yt Zt )
ln |H|
2 2 t=1
T
! T !1
X X
T0 = 0 0
E t t1 E t1 t1
t=1 t=1
t = (t Tt1 )
T
= 1
X
Q t t0 ]
E [
T t=1
T
! T
!1
0
X X
Z 0 = yt E [t ] E [t t0 ]
t=1 t=1
t)
t = (yt Z
MAXIMUM LIKELIHOOD 87
T
= 1
X
H t 0t ]
E [
T t=1
where:
E [t ] = at|T
0
= at|T a0t1|T + P(t,t1)|T
E t t1
T
X
ln L() = ln Lt ()
t=1
2 ln L()
IEX,H (0 ) = E
0
0
T T
2 ln Lt ()
X X
= EH(0 ) = E = EHt (0 )
0
t=1 0 t=1
d
T (M L 0 ) N (0, VEX (0 )) (5.1)
where
1
IEX,H (0 )
VEX (0 ) = VEX,H (0 ) = plimT
T
1
IEX,s (0 )
= VEX,s (0 ) = plimT
T
!1 !1
IEX,H (M L ) IEX,s (M L )
VEX,H (0 ) = VEX,s (0 ) =
T T
!1 !1
IOB,H (M L ) IOB,s (M L )
VOB,H (0 ) = VOB,s (0 ) =
T T
Under possible distributional misspecification (but still assuming correct conditional mean
and variance function specifications),
d
T (M L 0 ) N (0, VEX
m
(0 )) (5.2)
MAXIMUM LIKELIHOOD 89
where:
m
VEX (0 ) = VEX,H (0 )1 VEX,s (0 )VEX,H (0 )1
!1 ! !1
IEX,H (M L ) IEX,s (M L ) IEX,H (M L )
VEX
m
(0 ) =
T T T
!1 ! !1
IOB,H (M L ) IOB,s (M L ) IOB,H (M L )
VOB
m
(0 ) =
T T T
Sandwich Estimator
d
m
T (M L ) N (0, VEX ( )) (5.3)
where:
m
VEX ( ) = VEX,H ( )1 VEX,s ( )VEX,H ( )1
!1 ! !1
IEX,H (M L ) IEX,s (M L ) IEX,H (M L )
VEX
m
( ) =
T T T
90 CHAPTER 5
!1 ! !1
IOB,H (M L ) IOB,s (M L ) IOB,H (M L )
VOB
m
( ) =
T T T
where f (j ; ) is the spectral density and the 22 random variables are independent
across frequencies
2j T
j = , j = 0, 1, ...,
T 2
MGF of any one of the xj s is
1
Mx (t) =
1 2t
Let
f (j ; ) xj
yj = f(j ) =
2
f (j ; ) 1
My (t) = Mx t =
2 1 f (j ; ) t
f (j )
1
g(f(j ); ) = e f (j ;)
f (j ; )
T /2 T /2
X X f(j )
ln L(f; ) = ln f (j ; )
j=0 j=0
f (j ; )
T /2 T /2
X X
ln L(f; ) = ln |F (j ; )| trace F 1 (j ; ) F (j )
j=0 j=0
3. Method of scoring
Slight variation on Newton:
Use E(H (m) )1 rather than H 1(m)
4. Constrained optimization.
5.6 NOTES
Chapter Six
1. Statistically independent
2. Reproducible
3. Non-repeating
4. Quickly-generated
Example:
SIMULATION 93
Figure 6.1: Ripleys Horror Plots of pairs of (Ui+1 , Ui ) for Various Congruential Gener-
ators Modulo 2048 (from Ripley, 1987)
x0 = 1, x1 = 3, x2 = 9, x3 = 11, x4 = 1, x5 = 3, ...
Remarks
xt
1. xt [0, m 1], t. So take xt = m, t
3. The maximum period, m, can be attained using the mixed congruential generator if:
Figure 6.2: Transforming from U(0,1) to f (from Davidson and MacKinnon, 1993)
6.2.2 Box-Muller
x1 x1
y1 y2
f (y1 , y2 ) = f (x1 , x2 )
x2 x2
y1 y2
1 2 2
y2
Box-Muller (Continued) Here we have x1 = e 2 (y1 +y2 ) and x2 = 1
2 arctan y1
x1 x1
y1 y2
1 y12 /2 1 y22 /2
Hence = e e
x2 x2
y1 y2
2 2
Bivariate density is the product of two N (0, 1) densities, so we have generated two inde-
pendent N (0, 1) deviates.
Generating Deviates Derived from N(0,1)
21 = [N (0, 1)]2
Pd
2d = i=1 [Ni (0, 1)]2 , where the Ni (0, 1) are independent
N (, 2 ) = + N (0, 1)
p
td = N (0, 1)/ x2d /d, where N (0, 1) and 2d are independent
Fd1 ,d2 = 2d1 /d1 /2d2 /d2 where 2d1 and 2d2 are independent
Multivariate Normal
N (0, I) (N -dimensional) Just stack N N (0, 1)s
N (, ) (N -dimensional)
Let P P 0 = (P is the Cholesky factor of )
Let X N (0, I). Then P X N (0, )
To sample from N (, ), take + P X
Accept-Reject
(Naive but Revealing Example)
We want to sample x f (x)
Draw:
1 U (, )
96 CHAPTER 6
2 U (0, h)
1. Draw x0 g(x)
f (x0 )
2. Take x = x0 w.p. g(x0 )M ; else go to 1.
(Allows for blanket functions g() more efficient than the uniform)
Note that accept-reject requires that we be able to evaluate f (x) and g(x) for any x.
Mixtures
On any draw i,
x fi (x), w.p. pi
where
0 pi 1, i
N
X
pi = 1
i=1
SIMULATION 97
For example, all of the fi could be uniform, but with different location and scale.
3. Parametric II: Approximate realization via arbitrary startup value with early realiza-
tion discarded
4. Parametric III: Exact realization via drawing startup values from unconditional den-
sity
6.4 MORE
Slice Sampling
Copulas and Sampling From a General Joint Density
Monte Carlo
Key: Solve deterministic problems by simulating stochastic analogs, with the analytical
unknowns reformulated as parameters to be estimated.
Many important discoveries made by Monte Carlo.
Also, numerous mistakes avoided by Monte Carlo!
The pieces:
(I) Experimental Design
(II) Simulation (including variance reduction techniques)
(III) Analysis: Response surfaces (which also reduce variance)
98 CHAPTER 6
Objective
e.g., MSE of an estimator:
2 ] = g(, T )
E[( )
= g(, T )
a
(1 )
N ,
N ormal approximation :
N
" r #!
(1 )
P 1.96
= .95
N
r
0 (1 0 )
2 1.96 = .01
N
If 0 = .05, N = 7299
1
Strategy 2 (Use = 2 = argmax [(1 )]; conservative):
s
1 1
2 2
2 1.96 = .01 N = 38416
N
Strategy 3 (Use =
; the obvious strategy)
6.6.2 Simulation
(II) Simulation
Running example: Monte Carlo integration
R1
Definite integral: = 0 m(x)dx
Key insight:
R1
= 0 m(x)dx = E(m(x))
x U (0, 1)
Notation:
= E[m(x)]
2 = var(m(x))
Direct Simulation:
Arbitrary Function, Uniform Density
Generate N U (0, 1) deviates xi , i = 1, ..., N
Form the N deviates mi = m(xi ), i = 1, ..., N
N
1 X
= mi
N i=1
d
N ( ) N (0, 2 )
Z
= E(m(x)) = m(x)f (x)dx
100 CHAPTER 6
d
N ( ) N (0, 2 )
Z
= E(x) = xf (x)dx
d
N ( ) N (0, 2 )
d
N ( ) N (0, 2 )
SIMULATION 101
Z
f (y) = f (y/x)f (x)dx.
N
1 X
E(y)
b = f (y|xi )
N i=1
So importance sampling replaces a simple average of f (y|xi ) based on initial draws from
f (x) with a weighted average of f (y|xi ) based on initial draws from g(x), where the weights
wi reflect the relative heights of f (xi ) and g(xi ).
Indirect Simulation
Variance-Reduction Techniques
(Swindles)
Importance Sampling to Achieve Variance Reduction
Again we use:
Z
f (x)
= x g(x)dx,
g(x)
and again we arrive at
d
N ( ) N (0, 2 )
xf (x)
Key: Pick g(x) s.t. g(x) has small variance
Importance Sampling Example
Let x N (0, 1), and estimate the mean of I(x > 1.96):
Z
= E(I(x > 1.96)) = P (x > 1.96) = I(x > 1.96) (x) dx
| {z } |{z}
m(x) f (x)
N
X I(xi > 1.96)
= (with variance 2 )
i=1
N
g(x) = N (1.96, 1)
Z
I(x > 1.96) (x)
P (x > 1.96) = g(x) dx
g(x)
2
0.06
2
Antithetic Variates
We average negatively correlated unbiased estimators of (Unbiasedness maintained,
variance reduced)
The key: If x symmetric(, v), then xi are equally likely
e.g., if x U (0, 1), so too is (1 x)
e.g., if x N (0, v), so too is x
Consider for example the case of zero-mean symmetric f (x)
Z
= m(x)f (x)dx
N
1 X
Direct : = mi , ( is based on xi , i = 1, ..., N )
N i=1
1 1
Antithetic : = (x) + (x)
2 2
((x) is based on xi , i = 1, ..., N/2 , and
(x) is based on xi , i = 1, ..., N/2)
Antithetic Variates, Contd
SIMULATION 103
More concisely,
N/2
2 X
= ki (xi )
N i=1
where:
1 1
ki = m(xi ) + m(xi )
2 2
d
N ( ) N (0, 2 )
1 1 1
2 = var (m(x)) + var (m(x)) + cov (m(x), m(x))
4 4 2 | {z }
<0 f or m monotone incr.
Often 2 2
Z Z Z
= m(x)f (x)dx = g(x)f (x)dx + [m(x) g(x)]f (x)dx
Control function g(x) simple enough to integrate analytically and flexible enough to absorb
most of the variation in m(x).
We just find the mean of m(x)g(x), where g(x) has known mean and is highly correlated
with m(x).
Control Variates
Z N
1 X
= g(x)dx + [m(xi ) g(xi )]
N i=1
d
N ( ) N (0, 2 )
Related method (conditioning): Find the mean of E(z|w) rather than the mean of z. The
two are of course the same (the mean conditional mean is the unconditional mean), but
var(E[z|w]) var(z).
Control Variate Example
Z 1
f (x) = ex dx
0
104 CHAPTER 6
N
1 X xi
direct = e
N i=1
N
1 X xi
cv = 1.85 + [e (1 + 1.7xi )]
N i=1
var(direct )
78
var(CV )
Common Random Numbers
We have discussed estimation of a single integral:
Z 1
f1 (x)dx
0
But interest often centers on difference (or ratio) of the two integrals:
Z 1 Z 1
f1 (x)dx f2 (x)dx
0 0
The key: Evaluate each integral using the same random numbers.
Common Random Numbers in Estimator Comparisons
; true parameter 0
Two estimators ,
Compare MSEs: E( 0 )2 , E( 0 )2
Expected difference: E ( 0 )2 ( 0 )2
Estimate:
N
1 X 2
(i 0 )2 (i 0 )
N i=1
Variance of estimate:
1 1 2
var ( 0 )2 + var ( 0 )2 cov ( 0 )2 , ( 0 )2
N N N
Extensions...
= + = g(T ) +
g(T )(1 g(T ))
N 0,
N
Note the heteroskedasticity: variance of changes with T .
Example: Assessing Finite-Sample Test Size
Enforce analytically known structure on
.
Common approach:
p
!
i
12
X
= 0 + T c0 + ci T 2 +
i=1
0 is nominal size, which obtains as T . Second term is the vanishing size distortion.
Response surface regression:
106 CHAPTER 6
1 3
0 ) T 2 , T 1 , T 2 , ...
(
6.7.1 GMM
where
m1 () m
1
m2 () m
2
d() =
..
.
mr () m
r
Model moments for GMM may also be unavailable (i.e., analytically intractable)
MLE efficiency lost may be a small price for SMM tractability gained.
Under correct specification any consistent estimator (e.g., MLE or GMM/SMM) takes
you to the right place asymptotically, and MLE has the extra benefit of efficiency.
Under misspecification, consistency becomes an issue, quite apart from the secondary
issue of efficiency. Best DGP approximation for one purpose may be very different
from best for another.
In contrast, pseudo-MLE ties your hands. Gaussian pseudo-MLE, for example, is con-
sistent for the KLIC-optimal approximation (1-step-ahead mean-squared prediction
error).
The bottom line: under misspecification MLE may not be consistent for what you
want, whereas by construction GMM is consistent for what you want (once you decide
what you want).
where
1 () 1
2 () 2
d() =
..
.
d () d
T
1X
x
T = xt , 2 (x) = E(x )2
T t=1
!
xT )
(
u solves P u =
T
(x) (x)
(1+)/2 , x
xT u
I = [ (1)/2 ]
T u
T T
T
1 X
2 (x) =
T )2
(xt x
T 1 t=1
xT )
(
u
solves P
(x)
u
=
T
!
xT )
(
Root c.d.f. : H(z) = P z
T
(j) T
1. Draw {xt }t=1 with replacement from {xt }Tt=1
(j)
T
x xT
2. Compute
(x)
T
(j)
T
x xT
3. Repeat many times and build up the sampling distribution of
(x)
which is an
T
T
x
approximation to the distribution of
T
(j)
xT x
( T )
H(z) = P
(x)
z
T
(x) (x)
I = [ (1+)/2 , x
xT u (1)/2 ]
T u
T T
!
(j)
xT x
( T ) u ) =
where P
(x)
u
= H(
T
Percentile-t Bootstrap
xT )
(
S=
(x)
T
xT )
(
H(z) = P
(x)
z
T
(j)
xT x
( T )
H(z) =P (x(j) )
z
T
110 CHAPTER 6
(x) (x)
I = [ (1+)/2 , x
xT u (1)/2 ]
T u
T T
(j)
(
x T x
T)
P (j)
u
=
(x )
T
Bootstrap-world root:
d
S D (as T, N )
3. Monte Carlo indicates that bootstrap often does very well in finite samples (not un-
related to 2, but does not require 2)
(j) T
2. Draw {xt }t=1 with replacement from {xt }Tt=1
SIMULATION 111
Issues:
xt = c + xt1 + t , t iid
(j)
2. Draw {t }Tt=1 with replacement from {et }Tt=1
(j)
3. Draw x0 from {xt }Tt=1
(j) (j) + (j) , t = 1, ..., T
4. Generate xt = c + x t1 t
(j) (j)
5. Regress xt (c, xt1 ) to get c(j) and (j) , associated t-statistics, etc.
at+1/t = T at/t1 + T Kt vt
yt = Zat/t1 + vt
1. Estimate system parameters . (We will soon see how to do this.)
2. At the estimated parameter values , run the Kalman filter to get the corresponding 1-step-ahead
prediction errors vt (0, Ft ) and standardize them to u 1/2 vt (0, I), where
t = t
0 = Ft .
t t
112 CHAPTER 6
6.9.1 Local
Using MCMC for MLE (and Other Extremum Estimators)
Chernozukov and Hong show how to compute extremum estimators as mean of pseudo-posterior distri-
butions, which can be simulated by MCMC and estimated at the parametric rate 1/ N , in contrast to the
much slower nonparametric rates achievable (by any method) by the standard posterior mode extremum
estimator.
6.9.2 Global
Summary of Local Optimization:
Simulated Annealing
(Illustrated Here for a Discrete Parameter Space)
Framework:
1. A set , and a real-valued function lnL (satisfying regularity conditions) defined on . Let
be the set of global maxima of lnL
2. (m) , a set N ((m) ) (m) , the set of neighbors of (m)
3. A nonincreasing function, T (m) : N (0, ) (the cooling schedule), where T (m) is the temper-
ature at iteration m
4. An initial guess, (0)
Rk
lnL() is continuous
lnL( ) is the unique finite global max of lnL(),
H( ) exists and is nonsingular
is a local max
lnL()
Develop statistical inference for
Draw {i }Ni=1 uniformly from and form {lnL(i )}i=1
N
1. Convex relaxation.
Our approaches to global optimization involved attacking a nasty objective function with methods
involving clever randomization. Alternatively, one can approximate the nasty objective with a friendly
(convex) objective, which hopefully has the same global optimum. This is called convex relaxation,
and when the two optima coincide we say that the relaxation is tight.
6.12 NOTES
Chapter Seven
Carlo
Overarching Paradigm (T )
T ( ) N (0, )
Shared by classical and Bayesian, but interpretations differ.
Classical: random, fixed
Bayesian: fixed, random
Classical: Characterize the distribution of the random data () conditional on fixed true . Focus on
the likelihood max (M L ) and likelihood curvature in an -neighborhood of the max.
Bayesian: Characterize the distribution of the random conditional on fixed true data (). Examine
the entire likelihood.
Bayesian Computational Mechanics
Data y {y1 , . . . , yT }
Bayes Theorem:
f (y/)f ()
f (/y) =
f (y)
or
f (/y) = c f (y/)f ()
where c1 = f (y/)f ()
R
f (/y) f (y/)f ()
p(/y) L(/y)g()
posterior likelihood prior
Classical Paradigm (T )
1 !
d IEX,H ()
T (M L ) N 0,
T
or more crudely
T (M L ) N (0, )
(Enough said.)
Bayesian Paradigm (T )
116 CHAPTER 7
Or, a posteriori, T ( M L ) N (0, )
Bayesian estimation and model comparison
Estimation:
Model comparison:
p(Mi |y) p(y|Mi ) p(Mi )
=
p(Mj |y) p(y|Mj ) p(Mj )
| {z } | {z } | {z }
posterior odds Bayes f actor prior odds
T
Y
P (y) = P (y1 , ..., yT ) = P (yt |y1:t1 )
t=1
T
X
= lnP (y) = lnP (yt |y1:t1 )
t=1
T
X Z
= ln P (yt |, y1:t1 )P (|y1:t1 ) d
t=1
Bayesian model averaging:
Weight by posterior model probabilities:
As T , the distinction between model averaging and selection vanishes, as one goes to 0 and the
other goes to 1.
If one of the models is true, then both model selection and model averaging are consistent for the true
model. Otherwise theyre consistent for the X-optimal approximation to the truth. Does X = KLIC?
Metropolis-Hastings
We want to draw S values of from p(). Initialize chain at (0) and burn it in.
1. Draw from proposal density q(; (s1) )
2. Calculate the acceptance probability ((s1) , )
3. Set
w.p. ((s1) , ) accept
s
=
(s1) w.p. 1 ((s1) , ) reject
q(; (s1) ) = q ()
Acceptance probability:
" #
p( = ) q = (s1)
((s1) , ) = min , 1
p( = (s1) ) q ( = )
= (s1) +
Acceptance probability reduces to:
p( = )
((s1) , ) = min ,1
p( = (s1) )
7.3.3 More
Burn-in, Sampling, and Dependence
total simulation = burn-in + sampling
Questions:
How to assess convergence to steady state?
In the Markov chain case, why not do something like the following. Whenever time t is a multiple of m,
use a distribution-free non-parametric (randomization) test for equality of distributions to test whether the
unknown distribution f1 of xt , ..., xt(m/2) equals the unknown distribution f2 of xt(m/2)+1 , ..., xtm . If,
for example, we pick m = 20, 000, then whenever time t is a multiple of 20,000 we would test equality of the
distributions of xt , ..., xt10000 and xt10001 , ..., xt20000 . We declare arrival at the steady state when the
null is not rejected. Or something like that.
118 CHAPTER 7
Of course the Markov chain is serially correlated, but who cares, as were only trying to assess equality of
unconditional distributions. That is, randomizations of xt , ..., xt(m/2) and of xt(m/2)+1 , ..., xtm destroy
the serial correlation, but so what?
How to handle dependence in the sampled chain?
Better to run one long chain or many shorter parallel chains?
A Useful Property of Accept-Reject Algorithms
(e.g., Metropolis)
Metropolis requires knowing the density of interest only up to a constant, because the acceptance prob-
ability is governed by the RATIO p( = )/p( = (s1) ). This will turn out to be important for Bayesian
analysis.
Metropolis-Hastings (Discrete)
For desired , we want to find P such that P = . It is sufficient to find P such that i Pij = j Pji .
Suppose weve arrived at zi . Use symmetric, irreducible transition matrix Q = [Qij ] to generate proposals.
That is, draw proposal zj using probabilities in ith row of Q.
Move to zj w.p. ij , where:
j
1, if i 1
ij =
j otherwise
i
Useful if/when conditionals are known and easy to sample from, but joint and marginals are not. (This
happens a lot in Bayesian analysis.)
General Gibbs Sampling
We want to sample from f (z) = f (z1 , z2 , ..., zk )
Initialize (j = 0) using z20 , z30 , ..., zk0
Gibbs iteration j = 1:
a. Draw z11 from f (z1 |z20 , ..., zk0 )
b. Draw z21 from f (z2 |z11 , z30 , ..., zk0 )
c. Draw z31 from f (z3 |z11 , z21 , z40 , ..., zk0 )
...
k. Draw zk1 from f (zk |z11 , ..., zk1
1 )
Repeat j = 2, 3, ....
Again, limj f (z j ) = f (z)
Metropolis Within Gibbs
Gibbs breaks a big draw into lots of little (conditional) steps. If youre lucky, those little steps are simple.
If/when a Gibbs step is difficult, i.e., its not clear how to sample from the relevant conditional, it can be
done by Metropolis.
(Metropolis within Gibbs)
Metropolis is more general but also more tedious, so only use it when you must.
Composition
We may want (x1 , y1 ), ..., (xN , yN ) iid from f (x, y)
Or we may want y1 , ..., yN iid from f (y)
They may be hard to sample from directly.
But sometimes its easy to:
Draw x f (x)
Draw y f (y|x )
Then:
(x1 , y1 ), ..., (xN , yN ) iid f (x, y)
(y1 , ..., yN ) iid f (y)
T
M2
L
2T K
2
Bayesian Inference for /
Prior:
/ 2 N (0 , 0 )
g(/ 2 ) exp(1/2( 0 )0 1
0 ( 0 ))
Likelihood:
1 0
L(/ 2 , y) exp( 2 2 (y X) (y X))
Posterior:
120 CHAPTER 7
p(/ 2 , y) exp(1/2( 0 )0 1 1 0
0 ( 0 ) 2 2 (y X) (y X))
This is the kernel of a normal distribution (*Problem*):
/ 2 , y N (1 , 1 )
where 1
1 = 1 0 +
2 (X 0 X) (1
0 0 +
2 (X 0 X)M L )
1 = (1
0 +
2 (X 0 X))1
(Independent
2 / for completeness.)
of , but write
1 T /2
L 2 /, y 2 exp 21 2 (y X)0 (y X)
(*Problem*: In contrast to L(/ 2 , y) earlier, we dont absorb the ( 2 )T /2 term into the constant of
proportionality. Why?)
Hence (*Problem*):
v1 1
exp
2
p 12 /, y 12 1
2 2
or 12 /, y v21 , 21
v1 = v0 + T
1 = 0 + (y X)0 (y X)
Bayesian Pros Thus Far
1. Feels sensible to focus on p(/y). Classical relative frequency in repeated samples replaced with
subjective degree of belief conditional on the single sample actually obtained
2. Exact finite-sample full-density inference
1. From where does the prior come? How to elicit prior distributions?
2. How to do an objective analysis?
(e.g. what is an uninformative prior? Uniform?)
(Note, however, that priors can be desirable and helpful. See, for example, the cartoon at http:
//fxdiebold.blogspot.com/2014/04/more-from-xkcdcom.html)
3. We still dont have the marginal posteriors that we really want: p(, 2 /y), p(/y).
Problematic in any event!
BAYES 121
t = T t1 + Rt
yt = Zt + t
! !
t iid Q 0
N
t 0 H
T = (01 , . . . , 0T )0 , = (T 0 , R0 , Z 0 , Q0 , H 0 )0
Let
The key: Treat T as a parameter, along with system matrices
Recall the State-Space Model in Density Form
t |t1 N (T t1 , RQR0 )
yt |t N (Zt , H)
Recall the Kalman Filter in Density Form
Initialize at a0 , P0
State prediction:
t |
yt1 N (at/t1 , Pt/t1 )
at/t1 = T at1
Pt/t1 = T Pt1 T 0 + RQR0
State update:
t |
yt N (at , Pt )
at = at/t1 + Kt (yt Zat/t1 )
Pt = Pt/t1 Kt ZPt/t1
Data prediction:
yt |
yt1 N (Zat/t1 , Ft )
where yt = (y10 , ..., yt0 )0
Carter-Kohn Multi-move Gibbs Sampler
Let yT = (y10 , . . . , yT
0 )0
0. Initialize (0)
We have already seen how to do univariate regression. We can easily extend to multivariate regression.
The Gibbs sampler continues to work.
*********************fxd
Multivariate Regression
yit = x0t i + it
iid
(1,t , ...N,t )0 N (0, )
i = 1, ..., N
t = 1, ...T
or
Y = X B + E
|{z} |{z} |{z} |{z}
T N T K KN T N
vec(B)| N (B0 , 0 )
np1 1
p(1 |vec(B)) |1 | 2 exp( tr(1 V1 ))
2
X W 1 (n, V) X 1 W (n, V)
where
np1 1
W (X; n, V) |X| 2 exp tr(XV1 )
2
Bayesian Inference for B|
Prior:
1 0 1
p(vec(B)|) exp tr vec(B B0 ) V0 vec(B B0 )
2
Likelihood:
T
!
1X 0 0 1 0
p(Y, X|B, ) exp (Yt B Xt ) (Yt B Xt )
2 t=1
1 1 0
exp tr (Y XB) (Y XB)
2
1
0 1 X 0 X vec(B B)
exp vec(B B)
2
Posterior:
p(vec(B)|, Y )
1
0 1 X 0 X vec(B B)
+ vec(B B0 )0 V 1 vec(B B0 )
exp vec(B B) 0
2
BAYES 123
vec(B)|, Y N (B1 , V1 )
h i h i1
+ V 1 vec (B0 ) , V1 = 1 X 0 X + V 1
vec(B1 ) = V1 (1 X 0 X)vec(B) 0 0
and B = (X 0 X)1 (X 0 Y )
Bayesian Inference for |B
Prior:
np1 1
p(1 |vec(B)) |1 | 2 exp( tr(1 V1 ))
2
Likelihood:
T 1
p(Y, X|B, ) || 2 exp tr 1 (Y XB)0 (Y XB)
2
Posterior:
T +np1 1
p(1 |vec(B), Y ) |1 | exp tr 1 (Y XB)0 (Y XB) + V1
2
2
This is the kernel of a Wishart distribution:
7.8 NOTES
Chapter Eight
Random Walks
Random walk:
yt = yt1 + t
t W N (0, 2 )
Random walk with drift:
yt = + yt1 + t
t W N (0, 2 )
Properties of the Random Walk
t
X
yt = y0 + i
i=1
(shocks perfectly persistent)
E(yt ) = y0
var(yt ) = t 2
lim var(yt ) =
t
Properties of the Random Walk with Drift
t
X
yt = t + y0 + i
i=1
(shocks again perfectly persistent)
E(yt ) = y0 + t
var(yt ) = t 2
lim var(yt ) =
t
The Random Walk as a Building Block
Generalization of random walk: ARIM A(p, 1, q)
Beveridge-Nelson decomposition:
yt ARIM A(p, 1, q) yt = xt + zt
xt = random walk
zt = covariance stationary
So shocks to ARIM A(p, 1, q) are persistent, but not perfectly so.
126 CHAPTER 8
xt = b + xt1 + t
t W N (0, 2 )
Optimal forecast:
xT +h,T = bh + xT
Forecast does not revert to trend
Forecasting a Linear Trend + Stationary AR(1)
xt = a + bt + yt
yt = yt1 + t
t W N (0, 2 )
Optimal forecast:
xT +h,T = a + b(T + h) + h yT
Forecast reverts to trend
Some Language...
Random walk with drift vs. stat. AR(1) around linear trend
unit root vs. stationary root
Difference stationary vs. trend stationary
Stochastic trend vs. deterministic trend
I(1) vs. I(0)
Stochastic Trend vs. Deterministic Trend
BAYES 127
yt = yt1 + t
d
T (LS 1) DF
Superconsistent
Biased in finite samples (E < (0, 1])
Hurwicz bias Dickey-Fuller bias
Nelson-Kang spurious periodicity
Bigger as T 0 , as 1 , and as intercept, trend included
Non-Gaussian (skewed left)
DF tabulated by Monte Carlo
Studentized Version
1
= r
s PT 1 2
t=2 yt1
(yt ) = (yt1 ) + t
yt = + yt1 + t
where = (1 )
Random walk null vs. mean-reverting alternative
Studentized statistic
Deterministic Trend Under the Alternative
3. Construct yt
: yt = yt1 + et
: yt = c + yt1 + et
: yt = c + t + yt1 + et
i , i , i }N
5. Repeat N times, yielding { i=1
p
X
yt + j ytj = t
j=1
p
X
yt = 1 yt1 + j (ytj+1 ytj ) + t
j=2
Studentized statistic
Allowing for Nonzero Mean Under the Alternative
p
X
(yt ) + j (ytj ) = t
j=1
p
X
yt = + 1 yt1 + j (ytj+1 ytj ) + t
j=2
where = (1 + pj=1 j )
P
Studentized statistic
Allowing for Trend Under the Alternative
p
X
(yt a bt) + j (ytj a b(t j)) = t
j=1
p
X
yt = k1 + k2 t + 1 yt1 + j (ytj+1 ytj ) + t
j=2
p
X p
X
k1 = a(1 + i ) b ii
i=1 i=1
p
X
k2 = b (1 + i )
i=1
Pp
Under the null hypothesis, k1 = b i=1 ii and k2 = 0
Studentized statistic
k1
X
yt = 1 yt1 + j (ytj+1 ytj ) + t
j=2
k1
X
yt = + 1 yt1 + j (ytj+1 ytj ) + t
j=2
k1
X
yt = k1 + k2 t + 1 yt1 + j (ytj+1 ytj ) + t
j=2
k1
X
(yt yt1 ) = (1 1)yt1 + j (ytj+1 ytj ) + t
j=2
yt = x t + t
d
t RV (t diverges)
T
d
RV ( diverges)
T
Cointegration
Consider an N -dimensional variable x:
x CI (d, b) if
1. xi I(d), i = 1, . . . , N
Leading Case
130 CHAPTER 8
x CI(1, 1) if
(1) xi I(1), i = 1, . . . , N
(2) 1 or more linear combinations
zt = 0 xt s.t. zt I(0)
Example
xt = xt1 + vt , vt W N
yt = xt1 + t , t W N, t vt , t,
(yt xt ) = t vt = I(0)
Cointegration and Attractor Sets
xt is N -dimensional but does not wander randomly in RN
0 xt is attracted to an (N R)-dimensional subspace of RN
N : space dimension
R: number of cointegrating relationships
Attractor dimension = N R
(number of underlying unit roots)
(number of common trends)
Example
3-dimensional V AR(p), all variables I(1)
R = 0 no cointegration x wanders throughout R3
R = 1 1 cointegrating vector x attracted to a 2-Dim hyperplane in R3 given by 0 x = 0
R = 2 2 cointegrating vectors x attracted to a 1-Dim hyperplane (line) in R3 given by intersection
of two 2-Dim hyperplanes, 01 x = 0 and 02 x = 0
R = 3 3 cointegrating vectors x attracted to a 0-Dim hyperplane (point) in R3 given by the
intersection of three 2-Dim hyperplanes, 01 x = 0 , 02 x = 0 and 03 x = 0
(Covariance stationary around E(x))
Cointegration Motivation: Dynamic Factor Structure
Factor structure with I(1) factors
(N R) I(1) factors driving N variables
e.g., single-factor model:
y1t 1 1t
. . .
. = . ft + .
. . .
yN t 1 N t
ft = ft1 + t
R = (N 1) cointegrating combs: (y2t y1t ), ..., (yN t y1t )
(N R) = N (N 1) = 1 common trend
Cointegration Motivation: Optimal Forecasting
I(1) variables always co-integrated with their optimal forecasts
Example:
xt = xt1 + t
xt+h|t = xt
xt+h xt+h|t = h
P
i=1 t+i
p1
X
xt = xt1 + Bi xti + ut
i=1
Integration/Cointegration Status
Rank() = 0
0 cointegrating vectors, N underlying unit roots
(all variables appropriately specified in differences)
Rank() = N
N cointegrating vectors, 0 unit roots
(all variables appropriately specified in levels)
Rank() = R (0 < R < N )
R cointegrating vectors, N R unit roots
New and important intermediate case
(not possible in univariate)
xt V ECM xt CI(1, 1)
V ECM Cointegration
We can always write
Pp1
xt = i=1 Bi xti xt1 + ut
But under cointegration, rank() = R < N , so
0
=
N N N R RN
xt = p1 0
P
i=1 Bi xti xt1 + ut
Pp1
= i=1 Bi xti zt1 + ut
V ECM Cointegration
p1
X
xt = Bi xti 0 xt1 + ut
i=1
Premultiply by 0 :
p1
X
0 xt = 0 Bi xti 0 0 xt1 + 0 ut
|{z}
i=1
full rank
So equation balance requires that 0 x t1 be stationary.
132 CHAPTER 8
Stationary-Nonstationary Decomposition
0
(R N ) CI combs
M0
x
= x =
(N N ) (N 1)
com. trends
(N R) N
(Rows of to columns of )
Intuition Transforming the system by yields
p1
X
xt = Bi xti 0 0 xt1 + t
|{z}
i=1
0 by orthogonality
!
0
= 1 1 = 0
1
! !
0 1 1 0
M = =
1 0
! ! !
0 x1t u2t u1t x2t x1t
M = =
x2t x1t x1t
BAYES 133
I(1) : ( ) const
Frequency domain, 0
I(1) : f () 2
I(0) : f () const
ln f () = 0 + 1 ln + t
|{z}
2d
GPH estimator of d: Regress ln f () const, ln
So take d = 1 1 . GPH estimator
2
1. Applied modeling.
134 CHAPTER 8
2. Aggregation.
Granger (1980) shows that aggregation of a very large number of stationary ARMA
time series results, under regularity conditions (generalized in Robinson, 1991), in a
fractionally-integrated process. Thus, aggregation of short-memory processes results
in a long-memory process. Discuss this result in light of theorems on aggregation of
ARMA processes. In particular, recall that aggration of ARMA processes results in
new ARMA processes, generally of higher order than the components.
8.9 NOTES
Chapter Nine
t = T t1 + Rt
yt = Zt + t
Linear / Non-Gaussian
t = T t1 + Rt
yt = Zt + t
136 CHAPTER 9
t D , t D
Non-Linear / Gaussian
t = Q(t1 , t )
yt = G(t , t )
Non-Linear / Gaussian II
(Linear / Gaussian with time-varying system matrices)
t = Tt t1 + Rt t
yt = Zt t + t
t N , t N
Conditionally Gaussian
Whites theorem
Non-Linear / Non-Gaussian
t = Q(t1 , t )
yt = G(t , t )
t D , t D
t = Q(t1 ) + t
yt = G(t ) + t
NON-LINEAR NON-GAUSSIAN 137
t D , t D
t = Qt (t1 , t )
yt = Gt (t , t )
t Dt , t Dt
Vt = Vt1 t0 , t0 lognormal
St = max(Vt D, 0)
!
1t
yt = 0 + I(2t > 0) + (1, 0)
2t
1t N 1 2t N 2 1t 2t
Extensions to:
Richer 1 dynamics (governing the observed y)
138 CHAPTER 9
ht = + ht1 + t (transition)
rt = eht t (measurement)
t N (0, 2 ), t N (0, 1)
Stochastic Volatility Model (Linear/Non-Gaussian Form)
ht = + ht1 + t (transition)
or
ht = + ht1 + t
yt = ht + ut
t N (0, 2 ), ut Du
A signal plus (non-Gaussian) noise
components model for volatility
Realized and Integrated Volatility
IVt = IVt1 + t
RVt = IVt + t
represents the fact that RV is based on less than an infinite sampling frequency.
Microstructure Noise Model
**Hasbrouck
(Non-linear / non-Gaussian)
A Distributional Statement of the Kalman Filter ****
Multivariate Stochastic Volatility with Factor Structure
NON-LINEAR NON-GAUSSIAN 139
***
Approaches to the General Filtering Problem Kitagawa (1987), numerical integration
(linear / non-Gaussian) More recently, Monte Carlo integration
Extended Kalman Filter (Non-Linear / Gaussian)
t = Q(t1 , t )
yt = G(t , t )
t N, t N
where / 12 , y N (1 , 1 )
and 1
2 /y G( v21 , 21 ),
and derive expressions for 1 , 1 , v1 , 1
in terms of 0 , 0 , 0 , x, and y.
Moreover, the key marginal posterior
R
P (/y) = 0 p(, 12 /y)d 2 is multivariate t.
Implement the Bayesian methods via Gibbs sampling.
Linear Quadratic Business Cycle Model
Hansen and Sargent
Linear Gaussian state space system
Parameter-Driven vs. Observation-Driven Models
Parameter-driven: Time-varying parameters measurable w.r.t. latent variables
Observation-driven: Time-varying parameters measurable w.r.t. observable variables
Parameter-driven models are mathematically appealing but hard to estimate. Observation-
driven models are less mathematically appealing but easy to estimate. State-space models,
in general, are parameter-driven. Stochastic volatility models are parameter-driven, while
ARCH models are observation-driven.
Regime Switching
We have emphasized dynamic linear models, which are tremendously important in prac-
tice. Theyre called linear because yt is a simple linear function of past ys or past s.
In some forecasting situations, however, good statistical characterization of dynamics may
require some notion of regime switching, as between good and bad states, which is a
type of nonlinear model.
Models incorporating regime switching have a long tradition in business-cycle analysis,
in which expansion is the good state, and contraction (recession) is the bad state. This
idea is also manifest in the great interest in the popular press, for example, in identifying
and forecasting turning points in economic activity. It is only within a regime-switching
framework that the concept of a turning point has intrinsic meaning; turning points are
naturally and immediately defined as the times separating expansions and contractions.
Observable Regime Indicators
Threshold models are squarely in line with the regime-switching tradition. The follow-
ing threshold model, for example, has three regimes, two thresholds, and a d-period delay
regulating the switches:
(u)
c(u) + (u) yt1 + t , (u) < ytd
(m)
yt = c(m) + (m) yt1 + t , (l) < ytd < (u) .
(l)
(l) (l) (l)
c + yt1 + t , > ytd .
The superscripts indicate upper, middle, and lower regimes, and the regime operative
NON-LINEAR NON-GAUSSIAN 141
at any time t depends on the observable past history of y in particular, on the value of
ytd .
Latent Markovian Regimes
Although observable threshold models are of interest, models with latent states as opposed
to observed states may be more appropriate in many business, economic and financial con-
texts. In such a setup, time-series dynamics are governed by a finite-dimensional parameter
vector that switches (potentially each period) depending upon which of two unobservable
states is realized, with state transitions governed by a first-order Markov process. To make
matters concrete, lets take a simple example. Let {st }Tt=1 be the (latent) sample path of
two-state first-order autoregressive process, taking just the two values 0 or 1, with transition
probability matrix given by
!
p00 1 p00
M = .
1 p11 p11
The ij-th element of M gives the probability of moving from state i (at time t 1) to state
j (at time t). Note that there are only two free parameters, the staying probabilities, p00
and p11 . Let {yt }Tt=1 be the sample path of an observed time series that depends on {st }Tt=1
such that the density of yt conditional upon {st } is
!
2
1 (yt st )
f (yt |st ; ) = exp .
2 2 2
Thus, yt is Gaussian white noise with a potentially switching mean. The two means around
which yt moves are of particular interest and may, for example, correspond to episodes of
differing growth rates (booms and recessions, bull and bear markets, etc.).
Chapter Ten
Volatility Dynamics
10.2 GARCH
Prologue: Reading
Much of what follows draws heavily upon:
Andersen, T.G., Bollerslev, T., Christoffersen, P.F. and Diebold, F.X. (2012), Finan-
cial Risk Measurement for Financial Risk Management, in G. Constantinedes, M.
Harris and Rene Stulz (eds.), Handbook of the Economics of Finance, Elsevier.
Andersen, T.G., Bollerslev, T. and Diebold, F.X. (2010), Parametric and Nonpara-
metric Volatility Measurement, in L.P. Hansen and Y. Ait-Sahalia (eds.), Handbook
of Financial Econometrics. Amsterdam: North-Holland, 67-138.
Andersen, T.G., Bollerslev, T., Christoffersen, P.F., and Diebold, F.X. (2006), Volatil-
ity and Correlation Forecasting, in G. Elliott, C.W.J. Granger, and A. Timmermann
(eds.), Handbook of Economic Forecasting. Amsterdam: North-Holland, 778-878.
Prologue
Aggregation level
Portfolio-level (aggregated, univariate) Risk measurement
Asset-level (disaggregated, multivariate): Risk management
Risk management
Portfolio allocation
Asset pricing
Hedging
Trading
Risk Management
144 CHAPTER 10
r (, )
Portfolio returns:
rp = 0 r (0 , 0 )
minw w0 w
s.t. w0 = p
Importantly, w = f ()
If varies, we have wt = f (t )
Asset Pricing I: Sharpe Ratios
Standard Sharpe:
E(rit rf t )
Conditional Sharpe:
E(rit rf t )
t
Asset Pricing II: CAPM Standard CAPM:
(rit rf t ) = + (rmt rf t )
cov((rit rf t ), (rmt rf t ))
=
var(rmt rf t )
Conditional CAPM:
covt ((rit rf t ), (rmt rf t ))
t =
vart (rmt rf t )
Asset Pricing III: Derivatives
Black-Scholes:
PC = BS(, ...)
Ht = St + ut
cov(Ht , St )
=
var(St )
Dynamic hedging
Ht = t St + ut
covt (Ht , St )
t =
vart (St )
Trading
Some Warm-Up
Unconditional Volatility Measures
Variance: 2 = E(rt )2 (or standard deviation: )
Mean Absolute Deviation: M AD = E|rt |
Interquartile Range: IQR = 75% 25%
148 CHAPTER 10
12
10
8
True Probability, %
0
0 100 200 300 400 500 600 700 800 900 1000
Day Number
p
Z V aRT +1|T
p = PT (rT +1 V aRTp +1|T ) = fT (rT +1 )drT +1
Z p
ESTp+1|T = p 1
V aRT+1|T d
0
t2 = t1
2 2
+ (1 ) rt1
X
t2 = 2
j rt1j
j=0
j = (1 ) j
RM-VaRpT +1|T = T +1 1
p
Rigorous Modeling I
Conditional Univariate Volatility Dynamics from Daily
Data
Conditional Return Distributions
f (rt ) vs. f (rt |t1 )
Key 1: E(rt |t1 )
Are returns conditional mean independent? Arguably yes.
Returns are (arguably) approximately serially uncorrelated, and (arguably) approximately
free of additional non-linear conditional mean dependence.
Conditional Return Distributions, Continued Key 2: var(rt |t1 ) = E((rt )2 |t1 )
Are returns conditional variance independent? No way!
Squared returns serially correlated, often with very slow decay.
The Standard Model
(Linearly Indeterministic Process with iid Innovations)
X
yt = bi ti
i=0
150 CHAPTER 10
X
iid (0, 2 ) b2i < b0 = 1
i=0
X
E(yt+k | t ) = bk+i ti
i=0
rt |t1 N (0, ht )
2
ht = + rt1
E(rt ) = 0
2
E(rt E(rt )) =
(1 )
E(rt |t1 ) = 0
2 2
E([rt E(rt |t1 )] |t1 ) = + rt1
GARCH(1,1) Process
Generalized ARCH
rt | t1 N (0, ht )
VOLATILITY DYNAMICS 151
2
ht = + rt1 + ht1
E(rt ) = 0
2
E(rt E(rt )) =
(1 )
E(rt |t1 ) = 0
2 2
E([rt E(rt | t1 )] | t1 ) = + rt1 + ht1
T T
T p 1 X 1 X rt2
ln L (; rp+1 , . . . , rT ) ln(2) ln ht ()
2 2 t=p+1 2 t=p+1 ht ()
where wj = (1 )j
But in GARCH(1,1) we have:
2
ht = + rt1 + ht1
X
ht = + j1 rtj
2
1
Variance Targeting
Sample unconditional variance:
T
1X 2
2 =
r
T t=1 t
2
= (1 )
rt2 = + ( + )rt1
2
t1 + t ,
where t = rt2 ht .
Variations on the GARCH Theme
Regression with GARCH Disturbances
yt = x0t + t
t |t1 N (0, ht )
VOLATILITY DYNAMICS 153
2
ht = + rt1 + ht1 + 0 zt
is a parameter vector
z is a set of positive exogenous variables.
Asymmetric Response and the Leverage Effect I: TARCH
2
Standard GARCH: ht = + rt1 + ht1
2 2
TARCH:( ht = + rt1 + rt1 Dt1 + ht1
1 if rt < 0
Dt =
0 otherwise
positive return (good news): effect on volatility
r rt1
t1
ln(ht ) = + 1/2 + 1/2 + ln(ht1 )
h ht1
t1
iid
td
zt
std(td )
yt = x0t + t
t |t1 N (0, ht )
yt = x0t + ht + t
t |t1 N (0, ht )
A GARCH(1,1) Example
A GARCH(1,1) Example
A GARCH(1,1) Example
A GARCH(1,1) Example
After Exploring Lots of Possible Extensions...
Rigorous Modeling II
Conditional Univariate Volatility Dynamics from High-
Frequency Data
VOLATILITY DYNAMICS 155
Figure 10.10: Conditional Standard Deviation, History and Forecast, Daily NYSE Returns.
156 CHAPTER 10
Dependent Variable: R
Method: ML - ARCH (Marquardt) - Student's t distribution
Date: 04/10/12 Time: 13:48
Sample (adjusted): 2 3461
Included observations: 3460 after adjustments
Convergence achieved after 19 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(4) + C(5)*RESID(-1)^2 + C(6)*RESID(-1)^2*(RESID(-1)<0)
+ C(7)*GARCH(-1)
Variance Equation
100
50
0
1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
10
10
15
1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
Figure 10.12: S&P500 Daily Returns and Volatilities (Percent). The top panel shows daily S&P500
returns, and the bottom panel shows daily S&P500 realized volatility. We compute realized volatility as the
square root of AvgRV , where AvgRV is the average of five daily RVs each computed from 5-minute squared
returns on a 1-minute grid of S&P500 futures prices.
N ()
X 2
RVt () pt1+j pt1+(j1)
j=1
Z t
RVt () IVt = 2 ( ) d
t1
Microstructure Noise
State space signal extraction
AvgRV
Realized kernel
Many others
RV is Persistent
RV is Reasonably Approximated as Log-Normal
RV is Long-Memory
Exact and Approximate Long Memory
Exact long memory:
(1 L)d RVt = 0 + t
158 CHAPTER 10
100
5 4 3 2 1 0 1 2 3 4 5
Standard Normal Quantiles
QQ plot of Daily Realized Volatility
Quantiles of Input Sample
10
10
5 4 3 2 1 0 1 2 3 4 5
Standard Normal Quantiles
QQ plot of Daily log RVAVR
Quantiles of Input Sample
5
5 4 3 2 1 0 1 2 3 4 5
Standard Normal Quantiles
Figure 10.13: S&P500: QQ Plots for Realized Volatility and Log Realized Volatility. The top
panel plots the quantiles of daily realized volatility against the corresponding normal quantiles. The bottom
panel plots the quantiles of the natural logarithm of daily realized volatility against the corresponding normal
quantiles. We compute realized volatility as the square root of AvgRV , where AvgRV is the average of five
daily RVs each computed from 5-minute squared returns on a 1-minute grid of S&P500 futures prices.
Even better:
RV V aRTp+1|T = RV
d T +1|T 1
p ,
GARCH-RV
t2 = + t1
2
+ RVt1
Realized GARCH
HEAVY
Separating Jumps
0.6
Autocorrelation
0.4
0.2
0.6
Autocorrelation
0.4
0.2
Figure 10.14: S&P500: Sample Autocorrelations of Daily Realized Variance and Daily Re-
turn. The top panel shows realized variance autocorrelations, and the bottom panel shows return autocor-
relations, for displacements from 1 through 250 days. Horizontal lines denote 95% Bartlett bands. Realized
variance is AvgRV , the average of five daily RVs each computed from 5-minute squared returns on a 1-minute
grid of S&P500 futures prices.
where
Jt
X
2
JVt = Jt,j
j=1
Truncation:
N ()
X
T Vt () = p2t1+j I ( pt1+j < T )
j=1
Bi-Power Variation:
N ()1
N () X
BP Vt () = |pt1+j | |pt1+(j+1) |
2 N () 1 j=1
Minimum:
N ()1
N () X 2
M inRVt () = min |pt1+j |, |pt1+(j+1) |
2 N () 1 j=1
Portfolio risk change under a certain scenario involving price movements of set of
assets or asset classes?
How do optimal portfolio shares change if the covariance matrix moves in a certain
way?
Similarly, what about almost any other question in asset pricing, hedging, trading? Almost
all involve correlation.
Basic Framework and Issues I
N 1 return vector Rt
N N covariance matrix t
N (N +1)
2 distinct elements
rt = t zt
zt i.i.d.(0, 1)
Multivariate:
1/2
Rt = t Zt
Zt i.i.d.(0, I)
1/2
where t is a square-root (e.g., Cholesky factor) of t
Ad Hoc Exponential Smoothing (RM)
0
t = t1 + (1 ) Rt1 Rt1
VOLATILITY DYNAMICS 161
Assumes that the dynamics of all the variances and covariances are driven by a single
scalar parameter (identical smoothness)
But covariance matrix forecasts inherit the implausible scaling properties of the uni-
variate RM forecasts and will in general be suboptimal
Multivariate GARCH(1,1)
0
vech (t ) = vech (C) + B vech (t1 ) + A vech (Rt1 Rt1 )
0
vech (t ) = vech (C) + (I) vech (t1 ) + (I) vech (Rt1 Rt1 )
Mirrors RM, but with the important difference that the t forecasts now revert to
= (1 )1 C
Fewer parameters than diagonal, but still O(N )2
(because of C)
Encouraging Parsimony: Covariance Targeting
Recall variance targeting:
T
1X 2
2 =
r , 2 = 2
= take = (1 )
T t=1 t 1
162 CHAPTER 10
t1
Calculate standardized return vector, et = Rt D
PT
Estimate correlation matrix (assumed constant) as 1
T t e0t
t=1 e
t1
Calculate standardized return vector, et = Rt D
DECO
Time-varying correlations assumed identical across all pairs of assets, which implies:
t = (1 t ) I + t J ,
t = + ut + t1
VOLATILITY DYNAMICS 163
0.8
0.6
0.4
0.2
0
75 80 85 90 95 00 05 10
Figure 10.15: Time-Varying International Equity Correlations. The figure shows the estimated
equicorrelations from a DECO model for the aggregate equity index returns for 16 different developed
markets from 1973 through 2009.
Updating rule is naturally given by the average conditional correlation of the stan-
dardized returns,
PN PN
2 i=1 j>i ei,t ej,t
ut = PN
N i=1 e2i,t
DECO Example
Factor Structure
Rt = Ft + t
where
1/2
Ft = F t Zt
Zt i.i.d.(0, I)
t i.i.d.(0, )
= t = F t 0 + t
Rt = ft + t
164 CHAPTER 10
where
ft = f t zt
zt i.i.d.(0, 1)
t i.i.d.(0, 2 )
= t = f2 t 0 +
2
it = f2 t 2i + i
2
2
ijt = f2 t i j
Rigorous Modeling IV
Conditional Asset-Level (Multivariate) Volatility Dynam-
ics from High-Frequency Data
Realized Covariance
N ()
X
0
RCovt () Rt1+j, Rt1+j,
j=1
Z t
RCovt () ICovt = ( ) d
t1
15
10
5
Quantiles of Input Sample
10
15
15 10 5 0 5 10 15
Standard Normal Quantiles
Figure 10.16: QQ Plot of S&P500 Returns. We show quantiles of daily S&P500 returns from January
2, 1990 to December 31, 2010, against the corresponding quantiles from a standard normal distribution.
S = RCovt () + (1 ) t
t
t1 )
vech (t ) = vech (C) + B vech (t1 ) + A vech(
Rigorous Modeling V
Distributions
Modeling Entire Return Distributions:
Returns are not Unconditionally Gaussian
Modeling Entire Return Distributions:
Returns are Often not Conditionally Gaussian
Modeling Entire Return Distributions: Issues
Gaussian conditional VaR is somewhat better but left tail remains bad
166 CHAPTER 10
2
Quantiles of Input Sample
5
5 4 3 2 1 0 1 2 3 4 5
Standard Normal Quantiles
Gaussian conditional expected shortfall, which integrates over the left tail, would be
terrible
So we want more accurate assessment of things like V aRTp +1|T than those obtained
under Gaussian assumptions
Doing so for all values of p [0, 1] requires estimating the entire conditional return
distribution
More generally, best-practice risk measurement is about tracking the entire condi-
tional return distribution
rT +1 = T +1/T T +1
T +1 iid(0, 1)
Multiply T +1 draws by T +1/T (fixed across draws, from a GARCH model) to build up
the conditional density of rT +1 .
2
Quantiles of Input Sample
5
5 4 3 2 1 0 1 2 3 4 5
Standard Normal Quantiles
rT +1 = T +1 T +1
T +1 iid(0, 1)
rT +1 = T +1 T +1
T +1 iid(0, 1)
Multiply T +1 draws from N (0, 1) by T +1 draws (from a simulated RV model fit to log
realized standard deviation) to build up the conditional density of rT +1 .
168 CHAPTER 10
rT +1 = T +1/T T +1
T +1 iidN (0, 1)
But in the conditionally non-Gaussian case there is potential loss of generality in writing:
rT +1 = T +1/T T +1
T +1 iid(0, 1),
because there may be time variation in conditional moments other than T +1/T , and using
T +1 iid(0, 1) assumes that away
Multivariate Return Distributions
If reliable realized covariances are available, one could do a multivariate analog of the
earlier lognormal/normal mixture model. But the literature thus far has focused primarily
on conditional distributions for daily data.
Return version:
1/2
Zt = t Rt , Zt i.i.d., Et1 (Zt ) = 0 V art1 (Zt ) = I
where Dt denotes the diagonal matrix of conditional standard deviations for each of the
assets, and t refers to the potentially time-varying conditional correlation matrix.
Leading Examples
Multivariate normal:
Multivariate t:
(d+N )/2
e0 1 et
f (et ) = C (d, t ) 1+ t t
(d 2)
VOLATILITY DYNAMICS 169
Multivariate asymmetric t:
r
C d, t K d+N 1 (et )
0
d + (et ) 1 exp (et )
0 1
0
t t t
2
f (et ) = (d+N )
(d+N ) r
1 (e )
0
(et ) t 2 2
1+ t d + (et ) 1 (et )
0 1
0
d t t
More flexible than symmetric t but requires estimation of N asymmetry parameters si-
multaneously with the other parameters, which is challenging in high dimensions.
Copula methods sometimes provide a simpler two-step approach.
Copula Methods
Sklars Theorem:
N
N G(F1 (e1 ), ..., FN (eN )) Y
f (e) = = g (u) fi (ei )
e1 ...eN i=1
T
X T X
X N
= log L = log g(ut ) + log fi (ei,t )
t=1 t=1 i=1
Standard Copulas
Normal:
1 1 1 1
g(ut ; t ) = |t | 2 0 1
exp (ut ) (t I) (ut )
2
where 1 (ut ) refers to the N 1 vector of standard inverse univariate normals, and the
correlation matrix t pertains to the N 1 vector et with typical element,
Often does not allow for sufficient dependence between tail events.
t copula
Asymmetric t copula
Asymmetric Tail Correlations
Multivariate Distribution Simulation (General Case)
Simulate using:
1/2
Rt = t Zt
Zt i.i.d.(0, I)
Empirical
Gaussian
0.4 DECO
0.3
Threshold Correlation
0.2
0.1
0
1 0.5 0 0.5 1
Standard Deviation
Figure 10.19: Average Threshold Correlations for Sixteen Developed Equity Markets. The
solid line shows the average empirical threshold correlation for GARCH residuals across sixteen developed
equity markets. The dashed line shows the threshold correlations implied by a multivariate standard normal
distribution with constant correlation. The line with square markers shows the threshold correlations from a
DECO model estimated on the GARCH residuals from the 16 equity markets. The figure is based on weekly
returns from 1973 to 2009.
1/2 ZF,t
Ft = F,t
Ft + t
Rt =
Rigorous Modeling VI
Risk, Return and Macroeconomic Fundamentals
We Want to Understand the Financial / Real Connections
Statistical vs. scientific models
Returns Fundamentals
rf
Disconnect?
excess volatility, disconnect, conundrum, ...
r , r , f , f
Links are complex:
r r f f
Volatilities as intermediaries?
For Example...
VOLATILITY DYNAMICS 171
Table 10.1: Stock Return Volatility During Recessions. Aggregate stock-return volatility is quar-
terly realized standard deviation based on daily return data. Firm-level stock-return volatility is the cross-
sectional inter-quartile range of quarterly returns.
Table 10.2: Real Growth Volatility During Recessions. Aggregate real-growth volatility is quarterly
conditional standard deviation. Firm-level real-growth volatility is the cross-sectional inter-quartile range of
quarterly real sales growth.
f r
Return Volatility is Higher in Recessions
Schwerts (1989) failure: Very hard to link market risk to expected fundamentals (lever-
age, corporate profitability, etc.).
Actually a great success:
Key observation of robustly higher return volatility in recessions!
Earlier: Officer (1973)
Later: Hamilton and Lin (1996), Bloom et al. (2009)
Extends to business cycle effects in credit spreads via the Merton model
f r , Continued
Bloom et al. (2009) Results
f f
Fundamental Volatility is Higher in Recessions
More Bloom, Floetotto and Jaimovich (2009) Results
f r
Return Vol is Positively Related to Fundamental Vol
Follows immediately from relationships already documented
Moreover, direct explorations provide direct evidence:
Engle et al. (2006) time series
172 CHAPTER 10
90
80
70
P (% per year)
60
50
40
30
20
10
0
0 10 20 30 40 50 60 70 80 90 100
Real Stock Return Volatility and Real PCE Growth Volatility, 1983-2002
Rt = 0 + 1 Xt + 2 t + t
t2 = + rt1
2 2
+ t1
Reliable risk measurement requires conditional models that allow for time-varying
volatility.
Risk measurement may be done using univariate volatility models. Many important
recent developments.
Other tasks require multivariate models. Many important recent developments, espe-
cially for N large. Factor structure is often useful.
****************************
Models for non-negative variables (from Minchul)
Introduction Motivation: Why do we need dynamic models for positive values?
Alternative model
Harvey (2013)
Creal, Koopman, and Lucas (2013)
Autoregressive Gamma Processes
Autoregressive Gamma Processes (ARG): Definition Definition: Yt follows the autoregressive gamma
process if
Measurement:
Yt Zt Gamma( + Zt , c)
Transition:
Zt Yt1 P oisson(Yt1 )
176 CHAPTER 10
Transition:
Zt Yt1 P oisson(Yt1 )
Conditional moments:
where = c > 0.
When < 1,
The stationary ARG process features marginal over-dispersion.
The process may feature either conditional under- or over-dispersion, depending on the value of Yt1 .
Remark: ACD (autoregressive conditional duration) model assumes the path-indepednet over-dispersion.
Continuous time limit of ARG(1) The stationary ARG process is a discretized version of the CIR process.
p
dYt = a(b Yt )dt + Yt dWt
where
a = log
c
b=
1
2 log
2 = c
1
Corr(Yt , Yth |, c, ) = E (h )
The autocorrelation function features hyperbolic decay when the distribution assigns sufficiently large
probabilities to values close to one.
Figures 1
Transition:
Zt Yt1 P oisson(Yt1 )
Yt : Interquote durations of the Dayton Mining stock traded on the Toronto Stock Exchange in October
1998.
Estimation based on QMLE
Measurement
yt p(yt |ht , xt ; )
Transition
ht Gamma( + zt , c)
zt P oisson(ht1 )
Transition
ht Gamma( + zt , c)
zt P oisson(ht1 )
yt Gamma (, ht exp(xt ))
Transition
ht Gamma( + zt , c)
zt P oisson(ht1 )
Transition
ht Gamma( + zt , c)
zt P oisson(ht1 )
Recent extension: ARG-zero processes 1 Monfort, Pegoraro, Renne, Roussellet (2014) extend ARG process
to take account for zero-lower bound spells,
Recent extension: ARG-zero processes 2 Monfort, Pegoraro, Renne, Roussellet (2014) extend ARG process
to take account for zero-lower bound spells,
BAYES 179
Two modifications
= 0: As 0, Gamma(, c) converges to dirac delta function.
is related with a probability of escaping from the zero lower bound.
Characterization Probability density for ARG-zero is
X
p(Yt |Yt1 ; , , c) = g(Yt , Yt1 , , , c, z)1{Yt >0} + exp( Yt1 )1{Yt =0}
z=1
Conditional moments
E[Yt |Yt1 ] = c + Yt1
and
V (Yt |Yt1 ) = 2c2 + 2cYt1
where = c.
Figure: ARG-zero
Autoregressive conditional duration model (ACD) Yt follows the autoregressive conditional duration
model if
yt = t et , E[et ] = 1
t = w + t1 + yt1
Because of its multiplicative form, it is classified as the multiplicative error model (MEM).
Conditional moments
E[yt |y1:t1 ] = t
V (yt |y1:t1 ) = k0 2t
V (yt |y1:t1 )
= k0
E[yt |y1:t1 ]2
Dynamic conditional score (DCS) model Dynamic conditional score model (or Generalized Autoregressive
Score model) is a general class of observation-driven model.
Convenient and general modelling strategy. I will describe it within the MEM class of model.
DCS Example: ACD 1 Recall
yt = t et , E[et ] = 1
t = w + t1 + yt1
BAYES 181
Instead, we apply DCS principle: Give me conditional likelihood and time-varying parameters, then I
will give you a law motion
yt = t et , et Gamma(, 1/)
DCS Example: ACD 2
yt = t et , et Gamma(, 1/)
t = w + t1 + st1
t = w + t1 + yt1
which is ACD.
However, a law of motion will be different with different choice of distribution General-
ized Gamma, Log-Logistic, Burr, Pareto, and many other distributions
10.6 NOTES
Chapter Eleven
High Dimensionality
1. xxx
11.2 NOTES
Appendices
183
Appendix A
Nonparametrics
f smooth in [x0 h, x0 + h]
Z x0 +h
1 1
f (x0 ) f (u)du = P (x [x0 h, x0 + h])
2h x0 h 2h
1 #xi [x0 h, x0 + h]
fh (x0 ) =
2h N
N
1 X1 x0 xi
= I 1
N h i=1 2 h
Rosenblatt estimator
bandwidth: h
Standard conditions:
R
K(u)du = 1
K(u) = K(u)
Common Kernel Choices
u2
Standard normal: K(u) = 1 e 2
2
N
1 X x0 xi
fh (x0 ) = K
N h i=1 h
Rosenblatt-Parzen estimator
(So h 0 = bias 0)
(So N h = var 0)
Thus,
) p
h0
= fh (x0 ) f (x0 )
Nh
186 APPENDIX A
d
N h(fh (x0 ) f (x0 )) D
h = N 1/5
Corresponding Optimal Convergence Rate
Recall:
d
N h fh (x0 ) f (x0 ) D
h N 1/5
p d
N 4/5 fh (x0 ) f (x0 ) D
h = 1.06N 1/5
So use:
= 1.06
h N 1/5
=
h N 1/5
A.2 MULTIVARIATE
N
1 X x0 xi
fh (x0 ) = K
N h i=1 h
where Kh () = h1 K( h )
or Kh () = h1 K(h1 )
Multivariate Version (d-Dimensional)
Precisely follows equation (A.2):
N
1 X
fH (x0 ) = KH (x0 xi ),
N i=1
N
1 X x0 xi
= fh (x0 ) = K
N hd i=1 h
) p
h0
= fh (x0 ) f (x0 )
N hd
d
N hd
fh (x0 ) f (x0 ) D
1
h N d+4
q
d
d
N 1 d+4 fh (x0 ) f (x0 ) D
Curse of dimensionality
Silvermans Rule
1
d+4
= 4 1
h N d+4
d+2
where
d
2 1X 2
=
d i=1 i
M (x) (M (x + h2 ) M (x h2 ))
(x) = = lim
xj h0 h
f (u), u = y M (x)
Conditional Variance
Z
f (y, x)
var(y|x) = V (x) = y2 dy M (x)2
f (x)
Hazard Function
f (t)
(t) =
1 F (t)
2 lim (x + h2 ) (x h2 )
C(x) = (x) = ( )M (x) =
xj xj 2 h0 h
Using multivariate kernel density estimates and manipulating gives the Nadaraya-Watson
estimator:
N
" #
x0 xi
h (x0 ) =
X K h
M PN x0 xi
yi
i=1 i=1 K h
h 0, N h =
d
h (x0 ) M (x0 )) N (0, V )
N hd (M
M k (x0 ) = 1 P
k in(x0 ) yi (Locally Constant, uniform weighting)
k , k k (x0 ) P M (x0 )
0 M
N
k (M k (x0 ) M (x0 )) d D
Equivalent to Nadaraya-Watson kernel regression with:
1
K(u) = 2 I(|u| 1) (uniform)
and h = R(k) (distance from x0 to k th nearest neighbor)
Variable bandwidth!
yt = g(xt ) + t
Computation of g(x ) :
0 < 1
kT = int( T )
(xt , x )
vt (xt , x , xkT ) = C (x , x
k )
T
(
(1 u3 )3 f or u < 1
C(u) =
0 otherwise
M (x0 ) =
j=0 j j (x0 )
(the j are orthogonal basis functions)
M J (x0 ) = J j j (x0 )
j=0
P
J , J
N 0 M J (x0 ) M (x0 )
hjt = (jo + R
i=1 ij xit ), j = 1, ..., S (N euron j)
e.g. () can be logistic (regression), 0-1 (classification)
Ot = (0 + Sj=1 j hjt )
e.g. () can be the identity function
Universal Approximator: S , S
N 0 O(x0 ) p O(x0 )
to get:
x xN
N (x0 ) = (N 1)hd fN 1 (x0 )M
N 1 (x0 ) + YN K( 0 )
M x x
h
(N 1)hd fN 1 (x0 ) + K( 0 N )
h
Ot = (0 + Sj=1 j hjt )
Back substitution:
Ot = g(xt , xt1 , ..., x1 ; )
A.8 NOTES
Appendix B
Ait-Sahalia, Y. and Hansen, L.P. eds. (2010), Handbook of Financial Econometrics. Ams-
terdam: North-Holland.
Beran, J., Feng, Y., Ghosh, S. and Kulik, R. (2013), Long-Memory Processes: Probabilistic
Properties and Statistical Methods, Springer.
Box, G.E.P. and Jenkins, G.W. (1970), Time Series Analysis, Forecasting and Control,
Prentice-Hall.
Douc, R., Moulines, E. and Stoffer, D.S. (2014), Nonlinear Time Series: Theory, Methods,
and Applications with R Examples, Chapman and Hall.
Durbin, J. and Koopman, S.J. (2001), Time Series Analysis by State Space Methods, Oxford
University Press.
Efron, B. and Tibshirani, R.J. (1993), An Introduction to the Bootstrap, Chapman and
Hall.
Elliott, G., Granger, C.W.J. and Timmermann, A., eds. (2006), Handbook of Economic
Forecasting, Volume 1, North-Holland.
Elliott, G., Granger, C.W.J. and Timmermann, A., eds. (2013), Handbook of Economic
Forecasting, Volume 2, North-Holland.
Engle, R.F. and McFadden, D., eds. (1995), Handbook of Econometrics, Volume 4, North-
Holland.
194 APPENDIX B
Geweke, J., Koop, G. and van Dijk, H., eds. (2011), The Oxford Handbook of Bayesian
Econometrics, Oxford University Press.
Granger, C.W.J. and Newbold, P. (1977), Forecasting Economic Time Series, Academic
Press.
Hammersley, J.M. and Handscomb, D.C. (1964), Monte Carlo Methods, Chapman and Hall.
Hansen, L.P. and Sargent, T.J. (2013), Recursive Models of Dynamic Linear Economies,
Princeton University Press.
Harvey, A.C. (1989), Forecasting, Structural Time Series Models and the Kalman Filter,
Cambridge University Press.
Harvey, A.C. (2013), Dynamic Models for Volatility and Heavy Tails, Cambridge University
Press.
Hastie, T., Tibshirani, R. and Friedman, J. (2001), The Elements of Statistical Learning:
Data Mining, Inference and Prediction, Springer-Verlag.
Kim, C.-J. and Nelson, C.R. (1999), State-Space Models with Regime Switching, MIT Press.
Nerlove, M., Grether, D.M., Carvalho, J.L. (1979), Analysis of Economic Time Series: A
Synthesis, Academic Press.
Silverman, B.W. (1986), Density Estimation for Statistics and Data Analysis, Chapman and
Hall.
USEFUL BOOKS 195
Whittle, P. (1963), Prediction and Regulation by Linear Least Squares Methods, University
of Minnesota Press.
Aldrich, E.M., F. Fernndez-Villaverde, A.R. Gallant, and J.F. Rubio-Ramrez (2011), Tap-
ping the Supercomputer Under Your Desk: Solving Dynamic Equilibrium Models with
Graphics Processors, Journal of Economic Dynamics and Control , 35, 386393.
Aruoba, S.B., F.X. Diebold, J. Nalewaik, F. Schorfheide, and D. Song (2013), Improving
GDP Measurement: A Measurement Error Perspective, Working Paper, University of
Maryland, Federal Reserve Board, and University of Pennsylvania.
Nerlove, M., D.M. Grether, and J.L. Carvalho (1979), Analysis of Economic Time Series:
A Synthesis. New York: Academic Press. Second Edition.
Yu, Yaming and Xiao-Li Meng (2010), To Center or Not to Center: That is Not the Question
An Ancillarity-Sufficiency Interweaving Strategy (ASIS) for Boosting MCMC Efficiency,
Manuscript, Harvard University.