Professional Documents
Culture Documents
Textbook Multivariate Modelling of Non Stationary Economic Time Series 2Nd Edition John Hunter Ebook All Chapter PDF
Textbook Multivariate Modelling of Non Stationary Economic Time Series 2Nd Edition John Hunter Ebook All Chapter PDF
https://textbookfull.com/product/multivariate-time-series-
analysis-in-climate-and-environmental-research-1st-edition-
zhihua-zhang-auth/
https://textbookfull.com/product/sas-for-forecasting-time-series-
john-c-brocklebank/
https://textbookfull.com/product/spectral-analysis-for-
univariate-time-series-2nd-edition-donald-b-percival/
Time Series Analysis: Nonstationary and Noninvertible
Distribution Theory 2nd Edition Katsuto Tanaka
https://textbookfull.com/product/time-series-analysis-
nonstationary-and-noninvertible-distribution-theory-2nd-edition-
katsuto-tanaka/
https://textbookfull.com/product/time-series-analysis-1st-
edition-palma/
https://textbookfull.com/product/practical-time-series-
forecasting-with-r-a-hands-on-guide-2nd-edition-galit-shmueli/
https://textbookfull.com/product/world-literature-non-
synchronism-and-the-politics-of-time-filippo-menozzi/
https://textbookfull.com/product/design-and-analysis-of-time-
series-experiments-1st-edition-bartos/
PALGRAVE TEXTS IN ECONOMETRICS
Multivariate Modelling of
Non-Stationary Economic
Time Series
John Hunter
Simon P. Burke
Alessandra Canepa
Palgrave Texts in Econometrics
This is a series of themed books in econometrics, where the subject is interpreted as
including theoretical developments, applied econometrics and more specialized fields
of application, for example financial econometrics, the econometrics of panel data
sets, forecasting and so on. Each book in the series is directed to particular aspects of
the underlying and unifying theme.
Multivariate
Modelling of
Non-Stationary
Economic Time Series
John Hunter Simon P. Burke
Department of Economics and Finance Department of Economics
Brunel University University of Reading
Uxbridge, UK Reading, UK
Alessandra Canepa
Department of Economics and Finance
Brunel University
Uxbridge, UK
The authors felt that, rather than produce a second edition of their book
Modelling Non-Stationary Economic Time Series, there was a need to
focus on the multivariate nature of the problem and handle some recent
developments in time series analysis. They would like to thank the two
referees who reviewed an earlier version and provided extensive comments
on the proposed revisions. The present structure is heavily influenced by
these reviews and, subject to the limitations of space, the authors believe
that they have taken account of them.
The main focus of this book is still on cointegration, but new material
has been added, including a chapter contributed by Alessandra Canepa
on small-sample correction, chapters on the impact of volatility on
cointegration analysis, and a chapter on the impact of different orders
of integration. Conventional time series models are discussed in the
context of stationary data prior to any discussion of cointegration. There
is extended discussion of impulse responses and forecasting in both the
stationary and non-stationary contexts.
Models with expectations are considered along with alternative meth-
ods such as singular spectrum analysis, the Kalman filter and structural
time series. The latter are all considered in relation to cointegration. Single
equation methods are purely used to develop a topic or as examples of the
notion of cointegration or non-stationarity.
v
vi Preface
1 Introduction 1
References 16
3 Cointegration 77
3.1 Cointegration of the VMA, VAR and VECM 79
3.2 The Smith-McMillan-Yoo Form 86
3.3 Johansen’s VAR Representation of
Cointegration 104
3.4 Cointegration with Intercept and Trend 116
3.5 Alternative Representations of the
Cointegrating VAR, VMA and VARMA 121
ix
x Contents
References 491
Bibliography 493
Index 495
1
Introduction
when the residual for a regression of the stock or the log of the stock
against the futures contract has a unit coefficient and the equation a white
noise error. If the current futures contract is the prediction drawn from the
random walk model, it follows that the difference in the series is equivalent
to the differential between the spot and the futures. In the terminology
of financial markets, the differential between the spot and the futures is
called the basis. In practice the basis follows a stochastic process related
to the information driving the aggregate expectations encapsulated in the
futures contract and as a result this differential may not usually be white
noise. If agents are rational then the differential may often be defined by
a moving average process related to the extent to which agents are able to
respond to new information. This leads to the conclusion that the basis,
although it may not be a white noise process, ought to be stationary.
This gives rise to a natural example from the well-defined literature
on spot-futures efficiency where a series can in one instance be rendered
stationary by subtracting a lag or in the other instance by subtracting a
series with similar time series properties. This is more obvious when one
considers a stock price and a derivative, as the derivative is a function of
the previous pricing decisions related to the spot. The underlying stability
of these relations follows from arbitrage as an equilibrating mechanism
in financial markets. This occurs where the matched nature of futures
contracts through a market implies that disequilibrium is not seen via a
mismatch of quantities, but from mispricing. In product markets both
prices and quantities might be out of line and the more appropriate
concept is one of disequilibrium (Leijonhufvud 1968).
Another decomposition of the random walk via repeated substitution
is as a sum of past errors subject to some initial value. If the original series
were a random walk then the stock evolution is the sum of all past errors
or shocks that have impacted on the market price. This summation of
past errors that are themselves stationary gives rise to a cumulative process
that is non-stationary. Sir David Hendry in his inaugural lecture to the
London School of Economics (LSE) generated such a stochastic trend
using cumulative rainfall. This was then used as a counter-example to
explain price movements by this artificial variable to show the potential for
spurious relations in economics (Hendry 1980). Hendry’s concern related
to the extent to which very poor relations might provide a good fitting
4 Multivariate Modelling of Non-Stationary Economic Time Series
model, but with very poor properties for the error and out of sample
performance. The extreme model had a Durbin Watson (DW) statistic of
0.1 and a fit that deteriorates markedly as the sample evolves, suggesting
that persistent autocorrelation and poor out of sample performance are
indicative of models that make no sense. It is clearly shown by Hendry that
this arises, because the actual economic data are trended as is the newly
constructed cumulative rainfall series, though the simple diagnostics that
follow from the regression results imply the latter data has nothing to
do with the behaviour of the underlying price series. The cumulated
stochastic variable is termed a “stochastic trend”; by comparison, when a
sequence of unit values are summed without error, then this sum gives rise
to a “deterministic trend”. The stochastic trend arises when, in cumulating
the series, there is a random component; in the simplest case this is
explained by a random walk with drift.
Most primary economic and financial data can for all intents and
purposes be described by random walks. A significant literature exists
for the exchange rate; and the idea that financial markets follow random
walks goes back to Bachelier (1900). This suggestion is less obvious when
it comes to economic variables like national income and consumption,
inflation or product prices. However, the Itô process relating to finance
data and well known derivative pricing models (Black-Scholes) are gen-
erally known to follow a diffusion process; and there is a literature on
economic pricing models based on diffusion processes. Hence, when the
change in a price series follows a diffusion process, then the log of the price
follows a random walk. Of course economic prices and quantities are not
limited to following strict random walks, but this means that the error
that drives the stochastic trend has correlated stationary components.
Hence, economic series follow more sophisticated random walks that
include other autoregressive and moving average error terms. However,
the primary driver is the stochastic trend with the other components
being of a smaller order. The order of the other terms is however of
the same dimension as the time series restricted to a random walk when
differenced.
At the core of econometrics is the capacity and quality of measurement
and the existence of the measure. Conventional statistical measurement
and inference considered the behaviour of processes that are associated
1 Introduction 5
with distributions that are generally viewed as being fixed across the
sample. When economists started to apply statistical measurement to
economic data then the notion that the data were identically and inde-
pendently distributed (IID) had to be rejected. Regression was used to
measure the heterogeneity by estimating a mean conditional on exogenous
information, while the assumption that the data are IID was used to give
structure to the unknown error in the model. Essentially some form of
least squares regression became the method generally applied to explain
economic phenomena, but in the early and quite mature literature it is
hard to find reference to the notion of non-stationarity. One exception is
the book by Herman Wold with Lars Jureen entitled Demand Analysis,
which does consider the behaviour of stationary economic time series.
However, Wold and Jureen (1953) analysed data for the interwar years,
a period when price series fell in relative terms and growth of output
stagnated. Hence, any question of how demand models might be derived
when time series are non-stationary was, apart from some exceptions,
ignored. It is of interest to note that James Tobin in a study of the demand
for food estimated both a logarithmic inverse demand curve and, in an
attempt to remove serial correlation, the same relationship in differences.
The latter equation relates to the Rotterdam models that are seen as
approximations of known demand equations (Theil 1965; Barten 1969;
Deaton 1975). In the early 1970s, Box and Jenkins wrote a book that
became highly influential in the statistical analysis of time series data.
Box and Jenkins set out a methodology for building time series models
that firstly considers the appropriate degree of differencing required to
render a series stationary, then discusses the type of alternative models-
autoregressive (AR), moving average (MA) or autoregressive moving
average (ARMA)-that might be used to estimate univariate time series,
before considering the method of estimation. Fama (1970) suggests that
the observation that financial time series follow random walks is consistent
with the idea that markets are efficient. The random walk model implies
that financial time series are non-stationary and, following Box and
Jenkins, need to be differenced to make them stationary. The difference
in the log of the share price yields a continuously compounded return
and, when the financial market is efficient, returns are not supposed to be
predictable.
6 Multivariate Modelling of Non-Stationary Economic Time Series
The structure of time series models pre-dates Box and Jenkins. Yule
(1927) first estimated AR processes, and Kolmogorov in 1929 consid-
ered the behaviour of sums of independent random variables (see the
discussion in Wold and Jureen 1953). In the regression context, Sargan
(1964) used time series methods to derive what would subsequently be
termed an “error correction model” of UK wage inflation. The Sargan
model became the basis of most of the UK wage equations used in the
large macroeconomic models in the late 1970s and 1980s (Wallis et al.
1984). For demand analysis, approximation rather than non-stationarity
was behind the use of differencing, while developments in economic
theory related to the structure of demand equations was more interested in
issues of aggregation as compared with the possible time series structure of
the data (Deaton and Muellbauer 1980). Differencing time series became
common practice in modelling univariate series, and this approach was
also applied in finance where it was common to consider returns of
different assets rather than share prices. The market model relates the
return on a share to the return on the market. There was now a discrepancy
between the methods applied in statistics and finance to time series data,
and the approach predominantly used by economists.
However, the first oil shock precipitated a crisis in macroeconomic
model building as most of the world’s large models were unable to resolve
many of the problems that ensued from this shock. Forecasts and policy
simulations that provided governments with predictions of the future and
a practical tool for understanding the impact of policy on the economy
were unable to explain what had happened and what policies might
remedy the situation (Wallis et al. 1984). The UK Treasury’s inability to
forecast the balance of payments position led to a developed economy
being forced to borrow from the IMF. This is a remedy that would not
have been sought had reasonable estimates been available of the true
payments position as the debt was never realized and the facility was
redeemed by the end of the year. The whole approach to the econometric
modelling of economic time series was in doubt.
Econometric modelling was criticized on three grounds: the specifi-
cation of the models used, their forecast accuracy and their existence.
The model building approach adopted at the LSE was built on a time
series methodology developed by Sargan (1964). The Sargan approach
1 Introduction 7
1
Muellbauer (1983) shows that a random walk model of consumption with innovations in income
and interest rates can be nested in the ADL framework due to Davidson et al. (1978). However, the
tests used do not take account of the underlying series being non-stationary.
2
As will be discovered in the first part of Chap. 9, stationarity is overly strong. In addition the types
of model used by Sargent are excessively restrictive (Hunter 1989).
3
It should be noted that the impulse response function solved from the VAR is not unique (Lippi
and Reichlin 1994) and any findings on causality depend on the structure of the model estimated
(Hendry and Ericsson 1990).
1 Introduction 9
4
Keynes discusses the latent nature of expectations, the problems with dynamic specification,
measurement error, the role of forecast performance and structural breaks.
10 Multivariate Modelling of Non-Stationary Economic Time Series
differenced. In the former case the economy moves about a fixed long-run
equilibrium; in the latter case the long run relates to the rate of growth.
If series require further differencing then the long run can be defined in
terms of levels and differences. In practice it is rare to observe a unit root,
so most parameters estimated on finite samples would be strictly less than
unity and technically consistent with the series being stationary. However,
a substantial number of computed roots would lie close to the unit circle
and as a result it may be seen that the best approximation may arise by
looking at the differenced data. Such a distinction will only be observed
as the sample evolves over considerable periods of time.
When one considers unit roots, then the problem of inversion of the
time series becomes critical as the discrepancy between an AR or sum
of AR parameters approaches zero. As a result the long-run response is
slowed and the reaction of the system to shocks becomes highly persistent.
The notion of a unit root is associated with persistence and poorly
defined impulse response functions. Impulse response analysis is often
used to analyse the performance of a macroeconomic system in response
to shocks. One problem with this is the finding by Lippi and Reichlin
that impulse responses are non-unique with the exception of what is
termed by them a “fundamental time series representation”. This can
be associated with the multivariate ARMA time series form required by
Lütkepohl (2006) for identification. Here the suggestion by Lütkepohl for
analysis of impulse responses in the near non-stationary case is followed
along with the other approaches that give rise to unique impulse response
functions. To remain consistent with the Johansen approach, most of the
primary estimation is derived via multi-step multivariate regression. The
VAR and the VARMA can be estimated in this way and, in the strictly
stationary case, the results are not materially different from full maximum
likelihood (see Spliid 1983). This means that the primary approach to
estimation via the assumption of Gaussianity and regression are similar
for the cointegration and non-cointegration cases. The analysis of this
chapter is completed via consideration of variance decompositions.
In Chap. 3, the nature of multivariate time series are extended to
the cointegration case. It is explained how the VMA in differences
can be transformed into an error correction model using the Granger
representation theorem and the Smith-McMillan form due to Yoo (1986).
14 Multivariate Modelling of Non-Stationary Economic Time Series
References
Bachelier, L. (1900). Théorie de la speculation. Paris: Gauthier-Villars.
Barten, A. P. (1969). Maximum likelihood estimation of an almost complete set
of demand equations. European Economic Review, 1, 7–73.
Bauwens, L., & Hunter J. (2000). Identifying Long-Run Behaviour with Non-
Stationary data. Discussion Paper CORE, The Catholic University, Louvain-
La Nueve DP 2000/43.
Boswijk, H. P. (1996). Cointegration, identification and exogeneity: Inference
in structural error correction models. Journal of Business and Economics and
Statistics, 14, 153–160.
Burke, S. P., & Hunter, J. (2011). Long-run equilibrium price targeting. Quanti-
tative and Qualitative Analysis in Social Sciences, 5, 26–36.
Clements, M. P. (2005). Evaluating econometric forecasts of economic and financial
variables. Basingstoke: Palgrave-Macmillan.
Another random document with
no related content on Scribd:
— Hélas ! Seulement deux canards et une oie ; mais nous avons
vu un Patagon !
— Pas possible.
— Mais si, vraiment. Vu, de nos yeux vu, ce qui s’appelle vu…,
là-bas, sur le sommet de cette montagne, une grande silhouette, très
grande, qui a paru nous observer quelques minutes et qui, malgré
nos démonstrations pacifiques, s’est éclipsée. Nous avons voulu
l’atteindre, mais c’est trop loin et trop haut. Et nous vous sommons
de consigner l’incident sur vos tablettes ; il est assez extraordinaire
pour que le public en soit informé, d’autant plus qu’on affirme que
cette partie de la côte est déserte.
— Comment donc ! Mais certainement… Très intéressant. Oh ! je
le relaterai. Grande silhouette sur une grande montagne…; c’était
sûrement un Patagon. Que vous êtes heureux !
Avant de regagner la Junon, nous relevons les inscriptions
suivantes, plantées un peu partout, sur la terre et dans les îles :
Ramsès, 23/6/78. — Ariadne, 19 Jan. 78. — H.M.S. Penguin, Jan.
5th 78. — Aiguillette (les autres caractères effacés, la planche
trouvée à terre). — Patagonia, 8 nov. 73. — Canonera peruana
Pilcomayo, comm. D. A. S. Muñoz, 11 Diciembre 74. — Christopho
Columbo, 11/9/78.
La nuit est venue, et une fois réunis à bord, nous constatons
l’absence de notre camarade Ed. S…, un robuste enfant de l’Alsace,
voulant toujours « donner la main » aux manœuvres et toujours le
premier aux excursions. Nous l’avons surnommé le « matelot ». On
crie, on appelle. Pas de réponse… On hisse deux fanaux en tête du
mât. Pendant qu’on envoie un canot à terre pour l’attendre et allumer
un feu qui lui montre la direction, les commentaires vont leur train : il
a perdu son chemin, — il est tombé dans un précipice, — il a été
enlevé par les indigènes… — et mangé peut-être ! Enfin, un coup de
fusil se fait entendre du rivage, nous répondons au signal et
quelques minutes après notre ami est à bord. Mais dans quel état !
Le visage et les mains déchirés par les épines, les vêtements en
lambeaux, trempé, meurtri… Il nous raconte qu’ayant escaladé la
plus haute montagne, il s’est égaré dans les bois au retour et s’est
vu forcé de traverser presque à la nage un des lacs pour ne pas se
perdre de nouveau dans les fourrés.
— Alors, c’était bien vous qui étiez là-haut sur la montagne ?
— Oui, et je suis redescendu pour explorer le versant opposé.
Je m’adresse aux chasseurs :
— En narrateur fidèle, je crois, messieurs, qu’il convient de
rectifier l’apparition du Patagon.
— Hélas ! oui. Rectifiez, mais expliquez qu’avec ce gaillard-là on
ne sait jamais à quoi s’en tenir. Qualifiez-le de passager-matelot-
patagon, avec la réserve de bien d’autres qualifications qui lui seront
probablement données avant le retour.
Le 7, au point du jour, la Junon repartait, non sans avoir envoyé
le charpentier clouer sur un arbre bien en vue, à l’entrée de la baie,
une planchette avec l’inscription : Junon, vap. français, commandant
Biard, 7/10/78.
Le petit lac d’eau douce que nous avons découvert, étant à peine
indiqué sur les cartes et ne portant aucun nom, le commandant,
après en avoir relevé approximativement le contour, lui a donné le
nom de lac d’Aunet [7] .
[7] Mme Biard, née Léonie d’Aunet, a fait en 1839,
avec son mari, à bord de la corvette de l’État la
Recherche, un voyage au Spitzberg.
Il y a là, par 80° de latitude nord, une petite crique
qu’on a appelée alors l’anse Léonie ; en sorte que deux
points situés aux extrémités du monde portent aujourd’hui
le nom de la célèbre voyageuse.
Mme Biard est morte à Paris le 21 mars 1879.
En mer, 26 octobre.
A LA MÉMOIRE
DES VICTIMES IMMOLÉES PAR LE FEU
LE 8 DÉCEMBRE 1863
L’AMOUR ET LA DOULEUR INEXTINGUIBLES
DE LA POPULATION DE SANTIAGO