Albert Rex Bergstrom, Khalid Ben Nowman-A Continuous Time Econometric Model of The United Kingdom With Stochastic Trends (2007)

A Continuous Time Econometric Model of the United
Kingdom with Stochastic Trends

Over the last thirty years there has been extensive use of continuous
time econometric methods in macroeconomic modeling. This monograph
presents the first continuous time macroeconometric model of the United
Kingdom incorporating stochastic trends. Its development represents a
major step forward in continuous time macroeconomic modeling. The
book describes the new model in detail and, like earlier models, it is designed in such a way as to permit a rigorous mathematical analysis of its
steady state and stability properties, thus providing a valuable check on the
capacity of the model to generate plausible long-run behavior. The model
is estimated using newly developed exact Gaussian estimation method for
continuous time econometric models incorporating unobservable stochastic trends. The book also includes discussion of the application of the model
to dynamic analysis and forecasting.
The late ALBERT REX BERGSTROM was Emeritus Professor of Economics, a former Dean of the School of Social Studies, Chairman of the
Department of Economics and Pro Vice Chancellor at the University of
Essex and a Fellow of the Econometric Society. He was one of the worlds
leading authorities on continuous time econometric modelling. Professor Bergstrom was formerly Professor of Econometrics at the University of
Auckland and Reader at the London School of Economics. His professional
papers appeared in leading journals such as Econometrica and Econometric
Theory. Professor Bergstroms earlier books include The Construction and Use
of Economic Models (1967), Continuous Time Econometric Modelling (1990),
Statistical Inference in Continuous Time Economic Models (editor, 1976) and
Stability and Inflation: Essays in Memory of A. W. Phillips (edited with A. J. L.
Catt, M. H. Peston and B. D. J. Silverstone, 1978).
KHALID BEN NOWMAN is Professor of Finance at the Westminster
Business School, University of Westminster in London. He previously
worked at City University Business School, London Business School, University of Essex, Durham University, University of Kent and in the banking
sector at the Bank of England, First National Bank of Chicago and Barclays
Bank. Professor Nowmans papers have appeared in leading journals such
as Econometric Theory, Journal of Finance, Journal of Financial and Quantitative
Analysis, and the Journal of Economic Dynamics and Control, among others.
A Continuous Time
Econometric Model
of the United Kingdom
with Stochastic Trends
Albert Rex Bergstrom

University of Essex
Khalid Ben Nowman

University of Westminster
CAMBRIDGE UNIVERSITY PRESS
Paulo
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao
Cambridge University Press
32 Avenue of the Americas, New York, NY 10013-2473, USA
www.cambridge.org
Information on this title: www.cambridge.org/9780521875493
Estate of Albert Rex Bergstrom and Khalid Ben Nowman 2007

c Cambridge University Press 2007
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.
First published 2007
Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication Data
Bergstrom, A. R. (Albert Rex)
A continuous time econometric model of the United Kingdom with stochastic
trends / Albert Rex Bergstrom, Khalid Ben Nowman.
p. cm.
Includes bibliographical references and index.
ISBN-13: 978-0-521-87549-3 (hardback)
ISBN-10: 0-521-87549-8 (hardback)
1. Great Britain Economic policy Econometric models. 2. Finance Great
Britain Econometric models. 3. Econometric models.
4. Stochastic processes. I. Nowman, Khalid Ben, 1962 II. Title.
HC256.7.B47 2007
330.941 086 dc22
2006037265
ISBN 978-0-521-87549-3 hardback
Cambridge University Press has no responsibility for
the persistence or accuracy of URLs for external or
third-party Internet Web sites referred to in this publication
and does not guarantee that any content on such
Web sites is, or will remain, accurate or appropriate.
BN: To Miho
Contents
xi
xiii
xix
List of Figures and Tables
page
Foreword by Peter C. B. Phillips

Preface
Introduction to Continuous Time Modelling
1.1
1.2
1.3
1
3
1.4
1.5
1.6
1.7
1.8
Introduction
Why Model in Continuous Time
Introduction to General Continuous Time
Models
Continuous Time Models in Finance
Continuous Time Macroeconomic Modelling
Policy Analysis in Continuous Time
Macroeconomic Models
Stochastic Trends in Econometric Models
An Outline of Contents
vii
9
18
31
42
45
47
Contents
Continuous Time Econometrics with

Stochastic Trends
2.1
2.2
2.3
2.4
2.5
Introduction
The Continuous Time Model
The Exact Discrete Model and Its VARMAX
Representation
Estimation and Forecasting
Conclusion
Appendix A: Formulae for the Coefficient Matrices
of Exact Discrete Model
Appendix B: Formulae for the Autocovariance
Matrices
50
50
53
58
67
79
80
85
Model Specification
114
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
3.10
3.11
3.12
3.13
3.14
3.15
3.16
114
116
125
128
130
134
136
138
141
144
145
146
147
147
149
150
Introduction
Equations and General Properties of the Model
Private Consumption
Residential Fixed Capital
Employment
Private Non-Residential Fixed Capital
Output
Price Level
Wage Rate
Interest Rate
Imports
Non-Oil Exports
Transfers Abroad
Real Profits Interest and Dividends from Abroad
Cumulative Net Real Investment Abroad
Exchange Rate
viii
Contents
3.17 Stocks
3.18 Conclusion
Appendix A: Derivation of General Adjustment
Equations
Appendix B: Distributed Lag Relations
151
152
Steady State and Stability Analysis
173
4.1
4.2
4.3
4.4
4.5
173
175
180
192
197
197
203
Introduction
The Steady State
Stability Analysis
Stability and Bifurcations
Conclusion
Appendix A: Steady State Level Parameters
Appendix B: Transformed Model
152
164
Empirical Estimation of the Model

and Derived Results
213
5.1
5.2
5.3
5.4
5.5
5.6
213
214
223
232
240
248
Introduction
Estimation from UK Data
Time Lag Distributions
Steady State and Stability Properties
Post-Sample Forecasting Performance
Conclusion
Appendix A: Linear Approximation
about Sample Means
Appendix B: Data
249
262
269
285
288
References
Author Index
Subject Index
ix
Figures
1.1
UK and US Interbank Interest Rates
1.2
Japan and Singapore Interbank Interest Rates
5.1
Lag Distribution Private Consumption
226
5.2
Lag Distribution Employment
227
5.3
Lag Distribution Residential Fixed Capital
227
5.4
Lag Distribution Private Non-Residential Fixed Capital
228
5.5
Lag Distribution Output
228
5.6
Lag Distribution Interest Rate
229
5.7
Lag Distribution Imports
229
5.8
Lag Distribution Non-Oil Exports
230
5.9
Lag Distribution Current Transfers Abroad
230
xi
page 22
23
5.10 Lag Distribution Profits Interest and Dividends from

Abroad
231
5.11 Lag Distribution Investment Abroad
231
Tables
1.1
Single Factor Continuous Time Interest Rate Models
21
5.1
Parameter Estimates from United Kingdom Quarterly

Data
219
5.2
Estimated Time Lag Parameters
225
5.3
Values of Exogenous Variable Time Path Parameters

Assumed in Steady State Analysis
235
5.4
Derived Steady State Growth Rates
235
5.5
Derived Steady State Levels
236
5.6
Eigenvalues for Linear Approximation about Steady

State
239
Root Mean Errors of Post-Sample Multi-Period

Forecasts
242
Post-Sample Multi-Period Forecasts
243
5.7
5.8
xii
Foreword
The daunting complexity of aggregate economic behaviour

has led recent generations of macroeconomists to pursue
a rich diversity of approaches in modelling. Some have
focused on the development of tightly specified small scale
systems that embody rational expectations, real business
cycle mechanisms and intertemporal optimization principles
to strengthen the economic foundations of the models and
furnish meaningful prior restrictions. Others have preferred
the use of model formulations that are more convenient in
coping with the time series properties of the observed series,
using both parametric and semiparametric approaches, and
allowing for mechanisms that can accommodate the regime
changes that can occur in practice. Still others have begun
to work with large dimensional panels and dynamic factor
decompositions.
xiii
Foreword
One unifying force amongst the growing diversity of empirical modelling research in macroeconomics has been
acknowledgement of the importance of long-run behaviour
and the recognition that trending mechanisms in economics
are stochastic. A second area of commonality lies in the use
of nonlinear dynamics, both systems dynamics and volatility
dynamics, the latter being especially important in financial
market applications.
The present volume reports the construction and implementation of a new macroeconomic model of the UK economy that embodies most of these themes. While the model
differs from much of the mainstream of modern macroeconomics in terms of its genesis and form, it shares some
commonality with mainstream work in its detailed attention to nonlinear dynamics, its concern for stochastically
trending data and its use of dynamic optimization principles
in the derivation of adjustment relations. The model developed here is capable of describing diverse patterns of cyclical behaviour in economic aggregates and long-run growth.
It is formulated in continuous time as a system of mixedorder stochastic differential equations. It synthesizes real,
monetary, financial, labour and production sectors of the
economy. It allows for market disequilibria in a systematic way, with parameterized adjustment mechanisms that
xiv
Foreword
measure responsiveness to deviations from partial equilibria whose values are internally consistent and determined
by economic theory. The models long-run properties are
explored analytically, an explicit steady state growth path is
obtained and plausible regions of the parameter space are
determined a priori, thereby providing useful guidelines for
empirical model evaluation. In many of these respects, the
model is very different from prevailing traditions in empirical macroeconomic modelling.
The model is the latest and most sophisticated in a succession of empirical econometric models of the United Kingdom that have been developed under the leadership of the
senior author Rex Bergstrom. The present model therefore
represents the culmination of a long research agenda in
which the cardfile of models has steadily increased in complexity. The incorporation of internally embedded stochastic
trends, the allowance for temporally aggregated flow data
and the introduction of higher order and more complex
lag responses distinguish the present contribution, giving
the new model important elements of realism and features
that are comparable to and in some ways exceed those of
other empirical macro models. Indeed, the reduced form
discrete time version of the present model is a higher order
vector autoregressive (VAR) moving average model with
xv
Foreword
exogenous variables that involves 18 endogenous variables,

27 speed of adjustment parameters, 33 long-run coefficients
and 3 trend parameters. In this respect, it has time series
dynamics that are considerably more complex than those of
most structural and unrestricted VAR specifications, many
other time series models and dynamic structural equation
macroeconomic models.
Bergstrom and Wymer (BW, 1976) developed an early
model in this general class for the UK economy. The BW
model is widely recognized as the first aggregate model of
economic activity that was formulated in continuous time
as a set of interdependent stochastic differential equations.
That model was highly successful as a parsimonious representation of economic activity in a modern industrialized
economy and it has since served as a prototype for similar
models in many other countries. The present volume continues the BW tradition of continuous time macroeconometrics; it integrates modelling enhancements that were
introduced in the subsequent higher order model reported
in Bergstrom, Nowman and Wymer (BNW, 1992), and it
breaks new ground in empirical continuous time modelling
by embedding stochastic trend elements into the system and
growth paths.
xvi
Foreword
This book carries the indelible signature of Bergstroms

superb scholarship. The theoretical model is developed with
great attention to underlying economic ideas, the econometric methodology is systematically built on the extraction of
an exact discrete model and an algorithm that constructs
the Gaussian likelihood, the empirical implementation is
painstakingly conducted, the size of the system and its
complex transcendental matrix nonlinearities push to the
limits of present computational capacity and the empirics involve specification testing and prediction evaluation
against a highly competitive VAR system with exogenous
inputs.
Ragnar Frisch, one of the founders of the Econometric
Society and a co-recipient of the first Nobel Prize for economics, believed that the ultimate strength of econometrics as a scientific tool for understanding economic activity
comes through the combined use of three elements mathematics, statistical analysis and economic theory. This Frisch
mantra is easy to cite as a guiding principle in quantitative
economic research, yet it is far more difficult to implement
in practice. The present work makes a substantive contribution to econometric knowledge but is of wider methodological interest because it succeeds in meeting Frischs criteria
xvii
Foreword
and aspirations for productive research in econometrics. In

so doing, it lays out a paradigm for others to follow.
This volume sadly concludes Rex Bergstroms distinguished work on continuous time econometric modelling,
the manuscript being completed just a few months before
his unexpected death on 1 May 2005 and subsequently
revised by his former student and co-author, Ben Nowman.
We hope that the research agenda will continue into the
future and attract further generations of researchers as well
as those who were themselves trained by Bergstrom or his
students. The present volume will remain a distinguished
magnum opus that concludes Bergstroms own work, a testimony to his sustained accomplishments in founding this
field, a resource for present and future researchers and a
basis for further work as the frontiers of macroeconometric
modelling and continuous time econometrics continue to
unfold.
Peter C. B. Phillips
Yale University
May 2006
xviii
Preface
Over the last 30 years there has been a growing use of continuous time econometric methods in macroeconomic modelling. Economy-wide continuous time macroeconometric
models have been developed for most of the leading industrial countries of the world. This monograph presents the
first continuous time macroeconometric model of the United
Kingdom incorporating stochastic trends. Its development
represents a major step forward in continuous time macroeconomic modelling. The book describes the new model in
detail and like earlier models it is designed in such a way
as to permit a rigorous mathematical analysis of its steady
state and stability properties, thus providing a valuable check
on the capacity of the model to generate plausible long-run
behaviour. The model is estimated using exact Gaussian estimation methods for continuous time econometric models
xix
Preface
incorporating unobservable stochastic trends. The book also

includes discussion on the application of the model to
dynamic analysis and forecasting.
An earlier version of this work was presented at the
Cowles Foundation Conference on Recent Developments
in Time Series Econometrics, Yale University, 1999; the 3rd
Japan Economic Policy Association International Conference, Tokyo, 2004; Westminster Business School, 2004; and
the 3rd Nordic Econometric Meeting, Helsinki, 2005. We
are grateful for comments from these presentations. Ben
Nowman would like to give special thanks to Peter Phillips
for his extensive comments and support in helping him to
revise the book. I also thank Peter Phillips for kindly sending me the photo of Rex Bergstrom and for permission to
use it. He also thanks his colleagues at Westminster Business
School for their support and especially Len Shackleton. I also
thank William Barnett and Yijun He for sending me research
papers. We are very grateful to Marcus Chambers for his
comments and constant support over many years in our
work. We would like to thank Scott Parris, the economics
editor of Cambridge University Press, for his encouragement
and support in this project and two anonymous referees for
valuable comments. Lastly we both give our warmest thanks
to Miho Takashima for her constant support during the
xx
Preface
writing of this book. Royalties from this volume will be used

to contribute to the A. R. Bergstrom Prize in Econometrics
established in 1993 for research excellence in econometrics.
A. R. B. and K. B. N.
January 2005
xxi
CHAPTER ONE
Introduction to Continuous
Time Modelling
1.1 Introduction
he worlds economies and many of its financial markets

operate 24 hours a day and the formulation of realistic
models to represent the economy and these markets is one of

the major challenges facing economists and finance specialists today. Given that the economy is made up of millions of
agents making decisions continuously the use of a continuous time model that allows for these interactions to be incorporated will be a more realistic description of the underlying phenomena we are trying to model. In response to this
challenge, the econometric estimation of continuous time
models in economics and finance has been a major ongoing development over the last 30 years. The major problem
faced in such econometric modelling is that we do not have a
1
Continuous Time Econometric Model of UK with Stochastic Trends
continuous time record of observations and most economic

data are available only at discrete time periods (annually,
monthly, quarterly). In finance, data are more frequently
available at a higher frequency, for example, daily, hourly
and more recently tick-by-tick transactions data. The problem facing an econometric investigator, therefore, is how
best to utilize such data in the estimation of an underlying
continuous time system so that the fitted model can then be
used in economic forecasting, policy analysis and derivative
pricing.
In financial markets continuous time models have found
widespread applications in financial asset and option pricing since the seminal work of Black and Scholes [1973] and
Merton [1973a] and in the general area of continuous time
term structure modelling and bond valuation over the last
20 years. These developments have opened up an explosion of research activity developing new continuous time
models in finance and associated econometric estimation
techniques for fitting continuous systems with the available
discrete data.
The history of the earliest work on continuous time
stochastic processes (for example, Bachelier [1900], Einstein
[1906], Wiener [1923] and Ito [1946, 1951]) and continuous time econometric modelling (Bartlett [1946], Koopmans
2
[1950], Phillips [1959], Durbin [1961]) has been well documented in survey papers by Bergstrom [1984a, 1988, 1990,
Ch. 1, 1996]. Our aims in this chapter are to give a brief
introduction to continuous time modelling and econometrics in areas that are important to economists and finance
specialists working in academia, government and financial
markets.
This chapter is organized as follows. Section 1.2 explains
the advantages and problems of continuous time modelling.
Section 1.3 introduces the general idea of a continuous
time model and the estimation of its parameters with available discrete time data. Section 1.4 discusses the application of continuous time models to term structure models in
finance and Section 1.5 introduces continuous time macroeconomic modelling. Section 1.6 discusses policy analysis and
Section 1.7 introduces the idea of stochastic trends in econometrics. Lastly, Section 1.8 outlines the contents of the book.
1.2 Why Model in Continuous Time

In economic and financial modelling, researchers face a fundamental choice between modelling in continuous and discrete time. This choice raises the important question of why
models should be formulated in continuous time rather than
3
discrete time. There are a number of advantages of modelling in continuous time that we briefly summarize here
(see Bergstrom [1996] for an extensive discussion). First,
since the economy and financial markets are continuously
operating and the underlying decision processes we are trying to model involves millions of decisions by economic
agents within the recorded data observation interval, realistic models will depend on the continuous passage of time.
A continuous system will therefore be sympathetic to the
underlying interactions being modelled and captured in the
data, whereas traditional discrete time models are inherently
less flexible because they restrict the underlying decisionmaking lag structures to match the observation interval in
the data exactly (daily, monthly, quarterly), so that at best
an economys equilibrium and disequilibrium characteristics are being captured at successive points of time. The
potential for misspecification therefore seems greater in a
discrete time system when the underlying phenomenon is
continuous.
Secondly, economic and financial models typically comprise two types of variables. Stock variables observed at
points in time (for example, the money stock, inventories,
fixed capital) and flow variables (for example, consumption,
exports, output and imports) observed as aggregations over
4
the unit observation period. Continuous time econometric

methods, as applied in this monograph, allow for the correct
treatment of these different types of variables in the direct
and exact estimation of the models. This solves the temporal
aggregation bias that commonly occurs in using discrete time
models with flow variables, where no distinction is given to
stock and flow variables generally.
Thirdly, when economists in government and central
banks develop econometric models for short- and long-term
economic forecasting, the construction of a quarterly or
monthly model is commonplace due to the availability of
national accounts data of that frequency. This dependence
of the formulation of the model on the observed data frequency is a major disadvantage of discrete time models.
For example, the construction of a model on monthly data
will be different from one based on daily data. In contrast,
the specification of a continuous time model in economics
and finance is independent of the available data frequency.
Related to the previous point, the fourth advantage of continuous time models is in economic and financial forecasting. The use of continuous time models allows one to obtain
continuous time paths of the variables that can be used to
make forecasts at shorter intervals to the available data used
in the estimation of the model. For example, the continuous
5
time path forecasts of manufacturing production and gross

domestic product in the economy would be useful to government Treasury departments and central banks in their
policy setting. In the business sector, companies would find
it useful to have the continuous time path of retail sales
projections for setting future production levels for example. A discrete time model is less flexible where a quarterly
macroeconomic model would generate forecasts of gross
domestic product at quarterly intervals and if we wanted
forecasts on a weekly or monthly interval would have to be
interpolated.
A fifth advantage of continuous time modelling is concerned with the modelling of dynamic adjustment mechanisms in the economy where continuous systems allow for
more realistic specifications of the partial adjustment processes compared to a discrete time model. Typically, continuous systems in formulating partial adjustment (error correction) mechanisms allow the dependent variable to adjust
continuously in response to deviations from its partial equilibrium level, which may be set to continuously depend on
other variables in the model according to some underlying economic theory. These partial adjustment equations
usually take the form of first- or second-order differential
equations but could, of course, be much more general. The
6
main reason why economic variables adjust gradually, rather

than instantaneously, to their partial equilibrium level is
that there are adjustment costs, which depend on the rate
of change and, possibly, the acceleration of the adjusting
variable. The resulting mechanisms can be formally derived
as part of the solution of a dynamic optimization problem,
which takes account of the costs of adjustment, as is done
in the present monograph.
The sixth advantage of continuous time models is their natural application in the development of finance theory and
derivative pricing models since the pathbreaking work of
Merton [1969, 1971, 1973a,b] and Black and Scholes
[1973]. Sundaresan [2000] provides a major review of these
developments, continuous time models in finance and their
many applications. There are a number of other advantages
of continuous time modelling and the reader is referred to
Bergstrom [1990, Ch. 1, 1996] and Gandolfo [1981, 1993]
for further discussion and motivation.
One of the disadvantages of continuous time modelling is
that the estimation of these models has involved the development of complicated and sometimes specialized econometric methods that have in the past largely remained in
the province of econometric theorists and finance specialists. But over the last 10 years there has been a substantial
7
increase in the number of Ph.D. educated graduates in

econometrics, finance, physics and mathematics who are
capable of and interested in using more advanced econometric techniques in financial and macroeconomic applications. These graduates have taken up employment in government departments and the banking and financial sectors,
as well as academia, and are busy working on quantitative research, risk management, asset price modelling, hedge
fund research and trading floor activities, where advanced
econometric methods are now being more commonly used.
Secondly, the specification of a macroeconomic model in
continuous time typically involves a system of structural
equations derived from economic theory and detailed mathematical analysis of its long-run properties and steady state
analysis. This work is very time consuming mathematically and re-specification and analysis is seldom easy to do.
This compares to discrete time macroeconomic modelling in
which a typical model is constructed of individually specified equations and least squares or instrumental variable
estimation methods are usually employed. Re-specification
and updating of estimation results as new quarterly data
come available for a quarterly forecasting round can be individually done and easily achieved. On the other hand, the
8
explosive development of continuous time modelling in economics, and more especially finance over the last 20 years,
together with the enormous increases in computational
capability has made the specification and estimation of continuous time econometric models a much more plausible
practical enterprise.
1.3 Introduction to General Continuous

Time Models
We present in this section a review of the general form of
continuous time models and the basic ideas of econometric
estimation using discrete time data. The general model considered by Bergstrom [1966a] is a first-order continuous
time system specified as
d x(t) = {A()x(t) + b()}dt + (dt) (t 0),
(1.1)
where x(t) = {x1 (t), . . . , xn (t)} is a n-dimensional continuous time random process, A() is an n n matrix whose elements are functions of a vector = [1 , . . . , p ] of unknown
structural parameters ( p n(n + 1)) and b() is a vector
that is a function of . The error term (dt) is assumed
to be a vector of white noise innovations (see Bergstrom
9
[1984a] for a precise definition and interpretation of this

system).
We assume the continuous time model generates equispaced discrete data observed as the sequence {x(0),
x(1), . . .} and our objective is to estimate the parameters
of the continuous time model. Two approaches that have
been used in the past have become known as the Discrete Approximation Method and the Exact Discrete Model
approach. The first approach was originally developed by
Bergstrom [1966a] (see also Houthakker and Taylor [1966],
Sargan [1974, 1976], Wymer [1972]). The discrete approximation is obtained by using a trapezoidal approximation that
gives an approximate simultaneous equations system of the
form
1
A(){x(t) + x(t 1)} + b( ) + ut
2

E(ut ) = 0, E ut ut = ,

E us ut = 0, s = t, s, t = 1, 2, . . . .
x(t) x(t 1) =
(1.2)
In using this model it was shown in Bergstrom [1966a] that

the estimates have a small asymptotic bias. The continuous
time model can be extended to include an m-dimensional
vector of exogenous variables z(t) = {z 1 (t), . . . , z m (t)} with
10
an associated n m coefficient matrix B() so the model has

the form:
d x(t) = {A()x(t) + B()z(t)} dt + (dt) (t 0).
(1.3)
The first element of the exogenous vector will typically represent the vector of constants in the system. The approximate simultaneous equations model is now given by
1
A(){x(t) + x(t 1)}
2
1
+ B(){z(t) + z(t 1)} + ut
2

E(ut ) = 0, E ut ut = ,

E z s ut = 0, s, t = 1, 2 . . . .
E us ut = 0, s = t,
x(t) x(t 1) =
(1.4)
The use of this approximate simultaneous equation model

has been extensive in continuous time empirical work in
many areas of economics, macroeconomic modelling, financial and commodity markets over the last 30 years. These
applications are listed and discussed in Bergstrom [1988,
1996].
The second and more complicated approach of using the
exact discrete model was suggested and explored in Phillips
[1972]. The exact discrete model is obtained from the
11
solution of the continuous time model (see Bergstrom

[1984a, Theorem 3, p. 1167]) and is given below for the
simple closed model case:
x(t) = F ()x(t 1) + g() + t
1
1
F = e A( ) = I + A + A2 + A3 +
2!
3!

g = e A( ) I A1 ()b()
1

E(t ) = 0, E t t =
e rA e rA dr = ,
0

E s t = 0,
s = t.
(1.5)
As can be seen from this formulation, the exact discrete

model coefficients are complicated functions of the underlying parameters of the continuous time model, and such was
the available computer power in the 1960s and 1970s that
it was infeasible to use this approach in general empirical
work. It was, however, shown in Phillips [1972] that this
exact model can be used to obtain consistent and asymptotically efficient estimates of the parameter vector , and
in an important Monte Carlo study Phillips [1972] found
that these estimates have finite sample properties superior
to those obtained by using the approximate simultaneous
model (see Phillips [1972] for details). This has important
implications for obtaining more accurate economic forecasts
from the estimated model in practice.
12
More complicated dynamic adjustment mechanisms in

the economy can be handled by using second or higher
order continuous time models. There is, of course, no theoretical economic reason for adjustment processes in the
economy to be of first order and the implied time lag distributions generated by higher order continuous time models
are more complicated (see Chapter 3 for precise forms). A
general second-order model was considered by Bergstrom
[1983, 1985, 1986, 1990] and may be presented as
d[D x(t)] = [A1 ()D x(t) + A2 ()x(t)
+ B()z(t)] dt + (dt),
x(0) = y1 ,
D x(0) = y2 ,
(t 0)
(1.6)
where {x(t), (t > 0)} is a n-dimensional vector of endogenous variables, z(t) is an m-dimensional vector of nonrandom functions (exogenous variables), A1 () and A2 ()
are n n system matrices and B( ) is an n m system matrix
whose elements are known functions of , y1 and y2 are nonrandom n 1 vectors, D is the mean square differential operator and (dt) is a vector of white noise innovations. The first
element of z(t) will normally be unity and so the first column of B will be a vector of constants in the system. Usually
dynamic models include terms for the rate of change, D z(t)
and acceleration, D 2 z(t) of the exogenous variables and this
13
can be done by using the interpolation formulae of Nowman [1991](see also Chambers [1991]). Bergstrom [1986]
provides a precise interpretation of this system.
Associated with these more complex dynamic structures
during the last 20 years, there have been enormous developments in computing technology that have facilitated the
estimation of such models and that use estimation methods, which take account of the exact restrictions on the
distribution of the discrete data implied by the continuous
time model. The most extensively used of these new methods is the exact Gaussian estimation method developed by
Bergstrom [1983, 1985, 1986, 1990] (see also Agbeyegbe
[1984, 1987, 1988], Chambers [1991, 1999], McCrorie
[2000, 2001, 2002], McGarry [2003] and Nowman [1991]).
These methods are applicable to models formulated as systems of higher order and mixed-order stochastic differential equations and under appropriate conditions, yields exact
maximum likelihood estimates of their structural parameters from a sample of mixed stock and flow data.
Typical economic models comprise both stock and flow
variables and we may assume that the stock variables are
observed at specific points in time such as the points 0, 1,
2, . . . , T and the flow variables are observed as integrals over
the unit intervals [0, 1], [1, 2], . . . , [T 1, T ]. We let the
14
elements of x(t), z(t), y1 and y2 be ordered (without loss of

generality) so that
x(t) =
y1 =
x s(t)
x f (t)
s
y1
f
y1

,
z(t) =
z s(t)
z f (t)
s
y2
y2 =
f ,
y2
where x s(t) is a vector of endogenous stock variables, x f (t)

is a vector of endogenous flow variables, z s(t) is a vector
of exogenous stock variables, z f(t) is a vector of exogenous
flow variables. We then define the observable vectors x t and
z t (t = 1, . . . , T ) by

x s(t) x s(t 1)
x t = t
(t = 1, . . . , T ),
f
t1 x (r )dr
1

s
s
{z
(t)
+
z
(t
1)}
(t = 1, . . . , T ).
z t = 2 t
f
t1 z (r )dr
(1.7)
(1.8)
We also define the observable part of the initial state vector

that comprises the level stock vector as y1s = x s (0). We define
y = [y1 , y2 ] as the unobservable part of the initial state vecf
tor that has to be estimated along with the other parameters

of the model. In addition to these parameters, we also have
to estimate the parameters of the covariance matrix of the
white noise errors E[(dt) (dt)] = dt(), where is a
15
matrix whose elements are known functions of . The total

parameters to be estimated then comprise [, , y]. Our aim
in empirical work is to estimate these underlying parameters of the second-order continuous time model. This is
achieved by deriving the exact discrete model corresponding to this second-order system, under certain assumptions
concerning the unobservable continuous time paths of the
exogenous variables (see Phillips [1974a, 1976]). This was
obtained in Bergstrom [1986] for the general case of stock
and flow variables, and is given below:
x 1 = G11 y1 + G12 y2 + E 11 z 1 + E 12 z 2 + E 13 z 3 + 1 ,
x 2 = C11 x 1 + G21 y1 + G22 y2 + E 21 z 1 + E 22 z 2 + E 23 z 3 + 2 ,
x t = F1 x t1 + F2 x t2 + E 0 z t + E 1 z t1 + E 2 z t2 + t ,
(t = 3, . . . , T ). (1.9)
The coefficient matrices and the error term covariance

matrix E( ) = were shown to be complicated functions of the underlying parameters of the continuous time
model (see Bergstrom [1986] for precise forms). Bergstrom
[1986] then derived the exact Gaussian likelihood function
to obtain the exact Gaussian estimates. Letting L(, , y) be
minus twice the logarithm of the Gaussian likelihood and by
using a Cholesky factorization = MM , where M is a real
lower triangular matrix with positive elements along the
16
diagonal, Bergstrom [1983, 1985, 1986] was able to define

the Gaussian likelihood function in a simpler form below:
L=
nT

i2 + 2 log mii
(1.10)
i=1
in which the term mii represents the ith diagonal element of

M and = [1 , . . . , nT ] is a vector whose elements can be
evaluated recursively from M = . The Gaussian estimates
,
[,
y ] of [, , y] are then obtained from the iterative procedure proposed by Bergstrom [1985, 1986]. It was shown
in a Monte Carlo study in Nowman [1993] using a secondorder version of an extended trade cycle model based on the
model of Phillips [1954] (see also Bergstrom [1967, Ch. 3]
and Phillips [1972]) that the Gaussian estimation method
performed well in finite samples. The model used is represented below:
d[DK(t)] = {5 DK(t) + 6 [2 Y(t) + 3 M(t) K(t)]} dt
+ 1 (dt)
d[DC(t)] = {1 DC(t) + 2 [1 Y(t) C(t)]} dt
(1.11)
+ 2 (dt)
d[DY(t)] = {3 DY(t) + 4 [DK(t) + C(t) + G(t) Y(t)]} dt
+ 3 (dt),
where Y(t) is real net national income or output at time t,

C(t) is real private consumption, K(t) is the amount of fixed
17
capital, G(t) is real government consumption, M(t) is real

money supply and 1 , 2 , 3 , 4 , 5 , 6 , 1 , 2 and 3 are the
parameters of the model.
Another general approach to estimating the parameters of
these continuous time models has been to use the Kalman
filter algorithm as discussed in Harvey and Stock [1985] (see
also Zadrozny [1988]). Readers are referred to Bergstrom
[1984a, 1996] for extensive reviews on the developments
and estimation of continuous time models in econometrics.
We now turn to a discussion of the application of continuous
time models in finance and their application to the important
area of interest rate modelling that is widely used in financial
markets today.
1.4 Continuous Time Models in Finance

The development and application of general continuous
time models in finance has been a major ongoing development originating in the important work of Merton [1969,
1971, 1973a, b] and Black and Scholes [1973] (see Merton
[1990] and Sundaresan [2000]). Over the last 30 years a
significant area of application of continuous time models
has been in the modelling of interest rates using linear and
nonlinear models for use in fixed income financial markets.
18
Typically, these models are used in the valuation of different

types of bonds and bond options in various bond markets
(for example, government, corporate, mortgage-backed).
The various features of interest rates we would like to
capture include a drift element and also the idea of mean
reversion in interest rates. When rates happen to be high,
they tend over time to revert to normal levels, and vice
versa. The other important element to incorporate is any
possible link between the level of rates and the volatility of
rates the so-called level-effect.
One important interest rate model was developed by
Chan, Karolyi, Longstaff and Sanders [1992, CKLS, hereafter], which had the advantage of nesting as special cases a
range of different short rate models. These special cases
included the models of Merton [1973a], Vasicek [1977],
Cox, Ingersoll and Ross [1985, CIRSR], Dothan [1978],
Brennan and Schwartz [1980, BRSC], Cox, Ingersoll and
Ross [1980, CIRVR], Constant Elasticity of Variance (CEV)
model of Cox [1975], Cox and Ross [1976] and the Geometric Brownian Motion (GBM) model (see Sundaresan [2000]
for further models). All these models differ in their specification of the conditional mean and diffusion components.
The correct specification is of importance to fixed income
traders to obtain the correct implied pricing of fixed income
19
securities. The CKLS general continuous time interest rate

model is presented below:
dr (t) = { + r (t)} dt + r (t)(dt) (t 0),
(1.12)
where {r (t), t > 0} is a short-term interest rate, and are

the unknown drift and mean reversion structural parameters, is the volatility of the short-term rate, is the proportional volatility exponent and (dt) is a white noise error
term. As can be seen, the conditional mean specification is of
the same form as Bergstrom [1966a] but the diffusion component is now time varying, which differs from the assumption of constant volatility originally formulated in Bergstrom
[1966a]. The special cases of the CKLS model are given in
Table 1.1.
There are many approaches available to estimate the parameters of these models with discrete data. Melino [1994]
provides an excellent early review of the estimation of continuous time models in finance. In CKLS they used the Generalized Method of Moments method of Hansen [1982] and
an approximate discrete time model. Their important findings for the US Treasury market was that term structure
models with volatilities more highly sensitive to the level of
rates have a closer fit to the data. In particular, the magnitude
20
Table 1.1. Single Factor Continuous Time Interest Rate Models
Model
CKLS
Merton
Vasicek
CIRSR
Dothan
GBM
BRSC
CIRVR
CEV
dr (t) = { + r (t)}dt + r (t) (dt)

dr (t) = dt + (dt)
dr (t) = { + r (t)}dt + (dt)
dr (t) = { + r (t)}dt + r 1/2 (t) (dt)
dr (t) = r (t) (dt)
dr (t) = r (t) dt + r (t) (dt)
dr (t) = { + r (t)}dt + r (t) (dt)
dr (t) = r 3/2 (t) (dt)
dr (t) = r (t) dt + r (t) (dt)
0
0
0
0
0
0
1/
2
1
1
1
1/
2
of the relationship was shown to be = 1.5 and far higher

than assumed by the other interest rate models.
Since banks trade in a range of international currencies
extensive empirical analysis has been carried out in other
markets. Examples include Dahlquist [1996], who considered rates for Denmark, Sweden, Germany and the United
Kingdom; Hiraki and Takezawa [1997] considered offshore
rates in Japan; Tse [1995] considered money market rates
for 11 countries and Adkins and Krehbiel [1999] looked at
London Interbank Offer rates. In Figures 1.1 and 1.2 the
relationship between the level of rates and the volatility of
rates for the United Kingdom, United States Japan and Singapore are displayed using weekly 1 month interbank rates
21
16
UK
US
10
14
8
12
6
10
8
6
2
4
2
0
1992 1994 1996 1998 2000 2002 2004
1992
UK volatility
3.0
2.5
1994 1996 1998 2000 2002 2004
US volatility
2.0
1.6
2.0
1.2
1.5
0.8
1.0
0.4
0.5
0.0
0.0
1992 1994 1996 1998 2000 2002 2004
1992
1994 1996 1998 2000 2002 2004
Figure 1.1 UK and US Interbank Interest Rates

22
Japan
Singapore
12
8
10
7
6
5
6
4
3
2
2
1
0
0
1992 1994 1996 1998 2000 2002 2004
1.6
Japan volatility
1992
1994 1996 1998 2000 2002 2004
Singapore volatility
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0
1992 1994 1996 1998 2000 2002 2004
1992
1994 1996 1998 2000 2002 2004
Figure 1.2 Japan and Singapore Interbank Interest Rates

23
obtained from DatastreamTM over the period from 1990 to

2005 where volatility is defined as |rt rt1 |. One of the
general conclusions obtained in these empirical studies is
the marked differences in the link between the level of rates
and the volatility of rates in different financial markets.
Another estimation approach is to use the general Gaussian estimation methods of Bergstrom [1983, 1985, 1986,
1990] and this was first demonstrated by Nowman [1997].
To apply these methods to the discrete model of Phillips
[1972] (see also Bergstrom [1984a]), which assumed a constant conditional second moment that is not usually satisfied by financial data, it was necessary to modify the discrete
model used in the estimation of the parameters of the continuous time interest rate models. Nowman [1997] assumed
that as an approximation to the true underlying model given
by Equation (1.12) that over the interval [0, T ], r (t) satisfied
the stochastic differential equation
dr (t) = { + r (t)} dt + {r (t 1)} (dt),
(1.13)
where t 1 is the largest integer less than t. Nowman

[1997] assumed that the volatility of the interest rate changes at the beginning of the unit observation period (day,
week, month) and then remained constant. This allowed the
modified discrete model of Phillips [1972] and Bergstrom
24
[1984a, Theorem 2] derived in Nowman [1997] to be used

to obtain the Gaussian estimates below
r (t) = e r (t 1) + (e 1) + t (t = 1, 2, . . . , T )
E(s t ) = 0 (s = t)

E(t2 )
(1.14)
(1.15)
e 2(t ) 2 {r (t 1)} d
t1
2 2
(1.16)
=
(e 1){r (t 1)}2 = m2tt .
2
The Gaussian estimates are then obtained from the Gaussian
likelihood function where = [, , , 2 ] and
r (t) e r (t 1) (e 1)
T

L() =
2 log mtt +
,
2
m
tt
t=1
(1.17)
which in Nowman [1997] can be defined as

L() =
T

2 log mtt + t2 .
(1.18)
t=1
The transformed residuals 1 , . . . , T can be computed

from mtt t = t . The approximate discrete model had the
advantage of reducing over the CKLS discrete approximation some of the temporal aggregation bias. The Nowman
[1997] estimator is a quasi-maximum likelihood estimator
25
since the CKLS, CIRSR and BRSC models have a diffusion

term that relates the level of rates to the volatility of rates
and is non-Gaussian. It was also shown in empirical work
on the US and UK financial markets in Nowman [1997]
that the volatility of rates was not sensitive to the level of
rates in the United Kingdom using monthly 1-month interbank rate with a level effect of = 0.2898, and using the
same monthly 1-month US Treasury bill rate data used in
CKLS a level effect of = 1.3610, the differences to the
CKLS estimate being due to the different estimation methods
used. Applications of this Gaussian estimation approach to
international markets include Nowman [1998a, 2002] using
monthly Eurocurrency data for the United States, Japan,
France, Italy and Certificate of Deposit and Gensaki rates for
Japan. An extensive study in Episcopos [2000] investigated
the level effect for 10 countries using interbank rates again
showing marked differences in the level effects.
More recently, Yu and Phillips [2001] made an important
contribution in deriving an exact likelihood function based
on a new Gaussian discrete time form of the CKLS model
to estimate the parameters. The method has the interesting property that it uses non-equispaced observations and a
time change transformation in the exact discrete model to
achieve Gaussianity. They used the idea of a time change in
26
the process to convert a continuous time martingale into a

Brownian motion using the Dambis, Dubins-Schwarz theorem to achieve this (see Yu and Phillips [2001, Lemma 3.1]
and Revuz and Yor [1999]). Yu and Phillips [2001] suggested
an estimation algorithm that involved using the Nowman
[1997] estimator and approximate discrete time model in
an intermediate stage and then the Yu-Phillips transformed
exact discrete model and exact Gaussian estimator. Monte
Carlo evidence reported by them found that the Nowman
estimator provided very good estimates of the diffusion term
in finite samples but that the Yu and Phillips [2001] approach
improves the estimation of the drift and mean reversion
parameters in a range of frequencies (see also Kawai and
Maekawa [2003] and Lo [2005]). Yu and Phillips [2001] also
presented some empirical results, using the UK data used in
Nowman [1997] and the US data used in CKLS, and found
that the Yu-Phillips estimates were close to the Nowman
estimates for the United Kingdom but not for the United
States. A recent application of the Yu-Phillips procedure to
guez
Japanese financial markets is given by Nowman and N
[2005a].
Due to the complicated nature of interest rate behaviour,
continuous time models with more complex nonlinear
drift and volatilities have been proposed for example by
27
At-Sahalia [1996]. Also many other estimation methods

have been put forward including At-Sahalia [1996], Stanton [1997], Jiang and Knight [1997], Bandi and Phillips
[2003] and Phillips and Yu [2005].
An important implication of these continuous time models for traders in fixed income financial markets is the
implied derivative prices. CKLS provided a comparison of
prices of bond options implied by the historically estimated
CKLS specification and some special cases on US data and
found significant variations in the call option values. Nowman and Sorwar [1999, 2005b] applied Gaussian estimation
to the CKLS, Vasicek, CIRSR and BRSC models on monthly
Eurocurrency rates for a number of countries and found that
default free bond prices and contingent claim prices are sensitive to the underlying interest rate model used in a range
of currencies. Phillips and Yus [2005] jackknifing method
provided substantial improvements in bond options pricing
over existing methods (see Phillips and Yus [2005] for further details).
Continuous time multi-factor models have found widespread application in financial market derivative valuation.
Examples of the factors include in addition to the short
rate, the long rate, volatility of the short rate, time varying central tendency of the short rate, spreads and news
28
streams. An important two-factor continuous time model

was developed by Brennan and Schwartz [1979] where
the first factor is the short rate and the second factor the
consol rate. To estimate their term structure model, they
used the estimation procedure of Malinvaud [1966] and the
exact discrete model (Phillips [1972]). Examples of other
multi-factor models include Beaglehole and Tenney [1991],
Longstaff and Schwartz [1992], Chen and Scott [1992], Hull
and White [1994], Chen [1996], Duffie and Kan [1996],
Andersen and Lund [1997] and Babbs and Nowman [1999].
The model of Babbs and Nowman [1999] was a general
multi-factor Generalized Vasciek model and a subclass of
the general Gaussian model of Langetieg [1980]. They dealt
with the most general case of n-factors and derived the general pure discount bond price for the model. In Babbs and
Nowman [1999] they assumed a possible description of the
instantaneous spot interest rate, r, is
r (t) = (t)
J

X j (t),
(1.19)
j=1
where is a constant representing the long-run average rate

and X1 (t), . . . , X J (t) are unobserved state variables representing the current effect of J streams of economic news.
The news streams for example may represent interest rate
29
rumours from the Monetary Policy Committee of the Bank

of England or the Bank of Japans Policy Board Meetings
on monetary policy and long-term economic news representing monthly and quarterly economic statistics that all
effect the yield curve. The arrival of each type of news is
modelled by the innovations of Brownian motions, which
may be correlated, while the impact of a piece of news
dies away exponentially as the time since it was received
increases. Babbs and Nowman [1999] assumed that X j (t)
are generated by the processes below:
d X j (t) = j X j (t) dt + c j j (dt),
(1.20)
where the terms are defined in Babbs and Nowman [1999].

They used the Kalman filter (see Harvey [1989]) to estimate
one-, two- and three- factor Generalized Vasicek models on
US data and found that the two- and three-factor models
were able to explain the US yield curve. In Bergstrom and
Nowman [1999] they applied Gaussian estimation methods
and derived the exact discrete model and the exact Gaussian
likelihood function for the Babbs and Nowman [1999]
model. More recently, Nowman [2001, 2003, 2006] considered the Gaussian estimation of multi-factor CKLS and
CIRSR models and empirical applications on Japanese interest rate data indicated various level effects for the CKLS
30
[2003]
model in the multi-factor cases. Recently, Saltoglu
applied Gaussian estimation to the Brennan and Schwartz
[1979] model. This area of application of continuous time
models continues to be one of the most active research areas
in finance. We now turn to a brief introduction to continuous time macroeconomic modelling.
1.5 Continuous Time Macroeconomic

Modelling
Continuous time modelling has also been extensively applied in macroeconomics since the development of the first
continuous time macroeconomic model by Bergstrom and
Wymer [1976]. Economy-wide continuous time macroeconometric models have been developed for most of the
leading industrial countries of the world, including the
United Kingdom (Bergstrom and Wymer [1976], Knight and
Wymer [1978], Bergstrom, Nowman and Wymer [1992]),
Australia (Jonson, Moses and Wymer [1977]), Canada
(Knight and Mathieson [1979]), Italy (Gandolfo and
Padoan, [1982, 1984, 1987a, 1990], Germany (Kirkpatrick [1987]), New Zealand (Bailey, Hall and Phillips
[1987]), the United States (Donaghy [1993]), Sweden
(Sjo o [1993]), Netherlands (Nieuwenhuis [1995]), Czech
31
Republic (Stavrev [2001]) and various other countries. A

complete list is provided in Bergstrom [1996].
The Bergstrom and Wymer [1976] model was a closed system of 13 interdependent differential equations (10 structural equations and 3 identities). See also Bergstrom [1966b,
1967] for some earlier prototype models. An interesting feature of the model was that it contained only one exogenous
variable a simple time trend. The trend terms were allowed
for the unobservable factors: technical progress, the growth
of labour supply and growth in demand for exports. The
model was based on extensive economic theory and had
35 key parameters, which comprised 16 long-run elasticities and propensities , 16 speed of adjustment parameters
and 3 trend parameters . The model in its deterministic
form is given below:

DC
1 Y
= 1 log
(1.21)
C
C

2 e 1 t {Y 4 3 K 4 }1/4
DL
= 2 log
(1.22)
L
L

Dp
Y 1+4
r +
+ 5 k (1.23)
Dk = 3 4 3
K
p
DY = 5 {(1 6 )(C + DK + E) Y}
+ 6 {7 (C + DK + E) S}
32
(1.24)
DI = 7 {6 (C + DK + E) I }
+ 8 {7 (C + DK + E) S}

DE
8 p 9 e 2 t
= 9 log
E
E

Y 4 (1+4 )/4
1 t
1 3 K
we
Dp
10 2
= 10 log
p
p
(1.25)
(1.26)
(1.27)

Dw
2 e 1 t {Y 4 3 K 4 }1/4
= 11 log
(1.28)
w
11 e 3 t

Dr
12 pY 13 r 14
= 12 log
(1.29)
r
M

E
16 e 3 t
Dm = 13 log
+ 14 log
15 I
L

E
16 e 3 t
+ 15 Dlog
+ 16 Dlog
(1.30)
15 I
L
DS = Y + I C DK E
DK
=k
K
DM
=m
M
where
C = real consumption
Y = real net income or output
K = amount of fixed capital
33
(1.31)
(1.32)
(1.33)
E = real exports
I = real expenditure on imports
S = stocks
L = employment
M = volume of money
r = interest rate
p = price level
w = wage rate
k = proportional rate of increase of fixed capital
m = proportional rate of increase in volume of money
t = time
D = d/dt and 0 < 1 < 1, 4 > 1, 0 < 6 < 1, all other
parameters positive (except possibly 3 ).
An important feature of nearly all existing continuous
time macroeconometric models, starting with the model of
Bergstrom and Wymer [1976], is that they are designed in
such a way as to permit a rigorous mathematical analysis
of their steady state and stability properties, thus providing a valuable check on the capacity of the model to generate plausible long-run behaviour. The general methodology used in this analysis was developed by Bergstrom
[1966b, 1967], using smaller models. For the model of Bergstrom and Wymer [1976] and the many continuous time
34
macroeconometric models for which that model has served

as a prototype including the model of Bergstrom, Nowman
and Wymer [1992] below, it was shown that, when the
exogenous variables satisfy certain conditions the model has
a steady state solution in which all variables grow at constant exponential rates (possibly zero). Moreover, for each
of these models, it was possible to obtain an explicit differential equation system in the logarithms of the ratios of the
variables to their steady state paths, and so that the nonlin
ear part satisfy the PoincareLiapounovPerron
conditions
(see Bellman [1953, p. 93] and Coddington and Levinson
[1955, p. 314]). The asymptotic stability of the steady state
solution can then be deduced from the eigenvalues of the
linear part of this system.
Most previous continuous time macroeconometric models have not been estimated by econometric methods that
take into account the exact restrictions on the distribution of the discrete data implied by the continuous time
model. They have used the discrete approximation method
with a transformation developed by Phillips [1974b] and
Bergstrom and Wymer [1976] for the temporal aggregation problem arising from flow variables with the full information maximum likelihood estimator. This leads to estimates that are not asymptotically efficient compared to the
35
exact discrete model approach. The Bergstrom and Wymer

[1976] model was estimated on quarterly UK data over
195566 using the exact discrete model, but even in this
case the exact restrictions were imposed only on the autoregressive coefficient matrices and therefore the final estimates
were also not asymptotically efficient as the moving average
component of the VARMA model was only approximately
allowed for using their transformation for flow variables.
The first macroeconomic application of the exact Gaussian
method of Bergstrom [1983, 1986] was in the estimation of
the second-order continuous time macroeconometric model
of the United Kingdom developed by Bergstrom, Nowman
and Wymer [1992]. The model based on the Bergstrom
and Wymer [1976] model also made intensive use of economic theory to obtain a parsimonious parametrization and
cross-equation restrictions. One of the innovations of the
Bergstrom, Nowman and Wymer [1992] model in addition
to being specified as a second-order system of differential
equations is that it had a more elaborate financial sector.
The model, in its deterministic form, is set out below. It
is a system of 14 second-order differential equations with
63 structural parameters, including 27 long-run parameters or elasticities, 33 speed of adjustment parameters,
36
and 3 trend parameters representing the rate of technical

progress, the rate of labour supply (the non-accelerating
inflation level of employment) and the target rate of growth
of the money supply. The variables and equations of the
model, in its deterministic form, are as follows:
D 2 logC = 1 (1 + 2 DlogC )

1 e {2 (r Dlog p)+3 Dlog p} (Q + P )
+ 2 log
T1 C
D 2 logL = 3 (2 DlogL)

4 e 1 t {Q 6 5 K 6 }1/6
+ 4 log
L
D 2 logK = 5 (1 + 2 DlogK )

5 (Q /K )[1+6 ]
+ 6 log
r 7 Dlog p + 8
(1.34)
(1.35)
(1.36)
D 2 logQ = 7 (1 + 2 DlogQ )

{1 9 (qp/pi )10 }(C + Gc + DK + E n + E o )
+ 8 log
Q
(1.37)
D 2 log p = 9 (Dlog(w/p) 1 )

11 4 T2 we 1 t {1 5 (Q /K )6 }(1+6 )/6
+ 10 log
p
(1.38)
37
D 2 logw = 11 (1 Dlog(w/p)) + 12 Dlog( pi /qp)

4 e 1 t {Q 6 5 K 6 }1/6
+ 13 log
12 e 2 t
(1.39)
D 2r = 14 Dr + 15 [13 + rf 14 Dlogq
+ 15 { p(Q + P )/M} r ]
(1.40)
D 2 logI = 16 (1 + 2 Dlog( pi I /qp))

9 (qp/pi )10 (C + Gc + DK + E n + E o )
+ 17 log
( pi /qp)I
(1.41)
D 2 logE n = 18 (1 + 2 DlogE n )

16 Yf 17 ( pf /q p)18
+ 19 log
En
D 2F = 20 DF + 21 {19 (Q + P ) F }
(1.42)
(1.43)
D 2P = 22 DP + 23 [{20 + 21 (rf Dlog pf )}K a P ]

(1.44)
D 2K a = 24 DKa + 25 [{22 + 23 (rf r )
24 Dlogq 25 dx }(Q + P ) K a ]
38
(1.45)

26 e3 t
D 2 logM = 26 (3 DlogM) + 27 log
M

En + E o + P F
+ 28 Dlog
( pi /qp)I

E n + E o + P F DKa
+ 29 log
( pi /qp)I

27 pf
D 2 logq = 30 Dlog( pf /qp) + 31 log
qp

En + E o + P F
+ 32 Dlog
( pi /qp)I

E n + E o + P F DKa
+ 33 log
( pi /qp)I
(1.46)
(1.47)
Endogenous variables:
C = real private consumption
E n = real non-oil exports
F = real current transfers abroad
I = volume of imports
K = amount of fixed capital
K a = cumulative net real investment abroad (excluding
change in official reserves)
L = employment
M = money supply
P = real profits, interest and dividends from abroad
39
p = price level
Q = real net output
q = exchange rate (price of sterling in foreign currency)
r = interest rate
w = wage rate
Exogenous Variables:
dx = dummy variable for exchange controls (dx = 1 for
197479, dx = 0 for 1980 onwards)
E o = real oil exports
Gc = real government consumption
pf = price level in leading foreign industrial countries
pi = price of imports (in foreign currency)
rf = foreign interest rate
T1 = total taxation policy variable ((Q + P )/T1 is real private
disposable income)
T2 = indirect taxation policy variable (Q /T2 is real output at
factor cost)
t = time
Yf = real income of leading foreign industrial countries.
The model was estimated using the exact Gaussian estimator for the general case of Bergstrom [1986] on UK
quarterly data over the period 197484 with the predictive
40
performance tested over 198586. As judged by the plausibility of the parameter estimates, the post-sample forecasting
performance and the analysis of its steady state properties,
this model was very successful.
One of the problems in estimating the parameters of continuous time models concerns the issue of the identification
of the structural parameters. We have the aliasing problem
where one cannot distinguish between structures generating cycles whose frequencies (per unit observation period)
differ by integers. For example, the case of half a cycle per
observation period and one and a half cycles per observation period. This phenomenon was originally discussed in
a study by Phillips [1973], where it was shown how identification could be achieved through Cowles Commission
type restrictions on the structural parameters. Hansen and
Sargent [1983] also investigated the aliasing problem. They
showed that there at most a finite number of structures that
are indistinguishable (on the basis of the observations) from
the true structure because of the aliasing phenomenon. One
of the implications of their theorem is that the true structure
will be identifiable provided that the unit observation period
is sufficiently short. In Bergstrom, Nowman and Wymer
[1992] they used an alternative approach to identification,
suggested by the last result of Hansen and Sargent [1983],
41
but as pointed out by them not dependent on their theorem.

This involved the use of bounds on the speed of adjustment
parameters and they were able to show how to achieve for
the complete vector of structural parameters of their model
identification (see Bergstrom, Nowman and Wymer [1992]).
1.6 Policy Analysis in Continuous Time

Macroeconomic Models
Continuous time macroeconomic models have been extensively used in policy analysis (see, for example, Gandolfo and
Padoan [1982, 1984], Gandolfo and Petit [1987b], Jonson
and Trevor [1981], Jonson, McKibbin and Trevor [1982],
Kirkpatrick [1987], Sassanpour and Sheen [1984], Stefansson [1981] and Tullio [1981]). The Bergstrom and Wymer
[1976] model was used for policy analysis in Bergstrom
[1978, 1984b], where an analysis of monetary and fiscal
policy feedbacks was made. This analysis was based on
feedbacks of the type introduced into economic models
by Phillips [1954] and Bergstrom [1967, Ch. 6]. A more
advanced approach is to use optimal control theory. In
Bergstrom, Nowman and Wandasiewicz [1994], they investigated monetary and fiscal policy feedbacks, using the
Bergstrom, Nowman and Wymer [1992] model. As the
42
models steady state computed from the estimated parameters was slightly unstable (although, statistically they could
not reject the hypothesis that it is stable), they investigated
the use of simple fiscal policy feedbacks and the use of
control theory considered in Bergstrom [1987] that could
reduce or eliminate the instability. Optimal feedbacks were
derived by minimizing an infinite horizon quadratic cost
function involving deviations of output, the exchange rate
and the policy instruments from their steady state paths and
also the rates of change of these variables. One of the findings was that fiscal policy had an important role to play in
the stabilization of the model than did monetary policy. A
discussion of the important contributions of Barnett and He
[1999, 2002] concerning bifurcation analysis in continuous
time macroeconometric models is given in Chapter 4.
Another important issue in macroeconomic modelling in
discrete and continuous time concerns the famous Lucas
[1976] critique and also the related issues of deep parameters, structural parameters and rational expectations in
macroeconomic modelling. Deep parameters are taken to
be the parameters of Euler equations and are directly
acquired from the parameters of tastes and technology. The
role of rational expectations in macroeconomic modelling
had an important influence on the Bergstrom, Nowman
43
and Wymer [1992] models formulation. The Bergstrom,

Nowman and Wymer [1992] model made systematic use of
the assumption of long-run rational expectations in a number of ways. The assumption of long-run rational expectations in many of the equations played an important role
in the identification of certain parameters of the model
that would not otherwise be identifiable. In particular, it
is assumed that the agents in the economy know the trend
parameters 1 and 2 for productivity and the labour supply and take these into account in forming their expectations. The assumption that the agents in the economy
know these parameters is, of course, more realistic than
the assumption that they know the parameters affecting
the short-run dynamics of the model, which would be
necessary to justify the assumption of full rational expectations. It is shown in Bergstrom, Nowman and Wymer
[1992] that the steady state growth rate of most of the
real variables in the model is 1 + 2 . The long-run rational expectations assumption is incorporated in the structural equations by setting the expectations equal to 1 + 2
derived from the steady state. The assumption of rational
expectations also played an important role in the wage
rate adjustment equation and in the foreign exchange
44
market. See Bergstrom, Nowman and Wymer [1992] for full

details.
1.7 Stochastic Trends in Econometric

Models
In further work on the Bergstrom, Nowman and Wymer
[1992] model, Nowman [1996] re-estimated the model on
more recent data with a different measure of the money supply. One of the conclusions in Nowman [1996] was that the
use of single deterministic time trends for technical progress
and labour supply had an impact on the forecasting performance of the model. In Nowman [1998b], the trends were
modified as segmented trends over 4-year periods and the
model re-estimated improving the forecasting performance
of the model. Nevertheless, a careful analysis of the estimates and forecasts of these studies suggests that the forecasting performance could be improved by replacing the
deterministic trends (simple non-random functions of the
time parameter) introduced to allow for such factors as technical progress and changes in the non-accelerating inflation
level of employment (which are unobservable) by stochastic trends. Indeed, there are good reasons for believing that
45
the replacement of deterministic trends by stochastic trends

could significantly improve the forecasting performance of
most macroeconometric models, formulated in either discrete or continuous time, that include deterministic trends
to allow for such long-run changes as those mentioned.
There has, recently, been extensive work on the problems of statistical inference in cointegrated systems (see,
for example, Engle and Granger [1987], Johansen [1988,
1991], Phillips and Durlauf [1986], Park and Phillips [1988,
1989], Phillips [1987, 1991a, 1991b, 1995], Franses [1996]
and Chambers [2001, 2003]). Nearly all of this work is concerned with models that are formulated in discrete time and
in which the stochastic trend variables are observable. There
has, however, been some work by Harvey and Stock [1988]
and Bergstrom [1997] on the estimation of the parameters of continuous time dynamic models with unobservable
stochastic trends (see also Corradi [1997] and Simos [1996]).
The article of Bergstrom [1997] extended the exact
Gaussian estimation method developed in the earlier articles, referred to above, so that it is applicable to models incorporating unobservable stochastic trends. It contains a detailed estimation algorithm and formulae for
its implementation, when the model comprises either a
closed or open system of mixed first- and second-order
46
stochastic differential equations, which incorporate unobservable stochastic trends, while the data is a mixture of
stocks (observed at a sequence of equispaced points of time)
and flows (observed as integrals over a sequence of adjoining
intervals of equal length). The algorithm is applicable, therefore, to models which, in addition to incorporating unobservable stochastic trends, involve all the complications in
existing continuous time macroeconometric models.
Such a model has been developed for the UK economy
by the authors of this monograph. Its parameters have been
estimated by the algorithm of Bergstrom [1997], the estimation being the first application of the new algorithm.
The model is the first continuous time macroeconometric
model incorporating stochastic trends, and its development
represents a major step forward in macroeconometric modelling. The purpose of this monograph is to describe the new
model, the estimation of its parameters and its application
to dynamic analysis and forecasting.
1.8 An Outline of Contents

Chapter 2 deals with the econometric methodology used in
the later chapters, with emphasis on its application rather
than its theoretical derivation. The aim of this chapter is
47
to explain the newly developed exact Gaussian estimation

method for continuous time econometric models incorporating unobservable stochastic trends. It is based on the
article of Bergstrom [1997], and readers are referred to
this article for proofs of the various results used here. An
Appendix to the chapter sets out the detailed formulae
needed for the present development and this includes previously unpublished formulae for the autocovariance matrices
of the innovations in the exact discrete model satisfied by the
observations. These formulae play an important role in the
estimation procedure.
The next two chapters are concerned with the specification of the model and the analysis of its steady state and stability properties. The complete model is set out in Chapter 3,
together with a discussion of the specification of the individual equations. The model is, essentially, a development of
the model of Bergstrom, Nowman and Wymer [1992]. The
most important change is that the deterministic trends incorporated in this model to allow for technical progress and
the gradual change in the non-accelerating inflation level
of employment are replaced by stochastic trends. Another
important change is that, whereas all the equations in the
earlier model are specified as second-order differential equations, a few equations in the new model (taking account of
48
the estimated speed of adjustment parameters in the earlier

model) are specified as first-order differential equations, all
other equations being of the second order. This complicates
the estimation procedure, particularly as the data are a
mixture of stocks and flows. However, the exact Gaussian
estimation algorithm developed by Bergstrom [1997] and
described in Chapter 2 allows for the combined complications of mixed-order differential equations and mixed stock
and flow data, as well as unobservable stochastic trends. For
the model developed in this monograph a full steady state
and stability analysis is carried out in Chapter 4. There we
derive closed formulae from which the exact steady state
paths of the variables and linear (in logarithms) approximations about these steady state paths can be computed. The
final chapter (Chapter 5) is concerned with the estimation
of the structural parameters of the model from UK data, the
computation of the post-sample forecasts and the derivation
from the parameter estimates, of the steady state solution
and its stability properties.
49
CHAPTER TWO
Continuous Time
Econometrics with
Stochastic Trends
2.1 Introduction
his chapter is concerned with the general econometric methodology used throughout the book. The main
objective is to exposit the newly developed exact Gaussian estimation method for continuous time econometric
models that incorporate unobservable stochastic trends. The
approach is based on the article of Bergstrom [1997], and
readers are referred to this article for proofs of the various
results used. Bergstrom [1997] extends the exact Gaussian
estimation method developed in earlier work (Bergstrom
[1983, 1985, 1986, 1990]) in two ways. Firstly, the algorithm is applicable to models incorporating unobservable
stochastic trends. These can provide a more flexible and
50
Continuous Time Econometrics with Stochastic Trends
realistic representation of such unobservable trend variables

as technical progress, which have often been represented
in econometric models in the past by deterministic trends.
Secondly, the algorithm is applicable to models formulated
as systems of mixed-order stochastic differential equations,
even when the discrete observations of the variables are a
mixture of stocks and flows. Hitherto, the problem of estimating mixed-order continuous time systems has been dealt
with only for the special case in which all the variables
are measured at equispaced points of time, i.e. there are
no flow variables (see Agbeyegbe [1984]). The new algorithm is applicable, therefore, to models that, in addition
to incorporating unobservable stochastic trends, involve all
the complications assumed in the earlier work on the exact
Gaussian estimation of continuous time macroeconometric
models.
The general method used in the new algorithm is similar to that used in the earlier algorithm of Bergstrom
[1986, 1990] for the exact Gaussian estimation of systems
of higher order stochastic differential equations without
stochastic trends. This involves the derivation of an exact
discrete model for the observations generated by the underlying continuous time model and the representation of this
51
exact discrete model in the form of a VARMAX (vector

autoregressive moving average with exogenous variables)
model. The VARMAX model is then used as a basis for the
estimation of the parameters of the underlying continuous time model and the generation of post-sample forecasts
of the discrete observations. Like the earlier algorithm of
Bergstrom [1986, 1990], the new algorithm is also applicable to models containing exogenous variables. In which case,
the exact discrete model is also in the form of a VARMAX
model.
In the remainder of the chapter the new algorithm will
be discussed in detail when the model is a mixture of
first- and second-order stochastic differential equations and
the discrete observations of the variables are a mixture of
stocks and flows. The continuous time model is presented
and explained in Section 2.2, the exact discrete model and
its VARMAX representation are presented in Section 2.3,
the use of this VARMAX model for the estimation of the
parameters of the continuous time model and the generation of post-sample forecasts of the discrete observations
is described in Section 2.4 and Section 2.5 concludes. The
formulae used in the exact discrete model are presented in
Appendices A and B.
52
2.2 The Continuous Time Model

The assumed model is
d x1 (t) = [A1 ()x 1 (t) + A2 ()x2 (t) + A3 ()D x2 (t) + B1 ( )z(t)
+ C 1 ()(t) + b1 (, )]dt + 1 (dt)
(t 0), (2.1)
d[D x 2 (t)] = [A4 ()x1 (t) + A5 ()x2 (t) + A6 ()D x 2 (t)

+ B2 ()z(t) + C 2 ()(t) + b2 (, )]dt
+ 2 (dt)
(t 0),
d(t) = dt + 3 (dt)

x 1 (0), x2 (0), D x2 (0) = y (0),
(2.2)
(t 0),
(0) = 0,
(2.3)
(2.4)
where {x 1 (t), t > 0} is a real n1 -dimensional continuous

time random process, {x2 (t), t > 0} is a real n2 -dimensional
continuous time random process, {(t), t > 0} is a real kdimensional continuous time random process, z(t) is an
m 1 vector of non-random functions of t (exogenous variables), y(0) is an (n1 + 2n2 )-dimensional non-random vector, is a p-dimensional vector of unknown structural
parameters, is a k-dimensional vector of unknown drift
parameters,
A1 , A2 , A3 , A4 , A5 , A6 , B1 , B2 , C 1 and C 2 are matrices
whose elements are known functions of , b1 and b2 are
53
vectors whose elements are known functions of and , D

is the mean-square differential operator, 1 (dt), 2 (dt) and
3 (dt) are white noise innovation vectors that are, precisely,
defined by Assumption 1.
Assumption 1:
1 , 2 and 3 are vectors of random mea-
sures defined on all subsets of the half line 0 < t <

with finite Lebesgue measure, such that E[i (dt)] =
0 (i = 1, 2, 3), E[i ( 1 ) j ( 2 )] = 0 (i, j = 1, 2, 3) for any
disjoint sets 1 and 2 on the half line 0 < t <
and E[i (dt) j (dt)] = dti j (i, j = 1, 2, 3), where 11 ,
22 and 33 are positive definite matrices and 13 and
23 are zero matrices (for a discussion of random measures and their application to continuous time stochastic
models, see Bergstrom, [1984a, p. 1157]).
In the preceding model, the elements of (t) are unobservable stochastic trends, whereas the elements of x 1 (t)
and x2 (t) are endogenous variables, which are observable,
either at discrete points of time (stock variables) or as integrals over discrete intervals of time (flow variables). At each
point of time, x 1 (t) is adjusting through a first-order differential equation system, and x2 (t) is adjusting through
a second-order differential equation system to the current
values of x1 (t), x2 (t), D x2 (t), z(t), (t) and the white noise
54
innovations. As (2.2) is a second-order equation and the

variable x2 (t) is not mean differentiable to the second order
the appropriate representation is in terms of the increment
in the first-order mean square derivative, i.e. d[D x 2 (t)].
Because, under Assumption 1, neither x1 (t) nor D x 2 (t) is
differentiable (even in the mean square sense), (2.1) and
(2.2) are, more precisely, interpreted as representing the
integral equations
t
x1 (t) x1 (0) =
[A1 x1 (r ) + A2 x2 (r ) + A3 D x2 (r )
0
t
+ B1 z(r )+C 1 (r )+b1 ]dr + 1 (dr ) (t > 0),
0
(2.5)
D x 2 (t) D x 2 (0) =
[A4 x1 (r ) + A5 x 2 (r ) + A6 D x2 (r )
+ B2 z(r ) + C 2 (r ) + b2 ]dr
t
2 (dr ) (t > 0),
+
(2.6)
where
t
0
i (dr ) = i [0, t] (i, j = 1, 2) and the first integrals
on the right-hand sides of (2.5) and (2.6) are defined in the

wide (mean square) sense (see Bergstrom [1984a, p. 1152]).
Similarly, (2.3) is interpreted as meaning that
t
(t) (0) = t +
3 (dr ) (t > 0).
0
55
(2.7)
In the statistical literature dealing with stochastic trends,

it is commonly assumed that both the level and drift of the
stochastic trend path are stochastic. For example, Harvey
and Stock [1988] allowed the drift to evolve according to a
Brownian motion with non-random drift. The model comprising (2.1)(2.4) could be generalized by adding another
equation, similar to (2.3), representing the stochastic evolution of the drift vector . But, for the sort of variables represented by trends in most economic models, this generalization is unrealistic. It implies, for example, that if one of the
elements of represents the rate of technical progress, then
the rate of technical progress will, almost certainly, exceed
any bound at some time in the future. The vector in (2.3)
is assumed, therefore, to be a vector of constants. Moreover, some empirical support for this assumption has been
provided by Harvey, Henry, Peters and Wren-Lewis [1986]
in their estimation of a discrete-time employment-output
equation for the UK manufacturing data. They incorporated
a stochastic trend to represent technical progress and estimated that the drift in this trend was constant; that is the
estimated variance of the innovation in the drift was zero.
There is also a fundamental difference between the way
in which the stochastic trends are introduced into the above
model and the way in which they are introduced into the
56
model of Harvey and Stock [1988], who are among the few
other econometricians to have dealt with the problem of
estimating structural parameters in continuous time models with unobservable stochastic trends. Harvey and Stock
[1988] assumed that each of the observable (at discrete
intervals) variables is the sum of two unobservable components, the first component being one of a set of completely
unobservable variables generated by an underlying stochastic differential equation system and the second being a linear combination of stochastic trends. In the above model the
stochastic trends are more deeply embedded in the structure
and help to derive the differential equation system, just as,
for example, technical progress helps to derive the economy.
The essential distinction between the two approaches is that
in the above model the stochastic trends play the role of random forcing functions, which help to generate the solution
of the differential equation system, whereas in the model of
Harvey and Stock [1988] the stochastic trends have no effect
on the solution of that system, but are added to the solution
(like observation errors) to obtain the observed variables.
Finally, it should be noted that is introduced as an argument in the vector functions b1 (, ) and b2 (, ) in (2.1) and
(2.2). This allows for the incorporation in the model of the
assumption of long-run expectations, that is, the assumption
57
that the various agents in the economy know the drifts and
take them into account when adjusting the variables that
they control.
2.3 The Exact Discrete Model and Its

VARMAX Representation
We turn now to our precise assumptions concerning the
observability of x1 (t), x2 (t) and z(t), (0 < t < ), and the
definitions of the observable vectors used in the formation
of the exact discrete model. To concentrate on the case of
practical importance, we assume that x 1 (t) includes flow
variables only whereas x2 (t) includes both stock and flow
variables. These assumptions are suggested by the parameter estimates for the model of Bergstrom, Nowman and
Wymer [1992]. All of the variables whose partial adjustment processes could, on the basis of those estimates, be
realistically represented by first-order equations were flow
variables such as output, consumption, exports and imports.
Stock variables, such as the stock of fixed capital and the
level of employment, all adjusted relatively slowly, and the
modes of the time lag distributions implied by their estimated speed of adjustment parameters were all quite large
(at least 9 months), indicating that a second-order equation
58
was more realistic for these variables. There were also some
flow variables, particularly wage and price variables (which
were measured as time averages and had, therefore, to be
treated as flow variables), for which a second-order adjustment equation appeared to be more realistic.
We assume that the stock variables are observed at the
time points 0, 1, 2, . . . , T and the flow variables as integrals
over the intervals [0, 1], [1, 2], . . . , [T 1, T ]. We let the
elements of x2 (t) and z(t) be ordered (without loss of generality) so that

x2 (t) =
x2s (t)
f
x2 (t)

,
z(t) =
z s (t)
z f (t)

,
f
where x2s (t) is an ns2 1 vector of stock variables, x2 (t) is

f
an n2 1 vector of flow variables, ns2 + n2 = n2 , z s (t) is an

ms 1 vector of stock variables, z f (t) is an m f 1 vector of
flow variables and ms + m f = m. We then define the observable vectors x 1t , x 2t , x t and zt (t = 1, . . . , T ) by
t
x 1t =
x1 (r )dr (t = 1, . . . , T ),
(2.8)
t1
x 2t
x2s (t) x2s (t 1)

= t f
x 2 (r )dr
t1
59
(t = 1, . . . , T ),
(2.9)

x t =
1
2
zt =
x 1t
x 2t

(t = 1, . . . , T ),
{z s (t) + z s (t 1)}
t
z f (r )dr
(2.10)
(t = 1, . . . , T ).
(2.11)
t1
It has been shown (Bergstrom [1997, Theorems 4 and

5]) that if x1 (t) and x 2 (t)(0 t T ) are generated by the
model (2.1)(2.4), then, under certain conditions, the vectors x 1 , x 2 , . . . , x T , z1 , z2 , . . . , zT defined by (2.10) and (2.11)
satisfy, exactly, the system:
x 1 = G10 + G11 y(0) + E 11 z1 + E 12 z2 + E 13 z3 + 1 ,
(2.12)
x 2 = F21 x 1 + G20 + G21 y(0) + E 21 z1

+E 22 z2 + E 23 z3 + 2 ,
(2.13)
x 3 = F32 x 2 + F31 x 1 + G30 + G31 y(0)

+ E 31 z1 + E 32 z2 + E 33 z3 + 3 ,
(2.14)
x t = F1 x t1 + F2 x t2 + F0 + E 0 zt + E 1 zt1
+ E 2 zt2 + t
where
(t = 4, . . . , T ),
(2.15)
F0 , F1 , F2 , F21 , F32 , F31 , G10 , G11 , G20 , G21 , G30 , G31 ,
E 0 , E 1 , E 2 , E 11 , E 12 , E 13 , E 21 , E 22 , E 23 , E 31 , E 32 , E 33 are matrix functions of the structural parameters of the continuous

time model, which are precisely defined in Appendix A,
60
and 1 , 2 , . . . , T are unobservable random vectors such

that
E(t ) = 0 (t = 1, . . . , T ),

E t t = 0 (t = 4, . . . , T ),

= 1 (t = 5, . . . , T ),
E t t1

E t t2 = 2 (t = 6, . . . , T ),

= 3 (t = 7, . . . , T ),
E t t3

E 2 1 = 21 ,
E 2 2 = 22 ,
E 1 1 = 11 ,

E 3 2 = 32 ,
E 3 3 = 33 ,
E 3 1 = 31 ,

E 4 1 = 41 ,
E 4 2 = 42 ,
E 4 3 = 43 ,

E 5 3 = 53 ,
E 6 3 = 63 ,
E 5 2 = 52 ,

E t t
j = 0 (t = 5, . . . , T ; j = 4, . . . , T ),
where 0 , 1 , 2 , 3 and the various i j in the above

expressions are matrix functions of the structural parameters of the continuous time model that are precisely defined
in Appendix B. Sufficient conditions for the discrete observations generated by the continuous time model to satisfy, exactly, the system (2.12)(2.15) are that matrix A
defined by
A1
A = 0
A4
A2
0
A5
61
A3
I
A6
is nonsingular, a certain submatrix of e A (see Bergstrom

[1997, Assumption 3]) is nonsingular and the elements of
z(t) are polynomials in t of degree not exceeding 2. Only
the last of these conditions is very restrictive. For a model
such as that estimated in Chapter 5, the conditions on A
and e A will be satisfied except for combinations of parameters that may arise in areas of the parameter space into
which an optimisation algorithm might stray by chance
that would violate these conditions. The condition that the
elements of z(t) are polynomials in t of degree not exceeding
2 is, of course, very restrictive. However, an important property of the system (2.12)(2.15), which will be referred to
as the exact discrete model, is that the coefficient matrices
E 0 , E 1 , E 2 and E i j (i, j = 1, 2, 3) depend only on the structural parameters of the continuous time model and not on
the parameters determining the time path of z(t). Its accuracy depends, therefore, only on the accuracy with which
the continuous time path of each exogenous variable can be
approximated by a sequence of quadratic functions on the
overlapping subintervals [0, 4], [1, 5], . . . , [T 4, T ] and it
will be approximately satisfied, even if the behaviour of the
exogenous variables varies greatly over the sample period,
provided that their continuous time paths are sufficiently
smooth. Moreover, it will be exactly satisfied if the model is
62
closed, that is if the matrices B1 and B2 in the continuous

time model are zero matrices. It has been shown (Bergstrom
[1997, Theorem 6]) that the system (2.12)(2.15) can also
be represented in the form of a VARMAX model:
x 1 G10 G11 y(0) E 11 z1 E 12 z2 E 13 z3 = M11 1 ,
(2.16)
x 2 F21 x 1 G20 G21 y(0) E 21 z1 E 22 z2 E 23 z3

= M21 1 + M22 2 ,
(2.17)
x 3 F32 x 2 F31 x 1 G30 G31 y(0) E 31 z1 E 32 z2 E 33 z3

= M31 1 + M32 2 + M33 3 ,
(2.18)
x t F1 x t1 F2 x t2 F0 E 0 zt E 1 zt1 E 2 zt2
= Mtt t + Mt,t1 t1 + Mt,t2 t2 + Mt,t3 t3
(t = 4, . . . , T ),
(2.19)
where the random vectors t (t = 1, . . . , T ) satisfy the conditions

E(t ) = 0 (t = 1, . . . , T ),

E t t = I (t = 1, . . . , T ),

E s t = 0 (s = t; s, t = 1, . . . , T ),
(2.20)
(2.21)
(2.22)
and the coefficient matrices M11 , M21 , M22 , M31 , M32 , M33 ,
Mtt , Mt,t1 , Mt,t2 and Mt,t3 (t = 4, . . . , T ) are obtained,
recursively, from (2.23)(2.44).

M11 M11
= 11 ,
63
(2.23)
1
M21 = 21 M11
,
(2.24)

M22 M22
= 22 M21 M21
,
(2.25)
1
M31 = 31 M11
,
(2.26)

1

M32 = 32 M31 M21
,
M22
(2.27)

M33 M33
= 33 M31 M31
M32 M32
,
(2.28)
1
M41 = 41 M11
,
(2.29)
1

M22
M42 = 42 M41 M21
,
(2.30)

1

M42 M32
,
M33
M43 = 43 M41 M31
(2.31)

M44 M44
= 0 M41 M41
M42 M42
M43 M43
,
(2.32)
1
M52 = 52 M22
,
(2.33)

1

M53 = 53 M52 M32
,
M33
(2.34)

1

M54 = 1 M52 M42
M53 M43
,
M44
(2.35)
64

M55 M55
= 0 M52 M52
M53 M53
M54 M54
,
(2.36)
1
,
M63 = 63 M33
(2.37)

1

,
M64 = 2 M63 M43
M44
(2.38)

1

M64 M54
,
M65 = 1 M63 M53
M55
(2.39)

= 0 M63 M63
M64 M64
M65 M65
,
M66 M66
(2.40)
and, for t = 7, . . . , T,

1
Mt,t3 = 3 Mt3,t3
,
(2.41)

1

,
Mt,t2 = 2 Mt,t3 Mt2,t3
Mt2,t2
(2.42)

1

Mt,t2 Mt1,t2
,
Mt,t1 = 1 Mt,t3 Mt1,t3
Mt1,t1
(2.43)

Mtt Mtt = 0 Mt,t3 Mt,t3
Mt,t2 Mt,t2
Mt,t1 Mt,t1
,
(2.44)
the matrices Mtt (t = 1, . . . , T ) being lower triangular

matrices.
The VARMAX model (2.16)(2.19) is more general than
the standard VARMAX model considered in the statistical
literature in that the moving average coefficient matrices
65
Mtt , Mt,t1 , Mt,t2 and Mt,t3 are time dependent. It can be

shown, however, that they converge to constant matrices
as t (see Bergstrom [1990, Ch. 7, Theorem 1] and the
comments before and after that theorem).
As we shall see in this chapter, the VARMAX model
(2.16)(2.19) provides a very convenient basis for both the
Gaussian estimation of the structural parameters of the continuous time model (2.1)(2.4) and the computation of the
optimal post-sample forecasts of the discrete observations
generated by that model. It is also of considerable interest in itself and, in particular a link with the extensive
recent literature on discrete time models with unit roots
or stochastic trends (see, especially, Phillips and Durlauf
[1986], Phillips [1987, 1991a, 1995], Engle and Granger
[1987], Johansen [1988, 1991], Park and Phillips [1988,
1989], Franses [1996]. In this connection, it should be noted
that except when one or more variables are affected, either
directly or indirectly (via other variables), by the stochastic trends, all of the other elements of the vector x t are
integrated of order 1. Then, provided that the k (n1 + n2 )
matrix [C 1 C 2 ] has rank k, there are n1 + n2 k linearly
independent cointegrating vectors (see Bergstrom [1997,
Appendix B]).
66
2.4 Estimation and Forecasting

Before deriving the Gaussian likelihood function, it will be
useful to review the complete set of parameters to be calculated. The vector of basic structural parameters and the
vector of drift parameters have, already, been explicitly
introduced in (2.1)(2.3), and these parameters will affect
all of the matrices and vectors of coefficients on both the
left- and right-hand sides of (2.16)(2.19). It is convenient,
at this stage, to introduce the (n1 + 2n2 ) (n1 + 2n2 ) matrix
defined by
11
= 0
21
0
0
0
12
0 .
22
(2.45)
This matrix and the matrix 33 both occur in formulae for
the i and i j matrices in Appendix B. We shall parametrize
and 33 by writing them as () and 33 (), where
and are vectors of structural parameters. If there are no
restrictions on these matrices, apart from those implied by
Assumption 1, then will have (n1 + n2 )(n1 + n2 + 1)/2 elements and will have k(k + 1)/2 elements.
In addition to the vector [, , , ] of genuine parameters, most of the initial state vector y(0) is unobservable
67
and must be treated as part of the complete vector of parameters to be estimated. The observable part of y(0) is the ns2 1
vector x 2s (0), that is the initial levels of the stock variables.
f
The unobservable part of y(0) is the (n1 + n2 + n2 ) 1 vector yn defined by
x 1 (0)
yn =
x2 (0) .
D x 2 (0)
(2.46)
Because the vector yn occurs in only the first three equations

of the VARMAX model (2.16)(2.19), we cannot expect to
obtain a consistent estimate of this vector. But, its inclusion
as part of the complete vector of parameters to be estimated
could have a significant effect on the estimate of [, , , ]
obtained from a small sample. We shall treat the vector
[, , , , yn ] , therefore, as the complete vector of parameters to be estimated.
The Gaussian likelihood function can be derived following the general approach of Bergstrom [1983, 1985, 1986,
1997] from (2.16)(2.19). This system can be regarded as
a linear transformation from the vector = [1 , 2 , . . . , T ]
to the vector x = [x 1 , x 2 , . . . , x T ] , and the Jacobian of the
T

transformation is
t=1 |Mtt |. Hence, if L(, , , , yn )
68
denotes minus twice the logarithm of the likelihood function (less a constant), then
T

L=
t t + 2log|Mtt | ,
(2.47)
t=1
where 1 , 2 , . . . , T are regarded as functions of the observations and the parameters and can be computed, recursively,
from (2.16)(2.19), that is, from the recursive equations
1
1 = M11
(x 1 G10 G11 y(0) E 11 z1 E 12 z2 E 13 z3 ),
(2.48)
1
(x 2 F21 x 1 G20 G21 y(0)
2 = M22
E 21 z1 E 22 z2 E 23 z3 M21 1 ),
(2.49)
1
3 = M33
(x 3 F32 x 2 F31 x 1 G30 G31 y(0)
E 31 z1 E 32 z2 E 33 z3 M31 1 M32 2 ),
(2.50)
t = Mtt1 ( x t F1 x t1 F2 x t2 F0 E 0 zt E 1 zt1
E 2 zt2 Mt,t1 t1 Mt,t2 t2 Mt,t3 t3 )
(t = 4, . . . , T ).
(2.51)
The exact Gaussian estimates of the parameters are

obtained by maximizing L with respect to (, , , , yn ).
These will be exact maximum likelihood estimates if the
innovations in the continuous time model (2.1)(2.4) are
Brownian motion (in which case the innovation vectors
69
1 , 2 , . . . , T in the VARMAX model (2.16)(2.19) will be

normally distributed) and the continuous time paths of the
exogenous variables are quadratic functions of t. Moreover,
they can be expected to provide very good estimates under
much more general circumstances (see the comments in
Bergstrom [1997, p. 486]).
The above estimation procedure can also be applied
to continuous time models containing the first and second derivatives of the exogenous variables. For this purpose, the vector z(t) is replaced by the vector z (t) =
[z (t), D z (t), D 2 z (t)] and a vector zt is defined in terms
of z (t) by an equation analogous to (2.11). Although zt is
not directly observable, it can be obtained from zt , zt1 and
zt2 using quadratic interpolation formulae given in Nowman [1991] (cf. Chambers [1991]), these formulae being
exact when the elements of z(t) are quadratic functions
of t.
As was noted in the previous section, the elements of
the vector x t are, in general, integrated of order 1 with
n1 + n2 k cointegrating vectors. Consequently, the classical theory of inference is not applicable to the exact Gaussian
estimates discussed above, even when these are exact maximum likelihood estimates.
70
In this connection, it should be noted, first, that the

behaviour of the estimates as T will differ as between
various types of parameter. In a model such as that discussed
in the remaining chapters of this monograph, the structural
parameter vector will include both long-run parameters
and speed of adjustment parameters. Whereas the estimates
of the speed of adjustment parameters will converge (in
probability) at the rate T 1/2 to the true values as T ,
the estimates of some of the long-run parameters (those
that are uniquely related to the cointegrating vectors) will
converge at the rate T .
It should be noted, also, that unless C 1 and C 2 are severely
restricted, the drift parameters (the elements of ) are not
identifiable. In fact, in a typical macroeconometric model,
all or nearly all of the elements of C 1 and C 2 will be known
constants. For example, the stochastic trend representing
technical progress will enter the employmentoutput relation and the pricewage relation (each formulated in the
logarithms of the variables) with the coefficient 1.
The distribution of the estimates discussed above will also
be different from the distribution of parameter estimates
obtained under classical assumptions. Whereas, under classical assumptions, the deviations of the estimates from the
71
true values of the parameters (each deviation being multiplied by T 1/2 ) are, in general, asymptotically normally distributed, the deviations of the above estimates from the
true values of the parameters (each deviation being multiplied by an appropriate power of T ) are asymptotically
distributed like very complicated expressions involving integrals of quadratic forms of Brownian motions and integrals
of Brownian motions with respect to Brownian motions.
This is as we should expect in view of the extensive work
of P. C. B. Phillips and his collaborators on the estimation of models with integrated processes (see Phillips and
Durlauf [1986], Phillips [1987, 1991a, 1991b, 1995], Park
and Phillips [1988, 1989]). The computational cost of using
such expressions in assessing the accuracy of the parameter estimates would, however, be very heavy except in very
small models.
In assessing the accuracy of the parameter estimates for
our UK model presented in Chapter 5, we will follow a
simpler Bayesian approach. In this connection, it should
be noted, first, that, if the innovations in the continuous
time model (2.1)(2.4) are Brownian motion and the exogenous variables are quadratic functions of t, then assuming a uniform prior distribution, L(, , , , yn ) is minus
twice the logarithm of the Bayesian posterior probability
72
density function of [, , , , yn ] (less a constant). Then let ,

,
y n ) denote the exact Gaussian estimator of
ting (,
,
,
y n ) ( , ,
,
y n )]1 is the
(, , , , yn ), 2[ 2 L/( , ,
,
,
covariance matrix of a normal approximation to this
Bayesian posterior probability density function in the
neighbourhood of its mode. In Chapter 5, we, formally,
present the square roots of the diagonal elements of
,
,
y n ) ( , ,
,
y n )]1 as the Standard
,
,
2[ 2 L/(,
Errors of the parameter estimates, although it must be
remembered that they should be interpreted in this approximate Bayesian sense.
A final point to be noted in connection with the estimation procedure described above is that the formulation of the
algorithm mainly in differences (rather than levels) of the
variables does not result in any loss of information or efficiency. All the information in the sample is contained in the
observable part x2s (0) of the initial state vector and the observations of the random vectors x 1 , x 2 , x 3 , x 4 , . . . , x T , that
is, the set of vectors in terms of which the estimation algorithm is formulated. Moreover, there is a one-to-one correspondence between the vectors [ x 1 , x 2 , x 3 , x 4 , . . . , x T ]
and [x 1 , x 2 , x 3 , x 4 , . . . , x T ]. It makes no difference to the
estimates, therefore, whether the likelihood function is
derived from the probability density function (p.d.f.) of the
73
former vector or the p.d.f. of the latter vector. The estimates

obtained by maximizing the two alternative forms of the
likelihood function with respect to the unknown parameters are exactly the same. When the innovations in the continuous time model (2.1)(2.4) are Brownian motion and
the elements of z(t) are quadratic functions of t, the estimates obtained by maximizing the function L defined by
(2.47) with respect to the parameter vector (, , , , yn )
are exact maximum likelihood estimates taking account of
all information in the sample and all restrictions implied by
the continuous time model.
The VARMAX model (2.16)(2.19) provides a very convenient basis, not only for the estimation of the parameters
of the continuous time model, but also for the generation of
optimal post-sample forecasts of the discrete observations.
The procedure to be followed is similar to that described
in Bergstrom [1989] for obtaining optimal forecasts from
a higher order continuous time model without stochastic
trends.
Suppose that we wish to obtain forecasts of x t for R postsample periods, that is, for t = T + 1, . . . , T + R.
Extending the VARMAX model (2.16)(2.19) up to the
period T + R and separating the autoregressive and moving
average parts, we have the system (2.52)(2.59).
74
x 1 = G10 + G11 y(0) + E 11 z1 + E 12 z2 + E 13 z3 + 1 ,
(2.52)
x 2 = F21 x 1 + G20 + G21 y(0) + E 21 z1 + E 22 z2 + E 23 z3 + 2 ,

(2.53)
x 3 = F32 x 2 + F31 x 1 + G30 + G31 y(0)
+ E 31 z1 + E 32 z2 + E 33 z3 + 3 ,
(2.54)
x t = F1 x t1 + F2 x t2 + F0 + E 0 zt + E 1 zt1
+ E 2 zt2 + t
(t = 4, . . . , T + R),
(2.55)
1 = M11 1 ,
(2.56)
2 = M21 1 + M22 2 ,
(2.57)
3 = M31 1 + M32 2 + M33 3 ,
(2.58)
t = Mtt t + Mt,t1 t1 + Mt,t2 t2 + Mt,t3 t3

(t = 4, . . . , T + R).
(2.59)
,
,
From this system, together with the estimator [,
y n ] of the parameter vector, we can obtain a set of opti,

mal forecasts of the vectors x T +1 , x T +2 , . . . , x T +R . They are
optimal in the sense that they will be exact maximum likelihood estimates of the conditional expectations of these vectors, conditional on all information in the sample, when the
75
innovations in the continuous time model (2.1)(2.4) are

Brownian motion and the elements of z(t) are quadratic
functions of t.
To show this we, first, note that, under the latter conditions, the conditional expectations of the vectors T +1 ,
T +2 , . . . , T +R , conditional on all information in the sample are zero vectors, that is,
E (T +r /x 1 , x 2 , . . . , x T ) = 0
(r = 1, . . . , R).
(2.60)
Now let T +1 , T +2 , . . . , T +R denote the exact Gaussian

estimates of the conditional expectations of T +1 , T +2 , . . . ,
T +R conditional on all information in the sample. Then,
using (2.59), we obtain
T +1 = M T +1,T T + M T +1,T 1 T 1 + M T +1,T 2 T 2 , (2.61)
T +2 = M T +2,T T + M T +2,T 1 T 1 ,
(2.62)
T +3 = M T +3,T T ,
(2.63)
T +r = 0
(r = 4, . . . , R),
(2.64)
where the matrices M T +1,T , M T +1,T 1 , M T +1,T 2 , M T +2,T ,

M T +2,T 1 and M T +3,T are obtained from the extension of
the recursive system (2.41)(2.44) up to t = T + 3 and the
estimated values of the parameters, that is,
76

1

3 M t3,t3
,
M t,t3 =
(2.65)

1

2 M t,t3 M t2,t3
M t,t2 =
M t2,t2
,
(2.66)

1

1 M t,t3 M t1,t3
M t1,t1
M t,t2 M t1,t2
,
M t,t1 =
(2.67)

0 M t,t3 M t,t3
M tt M tt =
M t,t2 M t,t2
M t,t1 M t,t1
(2.68)
(t = T + 1, . . . , T + 3).
The matrices M T 2,T 2 , M T 1,T 2 , M T 1,T 1 , M T,T 2 , M T,T 1 ,

and M T T in (2.65)(2.68) will have been computed in the
estimation procedure.
Next, let x T +1 , x T +2 , . . . , x T +R denote the exact
Gaussian estimators of the conditional expectations of the
vectors x T +1 , x T +2 , . . . , x T +R conditional on all information in the sample with the vectors z T +1 , zT +2 , . . . , zT +R
being treated as known. Then, using (2.55), we obtain
x T + 1 = F 1 x T + F 2 x T 1 + F 0 + E 0 z T +1 + E 1 z T
+ E 2 z T 1 + T +1 ,
(2.69)
x T +1 + F 2 x T + F 0 + E 0 z T +2
x T + 2 = F 1
+ E 1 z T +1 E 2 z T + T +2 ,
(2.70)
77
x T +3 = F 1 x T +2 + F 2 x T +1 + F 0 + E 0 z T +3
+ E 1 z T +2 E 2 z T +1 + T +3 ,
(2.71)
x T +r = F 1 x T +r 1 + F 2 x T +r 2 + F 0 + E 0 z T +r
+ E 1 z T +r 1 + E 2 z T +r 2 (r = 4, . . . , R), (2.72)
where F 0 , F 1 , F 2 , E 0 , E 1 and E 2 are obtained by substitut of the parameter vector [, ] into the
ing the estimate [ , ]
formulae for F0 , F1 , F2 , E 0 , E 1 and E 2 given in Appendix A.
Finally, we obtain the forecasts of the vectors x T +1 ,
x T +2 , . . . , x T +R from the accumulation formula
x T + r = x T +
r

x T +s
(r = 1, . . . , R).
(2.73)
s =1
The forecasts given by (2.73) are optimal post-sample forecasts of the discrete observations of the endogenous variables
in the sense mentioned above.
Summarizing the forecasting procedure, the main computational steps are as follows.
1. Compute the matrices M T +1,T , M T +1,T 1 , M T +1,T 2 ,
M T +2,T , M T +2,T 1 and M T +3,T from (2.65)(2.68).
2. Compute the vectors T +1 , T +2 and T +3 from (2.61)
(2.64).
78
3. Compute the vectors x T +1 , x T +2 , . . . , x T +R from

(2.69)(2.72).
4. Compute the vectors x T +1 , x T +2 , . . . , x T +R from (2.73).
In the above forecasting procedure, the post-sample
observations of the exogenous variables have been treated
as known. In practice they would, of course, be unknown at
the time when the post-sample forecasts of the endogenous
variables are made, and they would have to be estimated
independently. In Chapter 5 we will, however, set aside the
last 2 years (8 quarters) observations in the sample for the
purpose of testing the predictive performance of the model
(the parameters being estimated from the earlier observations) so that the true values of the exogenous variables
used in the computation of the forecasts will be known.
2.5 Conclusion
In this chapter, we have formulated an open mixed-order
continuous time dynamic model with mixed stock and flow
variables and unobservable stochastic trends. We have then
described a form of VARMAX model which, under certain
assumptions, is satisfied, exactly, by the discrete observations
generated by the continuous time model, and we have
79
shown how to compute the coefficient matrices of this

model for any given values of the parameters of the continuous time model. Finally, we have shown how to use the
VARMAX model to compute the exact Gaussian estimates of
the parameters of the continuous time model and optimal
forecasts of the post-sample discrete observations.
When the innovations in the continuous time model are
Brownian motion and the exogenous variables are quadratic
functions of the time parameter, the computational procedures described produce exact maximum likelihood estimates of the parameters of the continuous time model, and
the forecasts are exact maximum likelihood estimates of the
conditional expectations of the post-sample discrete observations generated by the continuous time model. Moreover, both the parameter estimates and the forecasts can
be expected to be very accurate under much more general
conditions.
Appendix A: Formulae for the Coefficient

Matrices of Exact Discrete Model
In this Appendix we present formulae for the coefficient
matrices of the discrete time model (2.12)(2.15) derived
in Bergstrom [1997]. These formulae make extensive use of
80
certain selection matrices, each of which transforms a vector

of appropriate dimension into a vector of lower dimension,
whose elements are a selection of the elements of the original vector. The selection matrices S1 , S2 , S3 , S4 , S5 and S are
defined as follows.
S1 = [I 0

0 0
S2 =
0 0

0 I
S3 =
0 0
0 0 0],

0 I 0
,
I 0 0

0 0 0
,
0 0 I
(2.74)
S4 = [I 0],
S5 = [0 I ],

S1
.
S=
S2
S1 is an n1 (n1 + 2n2 ) matrix with an n1 n1 identity

matrix in columns 1 to n1 . S2 is an n2 (n1 + 2n2 ) matrix
with an ns2 ns2 identity matrix in rows 1 to ns2 and columns
n1 + n2 + 1 to n1 + n2 + ns2 identity matrix in rows ns2 + 1
to n2 and columns n1 + ns2 + 1 to n1 + n2 . S3 is an n2
(n1 + 2n2 ) matrix with an ns2 ns2 identity matrix in rows 1
f
to ns2 and columns n1 + 1 to n1 + ns2 and an n2 + n2 identity

matrix in rows ns2 + 1 to n2 and columns n1 + n2 + ns2 + 1 to
n1 + 2n2 . S4 is an ms m matrix with an ms ms identity
81
matrix in columns 1 to ms . S5 is an m f m matrix with an

m f m f identity matrix in columns ms + 1 to m.
We shall, also, use matrix A defined in Section 3, matrix
W defined by

1
W = Se A S3 S3 Se A S3 S3 e A S3 S2 e A S3
S2 ,
and the (n1 + 2n2 ) 1 vector b, the (n1 + 2n2 ) m matrix

B and the (n1 + 2n2 ) k matrix C defined by

B1
C1
b1
b = 0 , B = 0 , C = 0 .
b2
B2
C2
The formulae for the coefficient matrices of the discrete
time model (2.12)(2.15) are as follows:
F0 = (S + W)A1 (e A I )C ,

1
,
F1 = Se A S1 , Se A S2 + Se A S3 S3 e A S3 S2 e A S3
F2 = We A S ,
F21 = Se A S ,
F31 = Se A S3 S3 e A S ,
F32 = F21 ,
G10 = S{[A2 (e A I) A1 ]b+[A3 (e A I ) A2 0.5A1 ]C },
G11 = SA1 (e A I ),
82

G20 = S A1 (e A I ) + e A S3 S3 [A2 (e A I ) A1 ]b

+ S A2 (e A I ) + 0.5A1 (e A I ) A1

+ e A S3 S3 [A3 (e A I ) A2 0.5A1 ] C ,
G21 = Se A S3 S3 A1 (e A I ),

G30 = S I + e A S3 S3 A1 (e A I )

+ e A S3 S3 e A S3 S3 [A2 (e A I ) A1 ] b

+ S (A2 + 1.5A1 )(e A I ) A1
+ e A S3 S3 [(A2 + 0.5A1 )(e A I ) A1 ]

+ e A S3 S3 e A S3 S3 [A3 (e A I ) A2 0.5A1 ] C ,
G31 = Se A S3 S3 e A S3 S3 A1 (e A I ),

E 0 = F1 SL10 + F2 SL20 SL00 ,
E 1 = F1 SL11 + F2 SL21 SL01 ,
E 2 = F1 SL12 + F2 SL22 SL02 ,
E 11 = G11 P1 SL22 ,
E 12 = G11 P2 SL21 ,
E 13 = G11 P3 SL20 ,
E 21 = G21 P1 + F21 SL22 SL12 ,
E 22 = G21 P2 + F21 SL21 SL11 ,
E 23 = G21 P3 + F21 SL20 SL10 ,
E 31 = G31 P1 + F32 SL12 + F31 SL22 SL02 ,
E 32 = G31 P2 + F32 SL11 + F31 SL21 SL01 ,
E 33 = G31 P3 + F32 SL10 + F31 SL20 SL00 ,
83
where
11
3
S
4
= A1 B 12 + A2 B + A3 B,
2
S5
1
S
4
= A1 B 6 2A2 B 2A3 B,
0
1
1
S
4
= A1 B 12 + A2 B + A3 B,
2
0
L 00
L 01
L 02
1
1
S
4
= A1 B
12 + A2 B + A3 B,
2
0
7
S
4
= A1 B 6 2A3 B,
S5
1
1
S
4
= A1 B 12 A2 B + A3 B,
2
0
L 10
L 11
L 12
L 20
L 21
1
1
S
4
= A1 B
12 A2 B + A3 B,
2
0
1
S
4
= A1 B 6 + 2A2 B 2A3 B,
0
84
11
3
S
4
L 22 = A1 B 12 A2 B + A3 B,
2
S5
7
S4
4
P1 = A1 B
2A2 B + A3 B,
11
S5
6
S4
P2 = A1 B 7 + 3A2 B 2A3 B,
S5
6
1
S4
4
P3 = A1 B
A2 B + A3 B.
1
S5
3
Appendix B: Formulae for the

Autocovariance Matrices
In this Appendix we present formulae for the autocovariance matrices of the unobservable random vectors
1 , 2 , . . . , T in the discrete time model (2.12)(2.15), that
is for the matrices 0 , 1 , 2 , 3 and the various matrices i j defined below (2.12)(2.15). It has been shown
(Bergstrom [1997, Theorems 1 and 2]) that the vectors
1 , 2 , . . . , T in the discrete time model are related to the
85
white noise innovations in the continuous time model (2.1)

(2.4) by the following stochastic integral formulae:
t =
K 1 (t r ) (dr ) +
t1

+
t1
K 2 (t 1 r ) (dr )
t2
t2
t3

+
N1 (t r )3 (dr ) +
t1
N2 (t 1 r )3 (dr )
t2
t2
1 =
1
K 11 (1 r ) (dr ) +
N11 (1 r )3 (dr ),
K 21 (1 r ) (dr ) +
K 22 (2 r ) (dr )
K 31 (1 r ) (dr ) +
N21 (1 r )3 (dr ) +
(t = 4, . . . , T )
3 =
N4 (t 3 r )3 (dr )
t4

t3
N3 (t 2 r )3 (dr ) +
t3
2 =
K 4 (t 3 r ) (dr )
t4
t1
t3
K 3 (t 2 r ) (dr ) +
N22 (2 r )3 (dr ),
1
2
K 32 (2 r ) (dr )

+
K 33 (3 r ) (dr ) +
N31 (1 r )3 (dr )
0
2

N32 (2 r )3 (dr ) +
86
N33 (3 r )3 (dr ).
Formulae expressing the weighting functions K 1 , K 2 , K 3 ,

K 4 , K 11 , . . . , K 33 , N1 , N2 , N3 , N4 , N11 , . . . , N33 in terms of
the parameters of the continuous time model are presented
in Bergstrom [1997 Theorems 1 and 2], but will not be
reproduced here. Instead, we present formulae expressing
0 , 1 , 2 , 3 and the various matrices i j as sums of integrals involving combinations of these weighting functions,
followed by formulae expressing these integrals in terms of
the parameters of the continuous time model. The latter formulae, which are previously unpublished, greatly, facilitate
the computation of the VARMAX representation of the discrete time model and the value of the likelihood function
for a given set of values of the parameters of the continu
ous time model. The formulae, which involve the matrix
defined by (2.45), are as follows:
0 =
0
K 1 (r ) K 1 (r )dr +
+
0

+
0
1
+
0
1
0
K 3 (r ) K 3 (r )dr +
N1 (r )33 N1 (r )dr
N3 (r )33 N3 (r )dr
K 2 (r ) K 2 (r )dr
K 4 (r ) K 4 (r )dr
1

+
87
N2 (r )33 N2 (r )dr

N4 (r )33 N4 (r )dr ,
where
1
0
K 1 (r ) K 1 (r )dr
= SA1 {SA1 } {SA2 (e A I )}{SA1 } {SA1 }

1

2 A

1
e rA e rA dr {SA1 } ,
{SA (e I )} + {SA }
0
1
0
K 2 (r ) K 2 (r )dr
= {SA1 (I + e A) WA1 }{SA1 (I + e A) WA1 }

+ {(W 2S)A2 (e A I )}{SA1 (I + e A) WA1 }
+ {SA1 (I + e A) WA1 }{(W 2S)A2 (e A I )}
1

1
+ {(W 2S)A }
e rA e rA dr {(W 2S)A1 } ,
0
1
0
K 3 (r ) K 3 (r )dr
= {WA1 (I + e A) SA1 e A}{WA1 (I + e A) SA1 e A}

+ {(S 2W)A2 (e A I )}{WA1 (I + e A) SA1 e A}
+ {WA1 (I + e A) SA1 e A}{(S 2W)A2 (e A I )}
1

+ {(S 2W)A1 }
e rA e rA dr {(S 2W)A1 } ,

0
0
1
K 4 (r ) K 4 (r )dr
= {WA1 e A}{WA1 e A} {WA2 (e A I )}
88
{WA1 e A} {WA1 e A}

{WA2 (e A I )} + {WA1 }
e rA e rA dr {WA1 } ,
1
0
N1 (r )33 N1 (r )dr = {SA2 }C 33 C {SA2 }

1
+ {SA1 }C 33 C {SA2 }
2
1
+ {SA2 }C 33 C {SA1 }
2
1
+ {SA1 }C 33 C {SA1 }
3
{SA3 (e A I )}C 33 C {SA2 }
{SA2 }C 33 C {SA3 (e A I )}

1
2
rA
SA
r e dr C 33 C {SA1 }
0

0
1
{SA1 }C 33 C {SA2
r e rA dr }
0
1

+ {SA2 } e rA C 33 C e rA dr {SA2 } ,
1
N2 (r )33 N2 (r )dr
= {SA2 (I + e A) SA1 WA2 }

C 33 C {SA2 (I + e A) SA1 WA2 }
1
+ {SA1 (I + e A) WA1 }
2
C 33 C {SA2 (I + e A) SA1 WA2 }
89
1
+ {SA2 (I + e A) SA1 WA2 }
2
C 33 C {SA1 (I + e A) WA1 }
1
+ {SA1 (I + e A) WA1 }
3
C 33 C {SA1 (I + e A) WA1 }
+ {(W 2S)A3 (e A I )}
C 33 C {SA2 (I + e A) SA1 WA2 }
+ {SA2 (I + e A) SA1 WA2 }
+ C 33 C {(W 2S)A3 (e A I )}

1
+ (W 2S)A2
r e rA dr
0
C 33 C {SA (I + e A) WA1 }

+ {SA1 (I + e A) WA1 }

2
C 33 C (W 2S)A
2
rA
r e dr
+ {(W 2S)A }
e rA C 33 C e rA dr {(W 2S)A2 } ,
0
1
0
N3 (r )33 N3 (r )dr
= {S(A1 A2 )e A + WA2 (I + e A)
WA1 }C 33 C {S(A1 A2 )e A
+ WA2 (I + e A) WA1 }
1
+ {WA1 (I + e A) SA1 e A}
2
90
C 33 C {S(A1 A2 )e A + WA2 (I + e A)

1
WA1 } + {S(A1 A2 )e A
2
2
+ WA (I + e A) WA1 }
C 33 C {WA1 (I + e A) SA1 e A}
1
+ {WA1 (I + e A) SA1 e A}
3
C 33 C {WA1 (I + e A) SA1 e A}
+ {(S 2W)A3 (e A I )}C 33 C {S(A1 A2 )e A
+ WA2 (I + e A) WA1 } + {S(A1 A2 )e A
+ WA2 (I + e A) WA1 }
C 33 C {(S 2W)A3 (e A I )}

1
2
rA
+ (S 2W)A
r e dr
0
C 33 C {WA1 (I + e A) SA1 e A}

+ {WA1 (I + e A) SA1 e A}

1
r e rA dr
C 33 C (S 2W)A2
+ {(S 2W)A2 }
0
1
e rA C 33 C e rA dr {(S 2W)A2 } ,
1
0
N4 (r )33 N4 (r )dr
= {W(A1 A2 )e A}C 33 C {W(A1 A2 )e A}

1
{WA1 e A}C 33 C {W(A1 A2 )e A}
2
91
1
{W(A1 A2 )e A}C 33 C {WA1 e A}
2
1
+ {WA1 e A}C 33 C {WA1 e A}
3
+ {WA3 (e A I )}C 33 C {W(A1 A2 )e A}
+ {W(A1 A2 )e A}C 33 C {WA3 (e A I )}

1
WA2
r e rA dr C 33 C {WA1 e A}
0
1 A
{WA e }C 33 C
+ {WA2 }
1
rA
WA
r e dr
0
e rA C 33 C e rA dr {WA2 } .
1 =
K 2 (r ) K 1 (r )dr +

+
+
0
K 4 (r ) K 3 (r )dr +
N3 (r )33 N2 (r )dr
K 3 (r ) K 2 (r )dr
N2 (r )33 N1 (r )dr

1
+
0
N4 (r )33 N3 (r )dr,
where
1
K 2 (r ) K 1 (r )dr = {SA1 (I + e A) WA1 }{SA1 }
0
{(W 2S)A2 (e A I )}{SA1 }

+ {SA1 (I + e A)WA1 }{SA2 (e A I )}
1

+ {(W 2S)A1 }
0
92
K 3 (r ) K 2 (r )dr
= {WA1 (I + e A) SA1 e A}{SA1 (I + e A) WA1 }

+ {(S 2W)A2 (e A I )}{SA1 (I + e A) WA1 }
+ {WA1 (I + e A) SA1 e A}{(W 2S)A2 (e A I )}
1

+ {(W 2S)A1 }
0

0
K 4 (r ) K 3 (r )dr
= {WA1 e A}{WA1 (I + e A) SA1 e A}

+ {WA2 (e A I )}{WA1 (I + e A) SA1 e A}
{WA1 e A}{(S 2W)A2 (e A I )} + {WA1 }
1
e rA e rA dr {(S 2W)A1 } ,
0

0
N2 (r )33 N1 (r )dr
= {SA2 (I + e A) SA1 WA2 }C 33 C {SA2 }

1
{SA1 (I + e A) WA1 }C 33 C {SA2 }
2
1
{SA2 (I + e A) SA1 WA2 }C 33 C {SA1 }
2
1
{SA1 (I + e A) WA1 }C 33 C {SA1 }
3
{(W 2S)A3 (e A I )}C 33 C {SA2 }
+ {SA2 (I + e A) SA1 WA2 }C 33 C {SA3 (e A I )}
93

(W 2S)A2

r e rA dr C 33 C {SA1 }
0
1
+ {SA (I + e ) WA }C 33 C
A
+ {(W 2S)A2 }

SA
1
rA
r e dr
0
r e rA C 33 C r e rA dr {SA2 } ,
1
0
N3 (r )33 N2 (r )dr
= {S(A1 A2 )e A + WA2 (I + e A) WA1 }

C 33 C {SA2 (I + e A) SA1 WA2 }
1
+ {WA1 (I + e A) SA1 e A}
2
C 33 C {SA2 (I + e A) SA1 WA2 }
1
+ {S(A1 A2 )e A + WA2 (I + e A) WA1 }
2
C 33 C {SA1 (I + e A) WA1 }
1
+ {WA1 (I + e A) SA1 e A}
3
C 33 C {SA1 (I + e A) WA1 }
+ {(S 2W)A3 (e A I )}
C 33 C {SA2 (I + e A) SA1 WA2 }
+ {S(A1 A2 )e A + WA2 (I + e A) WA1 }
C 33 C {(W 2S)A3 (e A I )}

1
2
rA
+ (S 2W)A
r e dr
0
94
C 33 C {SA1 (I + e A) WA1 }

+ {WA1 (I + e A) SA1 e A}

1

2
rA
C 33 C (W 2S)A
r e dr
+ {(S 2W)A2 }
r e rA
0
C 33 C e rA dr {(W 2S)A2 } ,

0
N4 (r )33 N3 (r )dr
= {W(A1 A2 )e A}C 33 C {S(A1 A2 )

1
e A + WA2 (I + e A) WA1 } {WA1 e A}
2
C 33 C {S(A1 A2 )e A +WA2 (I + e A) WA1 }
1
+ {W(A1 A2 )e A}C 33 C {WA1 (I + e A)SA1 e A}
2
1
{WA1 e A}C 33 C {WA1 (I + e A) SA1 e A}
3
+ {WA3 (e A I )}C 33 C {S(A1 A2 )e A
+ WA2 (I + e A) WA1 } + {W(A1 A2 )e A}
C 33 C {(S 2W)A3 (e A I )}

1
+ WA2
r e rA dr C 33 C {WA1 (I + e A) SA1 e A}
0
1 A
{WA e }C 33 C
+ {WA2 }
(S 2W)A
1
rA
r e dr
0
e rA C 33 C e rA dr {(S 2W)A2 } .
95
K 4 (r ) K 2 (r )dr
0
0
1
1

+
N3 (r )33 N1 (r )dr +
N4 (r )33 N2 (r )dr,
2 =
K 3 (r ) K 1 (r )dr
where
1
0
K 3 (r ) K 1 (r )dr
= {WA1 (I + e A) SA1 e A}{SA1 }

{(S 2W)A2 (e A I )}{SA1 }
+ {WA1 (I + e A) SA1 e A}{SA2 (e A I )}
1

1
+ {(S 2W)A }
0

0
K 4 (r ) K 2 (r )dr
= {WA1 e A}{SA1 (I + e A) WA1 }

+ {WA2 (e A I )}{SA1 (I + e A) WA1 }
{WA1 e A}{(W 2S)A2 (e A I )}
1

+ {WA1 }
0
1
0
N3 (r )33 N1 (r )dr
= {S(A1 A2 )e A + WA2 (I + e A) WA1 }

1
C 33 C {SA2 } {WA1 (I + e A) SA1 e A}
2
96
1
{S(A1 A2 )e A
2
+ WA2 (I + e A) WA1 }C 33 C {SA1 }
1
{WA1 (I + e A) SA1 e A}C 33 C {SA1 }
3
{(S 2W)A3 (e A I )}C 33 C {SA2 }
C 33 C {SA2 }
+ {S(A1 A2 )e A + WA2 (I + e A) WA1 }

1
r e rA dr
C 33 C {SA3 (e A I )} (S 2W)A2
1
0
1 A
C 33 C {SA } + {WA (I + e ) SA e }

1
r e rA dr + {(S 2W)A2 }
C 33 C SA2
A
r e rA C 33 C e rA dr {SA2 } ,
1
0
N4 (r )33 N2 (r )dr
= {W(A1 A2 )e A}C 33 C {SA2 (I + e A) SA1 WA2 }

1
{WA1 e A}C 33 C {SA2 (I + e A) SA1 WA2 }
2
1
+ {W(A1 A2 )e A}C 33 C {SA1 (I + e A) WA1 }
2
1
{WA1 e A}C 33 C {SA1 (I + e A) WA1 }
3
+ {WA3 (e A I )}C 33 C {SA2 (I + e A) SA1 WA2 }
+ {W(A1 A2 )e A}C 33 C {(W 2S)A3 (e A I )}

1
2
rA
+ WA
r e dr C 33 C {SA1 (I + e A) WA1 }
0
97
1 A
{WA e }C 33 C
(W 2S)A
1
rA
r e dr
0
+ {WA2 }
e rA C 33 C e rA dr {(W 2S)A2 } .
3 =
0
where
K 4 (r ) K 1 (r )dr +
1
0
N4 (r )33 N1 (r )dr,
K 4 (r ) K 1 (r )dr
= {WA1 e A}{SA1 } {WA2 (e A I )}{SA1 }

{WA1 e A}{SA2 (e A I )} + {WA1 }
1


0
0
1
N4 (r )33 N1 (r )dr
= {W(A1 A2 )e A}C 33 C {SA2 }

1
+ {WA1 e A}C 33 C {SA2 }
2
1
{W(A1 A2 )e A}C 33 C {SA1 }
2
1
+ {WA1 e A}C 33 C {SA1 }
3
{WA3 (e A I )}C 33 C {SA2 }
+ {W(A1 A2 )e A}C 33 C {SA3 (e A I )}
98

WA2

0
1 A
{WA e }C 33 C
2

SA
1
rA
r e dr
0
+ {WA }
e rA C 33 C e rA dr {SA2 } .

11 =
0
K 1 (r ) K 1 (r )dr +

0
N1 (r )33 N1 (r )dr,
where
1
1

0 K 1 (r ) K 1 (r )dr and 0 N1 (r )33 N1 (r )dr are evaluated as
in the formulae for 0 .
1
1

(r )dr + 0 N21 (r )33 N11
(r )dr
21 = 0 K 21 (r ) K 11
where
1
0

K 21 (r ) K 11
(r )dr

= SA1 e A Se A S3 S3 A1 {SA1 }
{(Se A S3 S3 S)A2 (e A I )}{SA1 }

+ SA1 e A Se A S3 S3 A1 {SA2 (e A I )}

1 rA rA
+ (Se A S3 S3 S)A1
e e dr {SA1 } ,
0
1
0

N21 (r )33 N11
(r )dr

= SA2 e A SA1 Se A S3 S3 A2 C 33 C {SA2 }
99

1 1 A
SA e Se A S3 S3 A1 C 33 C {SA2 }
2

1 2 A
SA e SA1 Se A S3 S3 A2 C 33 C {SA1 }
2

1 1 A
SA e Se A S3 S3 A1 C 33 C {SA1 }

3

(Se A S3 S3 S)A3 (e A I ) C 33 C {SA2 }

+ SA2 e A SA1 Se A S3 S3 A2 C 33 C {SA3 (e A I )}

1
(Se A S3 S3 S)A2

+ SA1 e A Se A S3 S3 A1 C 33 C SA2

Se A S3 S3 S A2
22 =
r e rA dr

K 22 (r ) K 22
(r )dr
1
1

+
N21 (r )33 N21
(r )dr +
N22 (r )33 N22
(r )dr,
0

K 21 (r ) K 21
(r )dr
where

0

K 21 (r ) K 21
(r )dr

= SA1 e A Se A S3 S3 A1 SA1 e A Se A S3 S3 A1

+ Se A S3 S3 S A2 (e A I ) SA1 e A Se A S3 S3 A1
100

+ SA1 e A Se A S3 S3 A1 Se A S3 S3 S A2 (e A I )

1 rA rA A

+ Se A S3 S3 S A1
e e dr Se S3 S3 S A1 ,
0
1
0

K 22 (r ) K 22
(r )dr =

0
K 1 (r ) K 1 (r )dr ,
the latter integral being evaluated as in the formula for 0 ,

1

N21 (r )33 N21
(r )dr
0

= SA2 e A SA1 Se A S3 S3 A2

C 33 C SA2 e A SA1 Se A S3 S3 A2

1 1 A
+
SA e Se A S3 S3 A1
2


1 2 A
+
SA e SA1 Se A S3 S3 A2
2

C 33 C SA1 e A Se A S3 S3 A1

1 1 A
SA e Se A S3 S3 A1
+
3


+ Se A S3 S3 S A3 (e A I )


+ SA2 e A SA1 Se A S3 S3 A2

C 33 C Se A S3 S3 S A3 (e A I )
101

Se A S3 S3 S A2
r e rA dr
0


+ SA1 e A Se A S3 S3 A1

1

r e rA dr
C 33 C Se A S3 S3 S A2
0

1 rA
+ Se A S3 S3 S A2
e
0

C 33 C e rA dr Se A S3 S3 S A2 ,

1
0

N22 (r )33 N22
(r )dr =
N1 (r )33 N1 (r )dr ,
the latter integral being evaluated as in the formula for 0 .

31 =
0

K 31 (r ) K 11
(r )dr +
1
0

N31 (r )33 N11
(r )dr ,
where
1
0

K 31 (r ) K 11
(r )dr
A 1 A

= Se S3 S3 A e e A S3 S3 A1 {SA1 }

Se A S3 S3 e A S3 S3 I A2 (e A I ) {SA1 }

+ Se A S3 S3 A1 e A e A S3 S3 A1 {SA2 (e A I )}

1 rA rA
+ Se A S3 S3 e A S3 S3 I A1
e e dr {SA1 } ,
0
102

0

N31 (r )33 N11
(r )dr

= SA1 (e A I ) + Se A S3 S3 A2 e A A1 e A S3 S3 A2

1 A 1 A
Se S3 S3 A e e A S3 S3 A1
2
1 1 A
C 33 C {SA2 }
SA (e I )
2

+ Se A S3 S3 A2 e A A1 e A S3 S3 A2 C 33 C {SA1 }
C 33 C {SA2 }

1 A 1 A
Se S3 S3 A e e A S3 S3 A1 C 33 C {SA1 }
3

Se A S3 S3 e A S3 S3 I A3 (e A I ) C 33 C {SA2 }

+ SA1 (e A I ) + Se A S3 S3 A2 e A A1 e A S3 S3 A2

C 33 C {SA3 (e A I )} Se A S3 S3 (e A S3 S3 I )A2

r e rA dr C 33 C {SA1} + Se A S3 S3 A1 e A e A S3 S3 A1

1

2
rA
C 33 C SA
r e dr + Se A S3 S3 e A S3 S3 I A2
32 =
0

K 31 (r ) K 21
(r )dr +
+
0

N31 (r )33 N21
(r )dr +
103

K 32 (r ) K 22
(r )dr

0

N32 (r )33 N22
(r )dr,
where
1
0

K 31 (r ) K 21
(r )dr

= Se A S3 S3 A1 e A e A S3 S3 A1 SA1 e A Se A S3 S3 A1

+ Se AS3 S3 e AS3 S3 I A2(e AI ) SA1e ASe A S3 S3 A1

+ Se A S3 S3 A1e Ae A S3 S3 A1 Se A S3 S3S A2 (e AI )

1

+ Se A S3 S3 e AS3 S3 I A1 e rA e rA dr Se AS3 S3S A1 ,

0

K 32 (r )33 K 22
(r )dr =

K 21 (r ) K 11
(r )dr ,

1

N31 (r )33 N21
(r )dr
0

= SA1 (e A I ) + Se A S3 S3 (A2 e A A1 e A S3 S3 A2 )


1 A 1 A
+
2


1 1 A
+
SA (e I )+ Se A S3 S3 A2 e A A1 e A S3 S3 A2
2


1 A 1 A
+
3

A A
3 A

+ Se S3 S3 e S3 S3 I A (e I )
104



C 33 C Se A S3 S3 S A3 (e A I )

1
A
2
A
rA
+ Se S3 S3 e S3 S3 I A
r e dr
0


+ Se A S3 S3 A1 e A e A S3 S3 A1

A
2 1 rA

Se S3 S3 S A
r e dr
C 33 C
0

1 rA
e
0

C 33 C e rA dr Se A S3 S3 S A2 ,
1
1

N32 (r )33 N22 (r )dr =
N21 (r )33 N11
(r )dr ,


1
1

33 =
K 31 (r ) K 31 (r )dr +
K 32 (r ) K 32
(r )dr ,
0
0
1
1

+
K 33 (r ) K 33
(r )dr +
N31 (r )33 N31
(r )dr
0
0
1
1

+
N32 (r )33 N32
(r )dr +
N33 (r )33 N33
(r )dr,
0
where
1
0

K 31 (r ) K 31
(r )dr
A 1 A

= Se S3 S3 A e e A S3 S3 A1
105

Se A S3 S3 A1 e A e A S3 S3 A1

+ Se A S3 S3 e A S3 S3 I A2 (e A I )

Se A S3 S3 A1 e A e A S3 S3 A1


Se A S3 S3 e A S3 S3 I A2 (e A I )

1 rA

e
0

e rA dr Se A S3 S3 e A S3 S3 I A1 ,
1
0

K 32 (r ) K 32
(r )dr =
1
0

K 21 (r ) K 21
(r )dr ,

1
1

K 33 (r ) K 33
(r )dr =
K 1 (r ) K 1 (r )dr ,
0

1

N31 (r )33 N31
(r )dr
0

= SA1 (e A I ) + Se A S3 S3 A2 e A A1 e A S3 S3 A2

C 33 C SA1 (e A I ) + Se A S3 S3

A2 e A A1 e A S3 S3 A2

1 A 1 A
+
2

C 33 C SA1 (e A I ) + Se A S3 S3

A2 e A A1 e A S3 S3 A2
106

1 1 A
SA (e I )+ Se A S3 S3 A2 e A A1 e A S3 S3 A2
2

C 33 C Se A S3 S3 A1 e A e A S3 S3 A1

1 A 1 A
+
3


+ Se A S3 S3 e A S3 S3 I A3 (e A I )

C 33 C SA1 (e A I )+ Se AS3 S3 A2 e AA1e A S3 S3 A2


C 33 C Se A S3 S3 e A S3 S3 I A3 (e A I )

1

r e rA dr
+



1

r e rA dr
C 33 C Se A S3 S3 e A S3 S3 I A2
0
A A
2
+ Se S3 S3 e S3 S3 I A
1

e rA C 33 C e rA dr Se A S3 S3 e A S3 S3 I A2 ,

0
1
0

N32 (r )33 N32
(r )dr =

N21 (r )33 N21
(r )dr ,

1
1

N33 (r )33 N33
(r )dr =
N1 (r )33 N1 (r )dr ,
0
107

41 = 3 .

42 =

K 4 (r ) K 21
(r )dr +

K 3 (r ) K 22
(r )dr
0
0
1
1

+
N4 (r )33 N21
(r )dr +
N3 (r )33 N22
(r )dr,
0
where
1
0

K 4 (r ) K 21
(r )dr

= {WA1 e A} SA1 e A Se A S3 S3 A1

+ {WA2 (e A I )} SA1 e A Se A S3 S3 A1

{WA1 e A} Se A S3 S3 S A2 (e A I )
1

1
+ {WA }
e rA e rA dr Se A S3 S3 S A1 ,
0

0

K 3 (r ) K 22
(r )dr =
1
0
K 3 (r ) K 1 (r )dr ,

1

N4 (r )33 N21
(r )dr
0

= {W(A1 A2 )e A}C 33 C SA2 e A SA1 Se A S3 S3 A2

1
{WA1 e A}C 33 C SA2 e A SA1 Se A S3 S3 A2
2

1
+ {W(A1 A2 )e A}C 33 C SA1 e A Se A S3 S3 A1
2
108

1
{WA1 e A}C 33 C SA1 e A Se A S3 S3 A1
3

+ {WA3 (e A I )}C 33 C SA2 e A SA1 Se A S3 S3 A2

+ {W(A1 A2 )e A}C 33 C Se A S3 S3 S A3 (e A I )

1

+ WA2
r e rA dr C 33 C SA1 e A Se A S3 S3 A1
0
1 A
{WA e }C 33 C
+ {WA2 }
Se
S3 S3
e rA C 33 C e rA dr

N3 (r )33 N22
(r )dr
1
rA
r e dr
0

S A

Se A S3 S3 S A2 ,
N3 (r )33 N1 (r )dr ,

1
1

43 =
K 4 (r ) K 31 (r )dr +
K 3 (r ) K 32
(r )dr
0
0
1
1

+
K 2 (r ) K 33
(r )dr +
N4 (r )33 N31
(r )dr
0
0
1
1

+
N3 (r )33 N32
(r )dr +
N2 (r )33 N33
(r )dr,
0
where

K 4 (r ) K 31
(r )dr
0

= {WA1 e A} Se A S3 S3 A1 e A e A S3 S3 A1

+ {WA2 (e A I )} Se A S3 S3 A1 e A e A S3 S3 A1
109

{WA1 e A} Se A S3 S3 e A S3 S3 A1 I A2 (e A I )
1

1
+ {WA }
e rA e rA dr Se A S3 S3 e A S3 S3 A1 I A1 ,

0

K 3 (r ) K 32
(r )dr

= {WA1 (I + e A) SA1 e A} SA1 e A Se A S3 S3 A1

+ {(S 2W)A2 (e A I )} SA1 e A Se A S3 S3 A1

+ {WA1 (I + e A)SA1 e A} Se A S3 S3 S A2 (e A I )
1

+ {(S 2W)A1 }
e rA e rA dr Se A S3 S3 S A1 ,

0

K 2 (r ) K 33
(r )dr =
1
0
K 2 (r ) K 1 (r )dr ,

1

N4 (r )33 N31
(r )dr
0

= {W(A1 A2 )e A}C 33 C SA1 (e A I ) + Se A S3 S3

1
A2 e A A1 e A S3 S3 A2 {WA1 e A}
2

C 33 C SA1 (e AI)+ Se A S3 S3 A2 e AA1 e A S3 S3 A2

1
+ {W(A1 A2)e A}C 33 C Se AS3 S3 A1e A e A S3 S3 A1
2

1
{WA1 e A}C 33 C Se A S3 S3 A1 e A e A S3 S3 A1
3

+ WA3 (e A I )}C 33 C {SA1 (e A I )

+ Se A S3 S3 A2 e A A1 e A S3 S3 A2
110

+ {W(A1 A2)e A}C 33 C Se A S3 S3 e AS3 S3 I A3 (e A I )

1

2
rA
+ WA
r e dr C 33 C Se A S3 S3 A1 e Ae A S3 S3 A1
0

1

{WA1 e A}C 33 C Se A S3 S3 e A S3 S3 I A2 r e rA dr
+ {WA2 }

0

0

e rA C 33 C e rA dr Se A S3 S3 e A S3 S3 I A2

N3 (r )33 N32
(r )dr
= {S(A1 A2 )e A + WA2 (I + e A) WA1 }

1
+ {WA1 (I + e A) SA1 e A}
2

1
+ {S(A1 A2 )e A + WA2 (I + e A) WA1 }
2

1
+ {WA1 (I + e A) SA1 e A}
3

C 33 C SA1 e A Se A S3 S3 A1 + {(S2W)A3 (I +e A)}

+ {S(A1 A2 )e A + WA2 (I + e A) WA1 }

C 33 C Se A S3 S3 A1 S A3 (e A I )

1

2
rA
+ (S 2W)A
r e dr C 33 C SA1 e A Se A S3 S3 A1
0
111
1 A
+ {WA (I + e ) SA e }C 33 C
rA
r e dr

Se A S3 S3 S A2
+ {(S 2W)A2 }
r e rA C 33 C e rA dr

Se A S3 S3 S A2 ,

N2 (r )33 N33
(r )dr =
N2 (r )33 N1 (r )dr ,

52 = 3 .

K 3 (r ) K 33
(r )dr
1
1

+
N4 (r )33 N32
(r )dr +
N3 (r )33 N33
(r )dr,
53 =

K 4 (r ) K 32
(r )dr
where

K 4 (r ) K 32
(r )dr =
1
0

K 4 (r ) K 21
(r )dr ,

1
1

K 3 (r ) K 33
(r )dr =
K 3 (r ) K 1 (r )dr ,
0

1
1

N4 (r )33 N32
(r )dr =
N4 (r )33 N21
(r )dr ,
0
112

1
1

N3 (r )33 N33 (r )dr =
N3 (r )33 N1 (r )dr ,
0

63 = 3 .
The integrals
1
0
e rA e rA dr and
1
0
e rA C 33 C e rA dr in the
above formulae can be evaluated by the method of Van Loan

1
[1978] and the integral 0 r e rA dr from the formula
1
r e rA dr = A1 e A A2 (e A I ),
0
which is, easily, obtained using the method of integration of

parts.
113
CHAPTER THREE
Model Specification
3.1 Introduction
In this chapter we discuss the new mixed-order continuous time macroeconometric model of the UK economy with
stochastic trends. The model is based on the Bergstrom,
Nowman and Wymer [1992] second-order continuous time
macroeconometric model. The model is the first continuous time macroeconometric model incorporating stochastic
trends, and its development represents a major step forward
in general continuous time macroeconometric modelling
over the last 30 years. The detailed specification of the structural equations and economic theory of previous continuous
time macroeconomic models of various economies of the
world is presented, for example, in Bergstrom and Wymer
[1976], Knight and Wymer [1978], Bergstrom, Nowman
114
Model Specification
and Wymer [1992], Jonson, Moses and Wymer [1977],

Knight and Mathieson [1979], Gandolfo and Padoan, [1982,
1984, 1987a, 1990], Kirkpatrick [1987], Bailey, Hall and
Phillips [1987], Donaghy [1993], Sjo o [1993], Nieuwenhuis
[1995] and Stavrev [2001], amongst others. In this chapter
we provide a rigorous analysis and discussion of each of the
macroeconomic structural relationships in the model and a
formal derivation of them as the solution of a dynamic optimization problem, which takes account of adjustment costs.
We also provide a discussion of the underlying assumptions
concerning firm behaviour and market structure.
This chapter is organised as follows. Section 3.2 present
the complete set of equations of the model where we
shall, also, discuss their general properties. Sections 3.3
and 3.4 present the private consumption and residential
fixed capital equations as they have similar specifications.
Sections 3.5 to 3.8 discuss the equations relating to the production sector of the economy, in particular, employment,
private non-residential fixed capital, output and the price
level. Section 3.9 presents the wage rate adjustment equation and Section 3.10 the interest rate equation. The last
group of equations for international flows are presented in
Sections 3.11 to 3.16 to explain imports, non-oil exports,
transfers abroad, real profits, interest and dividends from
115
abroad, cumulative net real investment abroad and the exchange rate. The stock identity is presented in Section 3.17
and conclusions in Section 3.18. This chapter, also, contains
two technical Appendices presenting the formal derivation
of the adjustment equations and the distributed lag relations.
3.2 Equations and General Properties

of the Model
The model, in its deterministic form, is set out below. It is a
system of 18 mixed first- and second-order nonlinear differential equations with 63 structural parameters, including a
vector of 33 long-run parameters, a vector of 27 speed
of adjustment parameters, and a vector of 3 trend or drift
parameters. The variables and equations of the model, in its
deterministic form, are as follows.
Endogenous Variables
C = real private consumption
E n = real non-oil exports
F = real current transfers abroad
I = volume of imports
K h = residential fixed capital
K = private non-residential fixed capital
116
Model Specification
K a = cumulative net real investment abroad (excluding

change in official reserves)
L = employment
P = real profits, interest and dividends from abroad
p = price level
Q = real net output
q = exchange rate (price of sterling in foreign currency)
r = interest rate
S = stocks
w = wage rate
Exogenous Variables
B = stock of bonds
dx = dummy variable for exchange controls (dx = 1 for
197479, dx = 0 for 1980 onwards)
E o = real oil exports
Gc = real government consumption
K p = public non-residential fixed capital
pf = price level in leading foreign industrial countries
pi = price of imports (in foreign currency)
M = money supply
rf = foreign interest rate
T1 = total taxation policy variable ((Q + P )/T1 is real private disposable income)
117
T2 = indirect taxation policy variable (Q /T2 is real output

at factor cost)
Y f = real income of leading foreign industrial countries
Unobservable Trend Variables

1 = productivity trend variable (d1 /dt is the proportional
rate of decrease in the amount of labour required to
produce a given output with a given amount of capital)
2 = labour supply trend variable (d2 /dt is the proportional
rate of growth in the non-accelerating inflation level
of employment)
3 = trend variable allowing for the growth in the use of
credit and charge cards (plastic money) (d3 /dt is the
proportional rate of growth in the use of cards)
Structural Equations

1 e {2 (r Dlog p)+3 Dlog p} (Q + P )
DlogC = 1 + 2 + 1 log
T1 C
(3.1)
D 2 logL = 2 (2 DlogL)

4 e 1 {Q 6 5 K 6 }16
+ 3 log
L
118
(3.2)
Model Specification
D 2 logK h = 4 (1 + 2 DlogK h )

7 e {8 (r Dlog p)+9 Dlog p} (Q + P )
+ 5 log
T1 K h
(3.3)
D 2 logK = 6 (1 + 2 DlogK )

5 (Q /K )(1+6 )
+ 7 log
r 10 Dlog p + 11
(3.4)
DlogQ
= 1 + 2 + 8 log

{112 (qp/pi )13 }{1+14 (1 +2 )}
(C +Gc +DK+DKh +DK p + En + E o )
Q

+ 9 log

14 (C + Gc + D K + DKh + DK p + E n + E o )
S
(3.5)
D 2 log p = 10 (Dlog(w/p) 1 )

15 4 T2 we 1 {1 5 (Q /K )6 }(1+6 )/6
+ 11 log
p
(3.6)
D 2 logw = 12 (1 Dlog(w/p)) + 13 Dlog( pi /qp)

4 e 1 {Q 6 5 K 6 }1/6
+ 14 log
16 e 2
119
(3.7)
D 2 r = 15 Dr + 16 [17 + 18rf 19 Dlogq

+ 20 { p(Q + P )/Me 3 } + 21 (B/M) r ]
(3.8)
DlogI
= 1 + 2

12 (qp/pi )13 {1 + 14 (1 + 2 )}
(C + Gc + DK + DKh + DK p + E n + E o )
+ 17 log
( pi /qp)I

14 (C + Gc + DK + DKh + DK p + E n + E o )
+ 18 log
S

(3.9)

DlogE n = 1 + 2 + 19 log
22 Y f 23 ( pf /qp)24
En

(3.10)
DF = 20 {25 (Q + P ) F }
(3.11)
DP = 21 [{26 + 27 (rf Dlog pf )}K a P ]
(3.12)
D 2K a = 22 DKa + 23 [{28 + 29 (rf r )30 Dlogq 31 dx }

(Q + P ) K a ]

32 pf
+ 25 log
qp

E n + E o + P F DKa
+ 26 (r 33 ) + 27 log
( pi /qp)I
(3.14)
pf
D logq = 24 Dlog
qp
2
(3.13)
120
Model Specification
DS = Q + ( pi /qp)I C DK DKh
DK p E n E o Gc
(3.15)
D1 = 1
(3.16)
D2 = 2
(3.17)
D3 = 3
(3.18)
The above model can be simplified by solving (3.16)(3.18)

subject to the normalization conditions 1 (0) = 0, 2 (0) =
0, 3 (0) = 0, to obtain 1 = 1 t, 2 = 2 t, 3 = 3 t, and
substituting these expressions into (3.2), (3.6), (3.7) and
(3.8). But, this simplification will not be possible in the
stochastic form of the model, where (3.16), (3.17) and (3.18)
will be replaced by stochastic differential equations and
white noise innovations will be added to each of the other
equations. The main innovative feature of the above model
is the introduction of stochastic trends to represent unobservable variables that, in earlier models, have been represented by deterministic trends. But, the model also retains
several important features of the earlier models of Bergstrom
and Wymer [1976] and Bergstrom, Nowman and Wymer
[1992].
121
One of these features is that nearly every equation has the

form of a partial adjustment (error correction) equation, in
which the dependent variable is adjusting, continuously, in
response to the deviation of its current level from its partial
equilibrium level, which is a function of other variables in
the model. These partial adjustment equations are a mixture
of first- and second-order differential equations. The firstorder equations are of the form
d x(t)
= + { f (z(t)) x(t)},
dt
(3.19)
where z(t) is a vector of other variables in the model that,

directly, influence x(t), f (z(t)) (a function of z(t)) is the
partial equilibrium level of x(t) at time t, is a speed of
adjustment parameter and is a parameter representing
the expected rate of increase in f (z(t)), which depends on
the expected rates of increase in the elements of z(t). The
second-order equations are of the form

d 2 x(t)
d x(t)
= 1
+ 2 { f (z(t)) x(t)}
dt 2
dt
(3.20)
in which 1 and 2 are speed of adjustment parameters and

and f (z(t)) have the same interpretation as in (3.19).
The main reason why economic variables adjust gradually, rather than instantaneously, to their partial equilibrium
level is that there are adjustment costs, which depend on the
122
Model Specification
rate of change and, possibly, the acceleration of the adjusting

variable. Moreover, equations of the form (3.19) and (3.20)
can be, formally derived as the solution of a dynamic optimization problem, which takes account of these adjustment
costs. This is done in Appendix A of this chapter. Equation
(3.19) (which is identical with (3.A15) of Appendix A) is
obtained by minimizing the integral over future time of a
quadratic cost function, whose value, at any future point
of time, depends on both the expected deviation of the
adjusting variable from its partial equilibrium value and its
expected rate of change at that point of time. Equation (3.20)
(which identical with (3.A37) of Appendix A) is obtained in
a similar way by using a cost function that takes account of
second-order adjustment costs, the value of the function, at
any future point of time, depending on the expected acceleration of the adjusting variable at that point of time, as well
as its expected rate of change and expected deviation from
its partial equilibrium value.
Another general feature of the model, which was also a
feature of the model of Bergstrom, Nowman and Wymer
[1992], is that it incorporates the assumption of long-run
rational expectations. In particular, it is assumed that the
agents in the economy know the trend parameters 1 and
2 for productivity and the labour supply and take these
123
into account in forming their expectations. The assumption that the agents in the economy know the parameters
is, of course, more realistic than the assumption that they
know the parameters affecting the short-run dynamics of
the model, which would be necessary to justify the assumption of full rational expectations. It is shown in Chapter 4
that the steady state growth rate of most of the real variables
in the model is 1 + 2 . The long-run rational expectations
assumption is incorporated in equations of the form (3.19)
or (3.20) relating to these variables, therefore, by setting the
expectations parameter equal to 1 + 2 .
The reference, in the preceding paragraph, to the steady
state growth rate leads us, naturally, to the discussion of
a third general feature of the model. Like the model of
Bergstrom and Wymer [1976] and the many models for
which that model served as a prototype (including the model
of Bergstrom, Nowman and Wymer [1992]), it is designed
in such a way as to permit a rigorous mathematical analysis of its steady state and stability properties, thus providing
a check on the capacity of the model to generate plausible
long-run behaviour. Because of the smallness of the samples available for the estimation and testing of macroeconometric models, it is very important to ensure, preferably by
the mathematical analysis of their dynamic properties, that
124
Model Specification
they are capable of generating plausible long-run behaviour

under realistic assumptions about the behaviour of the
exogenous variables. As was pointed out by Bergstrom and
Wymer [1976, p. 268], failure to generate plausible longrun behaviour could indicate a structural defect, such as
the omission of an important feedback. Such a defect could
seriously affect the predictive power of the model and its
usefulness for either medium-term forecasting or policy
analysis.
The final feature of the model that we shall mention is
that, like the models of Bergstrom and Wymer [1976] and
Bergstrom, Nowman and Wymer [1992], it makes intensive
use of economic theory to obtain a parsimonious parameterization and cross equation restrictions. This will be evident
from the discussion of the individual equations of the model,
to which we now turn.
3.3 Private Consumption

Real private consumption C is one of six variables in the
model whose adjustment equations take the form of firstorder differential equations, the adjustment equations of the
remaining equations being of the second order. In choosing
the variables whose adjustment equations were to take the
125
form of first-order differential equations, we relied not only

on our a priori knowledge of the ease with which the speed
of adjustment of each variable can be changed, but also on
the estimates of the speed of adjustment parameters in the
model of Bergstrom, Nowman and Wymer [1992]. Since the
time lag distribution generated by a first-order differential
equation is of the exponential form, as is shown in Appendix
B of this chapter, it can be a good approximation to the
true unknown adjustment equation only if the mode of the
time lag distribution generated by that equation is either
small or nonexistent. The six variables whose adjustment
equations take the form of first-order differential equations
in this model are those for which the estimated modal time
lags in the model of Bergstrom, Nowman and Wymer [1992]
are smallest (C, Q, E, I, F, P). The estimated modal time lag
of C in that model is only 0.78 quarters.
The consumption adjustment equation (3.1) assumes that
the proportional rate of increase in consumption depends
on the ratio of the partial equilibrium level of consumption 1 e {2 (r Dlog p)+3 Dlog p} (Q + P )/T1 to the current level
of consumption. The taxation policy variable T1 is defined
in such a way that (Q + P )/T1 is real private disposable
income (as normally measured); that is T1 is the ratio of
the net real national income at market prices to real private
126
Model Specification
disposable income (as normally measured). The partial

equilibrium level of consumption is assumed, therefore, to
depend on real private disposable income (Q + P )/T1 , the
real interest rate r Dlog p and the inflation rate Dlog p.
The reason for assuming that inflation has an independent
effect on consumption (in addition to its effect through the
real interest rate) is that inflation reduces the real disposable income of holders of government bonds (like a hidden
tax), and, hence, the variable (Q + P )/T1 overstates the true
level of real disposable income. The partial equilibrium level
of consumption is really assumed, therefore, to be a function of only the real interest rate and real disposable income
correctly measured. We expect both 2 and 3 to be positive,
implying that the propensity to consume is lower the higher
are both the real interest rate and the inflation rate. The
parameter 1 can be interpreted as the limit of the propensity to consume as the real interest rate and inflation rate
each tend to zero. Consumption is assumed to adjust gradually, rather than instantaneously, to its partial equilibrium
level because of adjustment costs. These costs can be either
physical or psychological.
Equation (3.1) implies that when consumption is at its
partial equilibrium level, it will increase at the rate 1 + 2 ,
which is its steady state growth rate, as will be shown in
127
Chapter 4. It is implicitly assumed, therefore, that consumers

expect the variables affecting consumption to change in such
a way that the partial equilibrium level of consumption
increases at its steady state. In this way, the equation incorporates the assumption of long-run rational expectations.
It should be noted, also, that the equation has
the
[1 e
form
of
(3.19),
{2 (r Dlog p)+3 Dlog p}
with
x(t) = logC , f (z(t)) = log
(Q + P )/T1 ] and = 1 + 2 . It can
be formally derived, therefore, as the solution of a dynamic

optimization problem, which takes account of first-order
adjustment costs, as in Appendix A of this chapter.
3.4 Residential Fixed Capital

It is convenient, at this stage, to discuss (3.3), which is
the residential (housing) fixed capital adjustment equation.
Although it is a capital adjustment equation, it is, essentially, a demand equation for residential services, since the
consumption of residential services is, implicitly, assumed
to be proportional to the stock of residential capital. The
equation is similar in form, therefore, to (3.1), except that
it is a second-order differential equation. It assumes that
the acceleration of the logarithm of the stock of residential
128
Model Specification
capital depends on the excess of its steady state growth rate

1 + 2 over its current growth rate and on the ratio of the
partial equilibrium stock of residential capital to the current
stock. The formulation as a second-order differential equation allows for the fact that adjustment costs will have a
more important influence on the adjustment of residential
capital and the consumption of the services of this capital
than on the general level of consumption.
The equation implies that, when the stock of residential capital is at its partial equilibrium level and growing
at its steady state rate, its acceleration is zero. It is implicitly assumed, therefore, that the partial equilibrium stock
of residential capital is expected to grow at its steady state
rate, thus incorporating the assumption of long-run rational
expectations.
It should be noted, also, that the equation has
the
[7 e
form
of
(3.20)
{8 (r Dlog p)+9 Dlog p}
with
x(t) = logK h , f (z(t)) = log
(Q + P )/T1 ] and = 1 + 2 . It can
formally derived, therefore, as the solution of a dynamic

optimization problem, which takes account of both firstand second-order adjustment costs, as in Appendix A of this
chapter. Like (3.1) and (3.3), most of the other equations in
the model can be expressed in the form of either (3.19) or
129
(3.20) with being chosen so that the equation is consistent

with the assumption of long-run rational expectations.
Since this will be fairly obvious, it will be unnecessary to
comment on this aspect of the equation in each case.
3.5 Employment
In the next few sections, we shall discuss the equations
relating to the production sector of the economy. These
are the adjustment equations for employment, private nonresidential fixed capital, output and the price level. It will be
useful, before discussing these equations in detail, to say
something about the underlying assumptions concerning
firm behaviour and market structure. It is assumed that the
economy is made up of a large number of monopolistic competitors with identical production functions and uniformly
differentiated products. Since the prices charged by the different firms in the economy are equal, it can be treated as
a single product economy, and it tends to a perfectly competitive economy as the degree of differentiation between
the products of the different firms tends to zero, that is as
the price elasticity of the partial demand function for each
firms product tends to minus infinity.
130
Model Specification
It would be possible to derive a set of adjustment equations for employment, output, capital and the price level
by a multidimensional generalization of the analysis in
Appendix A of this chapter, assuming that the agent has
control of a vector of variables, rather than a single variable.
But, to limit the number of parameters to be estimated, we
followed a simpler approach, which takes account of our a
priori knowledge of the relative speeds at which different
factors of production can be adjusted (or the relative costs
of adjusting them at a given speed). We know that it is easier to adjust output by varying the intensity with which the
employed labour force is used (varying the number of hours
per week worked by each employee) than to vary the number of persons employed, and it is easier to vary the number
of persons employed than to vary the stock of fixed capital. We assume, therefore, that, at each point of time, output is adjusting in response to sales, the number of persons
employed is adjusting in response to output and the stock
of capital is adjusting in response to the marginal product of
capital and the real interest rate.
The remainder of this section is devoted to a more detailed
discussion of the employment adjustment equation (3.2).
This is formulated as a second-order differential equation,
131
partly for a priori reasons mentioned in the preceding paragraph, and partly on the basis of the estimated speed of
adjustment parameters in the employment equation in the
model of Bergstrom, Nowman and Wymer [1992]. These
estimates imply that the time lag distribution with which
employment adjusts to output and capital has a mode of
3.48 quarters (the modal time lag). The exponential time
lag distribution generated by a first-order differential equation would, therefore, be a poor approximation to that
distribution.
Equation (3.2) assumes that the acceleration of the logarithm of employment depends on the excess of its steady
state growth rate 2 over its current growth rate and on
the ratio of the partial equilibrium level of employment to
the current level. The function 4 e 1 {Q 6 5 K 6 }1/6
defining the partial equilibrium level of employment is
a constant elasticity of substitution production function,
which incorporates the productivity stochastic trend variable 1 as well as output and capital. This function
plays a central role in the model. Its parameters occur
in several other equations, resulting in a very parsimonious parametrization and cross equation restrictions. The
elasticity of substitution between labour and capital is
1/(1 + 6 ) and can be derived as follows.
132
Model Specification
When the production sector of the economy is in a state

of partial equilibrium, employment is related to output and
capital by the equation
L = 4 e 1 {Q 6 5 K 6 }1/6 ,
(3.21)
from which we obtain

1+6
L
4 e 1 6
{5 6 K (1+6 ) }
{Q
5 K 6 } 6
=
K
6

(1+6 ) (1+6 )
= 5 (4 e 1)6 4 e 1{Q 6 5 K 6}1/6
K
(1+6 )
L
= 5 (4 e 1 )6
,
K
and hence
L
=
K
(4 e 1 )6
5
1
1+6
"
(3.22)
The term L/K (which is, of course, positive) is the

marginal technical rate of substitution between labour
and capital and is, approximately, equal to the number of units of labour released as a result of the addition of one unit of capital, when output is held constant.
Equation (3.22) shows that the ratio of labour to capital
employed is related to the marginal rate of substitution
between labour and capital through a function with constant
elasticity 1/(1 + 6 ). This elasticity is called the elasticity
133
of substitution between labour and capital. It is shown in

Chapter 4 that the steady state growth rate of employment
is 2 , which is the rate of growth of the non-accelerating
inflation level of employment. Equation (3.2) implies, therefore, that, when employment is at its partial equilibrium
level and growing at its steady state rate, its acceleration is
zero.
3.6 Private Non-Residential Fixed Capital

Equation (3.4) assumes that the acceleration of the logarithm of the stock of private non-residential fixed capital
depends on the excess of its steady state growth rate 1 + 2
over its current growth rate and on the ratio of the marginal
product of capital 5 (Q /K )(1+6 ) to the real rate of interest plus a risk premium. The marginal product of capital is
obtained from the relation
L/K
Q
=
,
K
L/Q
where L/K and L/Q are the partial derivatives of the
production function 4 e 1 {Q 6 5 K 6 }1/6 . We have
134
Model Specification

1+6

4 e 1 6
L
6 6
=
5 6 K (1+6 ) ,
{Q
5 K }
K
6

1+6

4 e 1 6
L
=
6 Q (1+6 ) ,
{Q
5 K 6 } 6
Q
6
and hence
{5 6 K (1+6 ) }
Q
=
= 5
K
{6 Q (1+6 ) }
Q
K
(1+6 )
.
We assume that investors are risk averse and that the

minimum value of the marginal product of capital required
to induce further investment in fixed capital must exceed
the real interest rate r Dlog p by a risk premium. Moreover, we assume that the risk premium depends on the
rate of inflation, since the degree of uncertainty concerning the average rate of inflation over the life of fixed capital, and concerning, therefore, the relative rates of return
from investment in fixed capital and bonds, is likely to
depend on the current rate of inflation. The sum of the
real interest rate and the risk premium is represented in
(3.4) by the term r 10 Dlog p + 11 . In the special case
where 10 = 1, the risk premium is a constant 11 . Equation (3.4) implies that, when the marginal product of capital
135
equals the real interest rate plus the risk premium and capital is growing at its steady state rate, its acceleration is
zero.
3.7 Output
Equation (3.5) assumes that the proportional rate of increase in output depends on the ratio of the partial
equilibrium level of output to the current level of output
and the ratio of the partial equilibrium level of stocks
(inventories) to the current level of stocks. The partial
equilibrium level of stocks is assumed to equal 14 (C +
Gc + DK + DKh + DK p + E n + E 0 ), where the term in brackets equals the total sales for private and public consumption, capital formation and exports, while the parameter 14
is the optimal ratio of stocks to sales. It is assumed that, in
the state of partial equilibrium, the total supply from output
and imports {Q + ( pi /pq)I } is just sufficient to meet current
sales and maintain the constant ratio 14 of stocks to sales,
when sales are growing at their steady state rate 1 + 2 .
The supply required to maintain the constant ratio 14 of
stocks to sales, when sales are growing at the rate 1 + 2 , is
14 (1 + 2 )(C + Gc + DK + DKh + DK p + E n + E 0 ), and the
total supply required to meet current sales and maintain
136
Model Specification
the ratio 14 of stocks to sales is, therefore, {1 + 14 (1 + 2 )}

(C + Gc + DK + DKh + DK p + E n + E 0 ).
We assume that the proportion of the total supply that
is provided by imports depends on the ratio of the price
of imports to the price of domestic goods and, in particular that it equals 12 (q p/pi )13 . In the state of partial equilibrium imports and output must, therefore, satisfy (3.23)
and (3.24).
( pi /pq)I = 12 (qp/pi )13 {1 + 14 (1 + 2 )}
(C + Gc + DK + DKh + DK p + E n + E 0 ),
(3.23)
Q = {1 12 (qp/pi )13 }{1 + 14 (1 + 2 )}
(C + Gc + DK + DKh + DK p + E n + E 0 )
(3.24)
From (3.23) we obtain

I = 12 (qp/pi )(1+13 ) {1 + 14 (1 + 2 )}
(C + Gc + DK + DKh + DK p + E n + E 0 ),
(3.25)
which implies that the price elasticity of demand for imports

is (1 + 13 ).
The partial equilibrium level of output is given by the
term on the right-hand side of (3.24). Equation (3.5)
implies that, when both output and stocks are at their
137
partial equilibrium levels, output will grow at its steady state

rate 1 + 2 .
3.8 Price Level

Equation (3.6) assumes that the acceleration of the logarithm of the price level depends on the excess of the current rate of increase in the real wage over the rate of
technical progress 1 and on the ratio of the partial equilibrium price level to the current price level. The partial
equilibrium price level 15 4 T2 e 1 {1 5 (Q /K )6 }(1+6 )/6
is equal to 15 T2 w(L/Q ), where L/Q is the partial
derivative, with respect to Q, of the production function
4 e 1 {Q 6 5 K 6 }1/6 . It follows from the definition
of the indirect tax variable T2 that it can be written as
T2 = 1 + t2 , where t2 is the average rate of indirect taxation. The term T2 w(L/Q ) is equal, therefore, to marginal
cost including indirect taxation. The parameter 15 is a measure of the degree of imperfection of competition, and it
is assumed to satisfy the condition 15 1, with 15 = 1 in
the limiting case of perfect competition. In that limiting case,
therefore, the partial equilibrium price level equals marginal
cost including indirect taxation.
138
Model Specification
In the more general case, 15 is defined by 15 = e/(1 + e),

where e is the own-price elasticity of the partial demand
function for each firms product. The partial equilibrium
condition p = 15 T2 w(L/Q ) is then equivalent to the
profit maximization condition that the marginal revenue
of the representative firm is equal to its marginal cost. To
show this let pi , Q i and L i denote the price, output and
amount of labour employed by firm i. The marginal revenue
( pi Q i )/ Q i of firm i is then given by

( pi Q i )
pi
= Qi
+ pi
Q i
Q i

Q i pi
= pi
+1
pi Q i

(1 + e)
= pi
e
pi
=
.
15
The condition that the marginal revenue of firm i equals
its marginal cost is given, therefore, by

L i
pi
= T2 w
15
Q i
from which we obtain
L i
pi = 15 T2 w
Q i
139

.
Hence, since the price, output and amount of labour

employed are assumed to be the same for all firms, we
obtain

L
p = 15 T2 w
Q

.
The first term on the right-hand side of (3.6) differs

from the corresponding term in most of the second-order
equations in the model. For most of the second-order
equations, this term is the deviation of the rate of change
of the variable, whose acceleration is being explained by
the equation, from the steady state rate of change of that
variable. The steady state rate of change in the price level
does not exist, however, unless the money supply grows at
a constant rate, and past variations in the rate of growth
of the money supply have been so great that it is unrealistic to assume that agents in the economy have any firm
expectations concerning the future rate of change in the
money supply. The first term on the right-hand side of (3.6)
can be written as 10 {(Dlogw 1 ) Dlog p}, in which the
term Dlogw 1 can be regarded as the expected rate of
change of the partial equilibrium price level in the medium
future. Then (3.6) implies that, when the price variable is
at its partial equilibrium level and growing at the expected
140
Model Specification
rate of increase of its partial equilibrium level, its acceleration is zero. Finally, it should be mentioned that the speed
of adjustment parameters in (3.6) depends, not only on
adjustment costs (which, for price variables, are relatively
minor), but also on the time lags in the competitive process,
in which each firm is adjusting its price in small steps, taking
account of the current prices of its competitors products (see
Chamberlin [1946, Ch. 5, Section 3]).
3.9 Wage Rate

Equation (3.7) assumes that the acceleration of the logarithm of the wage rate depends on the excess of the rate of
technical progress 1 over the current rate of increase in the
real wage, on the current rate of increase in the ratio of the
import price level to the domestic price level (both measured
in sterling), and on the ratio of the partial equilibrium level
of employment to the labour supply 16 e 2 . The labour supply 16 e 2 can be interpreted as the non-accelerating inflation level of employment, since (3.7) implies that, when the
real wage is growing at its steady state rate 1 and the real
exchange rate is constant, the acceleration of the nominal
wage rate will be zero if, and only if, the partial equilibrium level of employment (which depends on the current
141
levels of output and capital) equals 16 e 2 . The labour supply, so defined, is not directly measurable, since it depends,
not only on demographic factors, but also on conditions in
the labour market. It is assumed, therefore, to be an exponential function of the unobservable trend variable 2 with
constant drift parameter 2 and, in the stochastic version of
the model, constant volatility parameter.
The first term on the right-hand side of (3.7) has the same
form as the first term on the right-hand side of (3.6) (discussed above) with the opposite sign. The two equations
together imply that, if the rate of increase in the real wage
rate exceeds the rate of technical progress and all other terms
on the right-hand side of each equation are zero, then there
will be an acceleration of the price level and a deceleration
of the wage rate, while, if the rate of increase in the real
wage rate is less than the rate of technical progress and all
other terms on the right-hand side of each equation are zero,
then there will be an acceleration of the wage rate and a
deceleration of the price level. The term Dlog( pi /q p) allows
for pressure (by trade unions and other organizations) for
higher wages to compensate for the loss of welfare caused
by the fall in the real exchange rate (since, although w/p
is the measure of the real wage that is relevant for firms
employing labour, it is not a measure of welfare).
142
Model Specification
Equation (3.7), together with the other equations in the

model, implies that there is no long-run trade off between
unemployment and inflation. This can be seen by comparing two steady states, in one of which the money supply M
is growing at a faster rate than in the other. The only variables in (3.7) that will change as between the two steady
states are Dlogw, Dlog p and Dlogq. As we move to the
steady state with the higher rate of growth of M, Dlogw
and Dlog p will increase by the same amount as DlogM,
while Dlogq will decrease by the same amount. The first
two terms on the right-hand side of (3.7) will be unchanged,
therefore, and, since the left-hand side will be zero, the
third term on the right-hand side must also be unchanged.
In both steady states, all 3 terms on the right-hand side
of (3.7) are zero, and employment equals 16 e 2 (see
Chapter 4).
A final point to be noted concerning (3.7) is that we
have used the concept of the non-accelerating level of
employment, rather than the more commonly used concept
of the non-accelerating inflation level of unemployment.
The advantage of this is that employment is a more precisely defined variable than unemployment, since the latter
depends on the somewhat arbitrary definition of the labour
force.
143
3.10 Interest Rate

The interest rate adjustment equation (3.8) represents the
dynamic behaviour of the market for long-dated bonds. It
is, essentially, a portfolio balance equation, which allows for
substitution between money, domestic bonds and foreign
bonds. The acceleration of the interest rate (the yield on
long-dated bonds) is assumed to depend on the current rate
of change of the interest rate and the excess of the partial
equilibrium interest rate over the current rate. The partial
equilibrium interest rate is the rate at which the demand
for domestic bonds equals the current stock of these bonds,
given the return on the two alternative assets money and
foreign bonds.
The return from holding money comes mainly from the
facilitation of transactions. The marginal rate of return from
holding money is assumed, therefore, to depend on the
ratio of income to the stock of money and is represented
by the term 20 { p(Q + P )/Me 3 }, in which the stochastic
trend variable 3 allows for the growing use of plastic money
(credit and charge cards).
The rate of return on foreign bonds is represented by
the term 17 + 18rf 19 Dlogq, in which Dlogq is a proxy
for the expected rate of change in the exchange rate.
144
Model Specification
If there were no uncertainty, we would expect to have

18 = 19 = 1. Indeed, the partial equilibrium condition for
(3.8) includes, as a very special case, the uncovered interest rate parity condition r = rf Dlogq. But this cannot be
expected to hold in a world of uncertainty. The inclusion
of the term 21 (B/M) in (3.8) ensures, together with the
other terms, that, if the stock of bonds, the money stock and
income all increase in the same proportion and the foreign
interest rate and the rate of increase in the exchange rate
all remain unchanged, then the partial equilibrium level of
r will also remain unchanged.
3.11 Imports
Equation (3.9) assumes that the proportional rate of increase
in the volume of imports depends on the ratio of the partial
equilibrium real value of imports to the current real value
of imports and the ratio of the partial equilibrium level of
stocks to the current level of stocks. It implies that, when
stocks and the real value of imports are at their steady state
levels, the volume of imports will increase at its steady state
rate 1 + 2 . The partial equilibrium real value of imports
has been discussed in Section 3.7, where it was shown that
the price elasticity of demand for imports is (1 + 13 ).
145
Equations (3.5) and (3.9) together ensure that, in the steady

state, imports and domestic output are just sufficient to meet
total sales and keep stocks growing at the rate 1 + 2 . It
should be noticed that we have distinguished between the
volume of imports I and the real value of imports pi I /q. To
avoid introducing two many price variables (and to be consistent with our assumption of, essentially, a single product economy), all other real variables in the model (for
example, C ) have been defined as values at current prices
deflated by p the implicit price deflator of the gross domestic
product.
3.12 Non-Oil Exports

Equation (3.10) assumes that the proportional rate of
increase of non-oil exports depends on the ratio of the partial
equilibrium level of non-oil exports to the current level. It
implies that, when non-oil exports are at their partial equilibrium level, they will increase at their steady state rate
1 + 2 . The partial equilibrium level of non-oil exports is
assumed to be a function of foreign real income Yf and
the ratio of the foreign price level to the domestic price level
(both measured in sterling) pf /q p. This is a standard constant
146
Model Specification
elasticity form of demand function, in which the parameters

23 and 24 can be interpreted as the foreign income and
price elasticities of demand for UK products.
3.13 Transfers Abroad

Equation (3.11)(3.13) are the adjustment equations for the
real transfers abroad F, real profits, interest and dividends
from abroad P and cumulative net real investment abroad
K a . These equations are formulated in natural values rather
than logarithms, because each of the variables F, P and K a
can be either positive or negative. Equation (3.11) assumes
that the rate of increase of real transfers abroad F is proportional to the excess of the partial equilibrium level of this
variable over its current level and that the partial equilibrium level is proportional to real income Q + P .
3.14 Real Profits Interest and Dividends

from Abroad
Equation (3.12) has the same general form as (3.11). It
assumes that the rate of increase of real profits, interest and
dividends from abroad P is proportional to the excess of the
147
partial equilibrium level of this variable over its current level.

The partial equilibrium level is assumed to equal cumulative
real net investment abroad K a multiplied by a linear function of the real foreign interest rate rf Dlog pf . This is an
approximation, which takes account of the several different
types of income in the variable P.
Consider, first, the real dividend component of P. This
equals the real market value of the shares on which these
dividends are paid multiplied by the dividend yield. The market value of the shares will, generally, be somewhat greater
than the share component of K a , because of the growth in
the market value of the shares since their purchase by the
UK owners. On the other hand, the dividend yield, although
closely related to the real interest rate rf Dlog pf , will,
normally, be less than the latter variable, because of the
potential growth of dividends. These approximations are
reflected in the parameters 26 and 27 in (3.12).
Consider, next, the real interest component of P . This
equals the real market value of the bonds on which the
interest is paid multiplied by the foreign interest rate rf .
Because of inflation, the real market value of the bonds
will be less than the bond component of K a , which is the
real historic cost of the bonds. Since the nominal interest
paid on the bonds remains constant over their life, their
148
Model Specification
real market value falls during a period of rising prices. This

effect of inflation will be allowed for by multiplying the bond
component of K a by a linear function of the real foreign interest rate rf Dlog pf , as in (3.12), rather than
by rf .
3.15 Cumulative Net Real

Investment Abroad
Equation (3.13) assumes that the acceleration of K a depends
on its current rate of change and the excess of its partial equilibrium level over its current level. The partial equilibrium
level of K a is assumed to equal real income multiplied by
a linear function of the difference between the foreign and
domestic interest rates, the rate of change in the exchange
rate and a dummy variable to allow for exchange controls
(which were abolished in 1979). This is, essentially, a portfolio balance relation representing investors allocation of
capital between domestic and foreign assets, taking account
of the difference between the rates of return on these assets
and the expected change in the exchange rate. The parameter 30 reflects the speed with which the rate of change in
the exchange rate is expected to revert to the steady state
rate of change.
149
3.16 Exchange Rate

The exchange rate adjustment equation (3.14) represents
the dynamic behaviour of the foreign exchange market. The
acceleration of the logarithm of the exchange rate is assumed
to depend on the rate of decrease in the real exchange rate
Dlog( pf /qp), on the ratio of the expected steady state real
exchange rate 32 to the current real exchange rate, on the
excess of the current interest rate r over the expected steady
state interest rate 33 and on the ratio of the balance of payments surplus (including capital items other than the change
in the official reserves of gold and foreign exchange) to the
real value of imports. The last term allows for the effect of the
excess supply or demand for sterling resulting from a surplus or deficit in the balance of payments, while the previous
two terms allow for the effect of the speculative excess supply or demand for sterling resulting from the deviations of
the interest rate and real exchange rate from their expected
steady state levels.
As in the case of the price and wage adjustment equations
(3.6) and (3.7), the first term on the right-hand side of (3.14)
differs from the corresponding term in most of the secondorder equations in the model. Instead of the deviation of
150
Model Specification
the rate of change in the nominal exchange from its steady

rate of change (which depends on the rate of change in the
money supply), we have the deviation of the rate of change
of the real exchange rate from its steady state rate of change
(which is zero). The reason for this variation from the form
of most of the second-order equations is, as in the case of
(3.6) and (3.7), that it is unrealistic to assume that the agents
in the economy have any firm expectations concerning the
future rate of change in the money supply. When solving
the complete model for the steady state in Chapter 4, we
shall use the assumption of long-run rational expectations
by assuming that 32 and 33 are equal to the actual steady
state values of the real exchange rate and interest rate. But
this assumption will not be imposed when estimating the
parameters of the model.
3.17 Stocks
The stock adjustment equation (3.15) is, formally, an identity, although it is not exactly satisfied by the data, because of
errors in measurement. It will be treated as a stochastic differential equation with white noise innovation, therefore,
in the estimation of the parameters of the model.
151
3.18 Conclusion
In this chapter we have specified a new continuous time
econometric model of the United Kingdom. The main innovative feature of the new model is the incorporation of
stochastic trends to represent unobservable variables that,
in earlier models, have been represented by deterministic
trends. This and other general properties of the model have
been discussed in Section 3.2 and the specification of the
individual equations in Sections 3.33.17. Another important general property of the model is that, like the earlier
UK models of Bergstrom and Wymer [1976] and Bergstrom,
Nowman and Wymer [1992], it is designed in such a way as
to permit a rigorous mathematical analysis of its steady state
and stability properties. This analysis will be carried out in
Chapter 4, to which we now turn.
Appendix A: Derivation of General

Adjustment Equations
In this Appendix, we show how adjustment equations of the
form taken by most of the structural equations in the model
can be derived as the solution of a dynamic optimization
problem, which takes account of adjustment costs. We deal
152
Model Specification
first with the simplest case, in which the adjustment cost

depends only on the rate of change of the adjusting variables.
It will be shown, that, in this case, the optimal adjustment
equation is a first-order differential equation.
We assume that an economic agent has control of a single variable x(t) (where t is the time parameter), which
can be adjusted, continuously, in response to a vector z(t)
of other variables. The vector z(t) can include both exogenous variables and other endogenous variables in the model.
We define a partial equilibrium function f (z(t)), to which
x(t) would adjust instantaneously and follow continuously
if there were no adjustment costs; that is we would have
x(t) = f (z(t)) at all points of time if the cost of adjusting
x(t) were zero. The function f (z(t)) could be the value of
x(t) that maximizes some objective functiong(x(t), z(t)).
Assume, now, that there is an adjustment cost, which
depends only on the rate of change of x(t), but not on
its acceleration rate or any of its higher derivatives with
respect to t; that is we assume that the adjustment cost
depends only on d x(t)/dt. Then, if at time t, x(t) differs
from f (z(t)), the agent plans a path of x(t) that moves,
gradually, towards the expected path of f (z(t)), in such a
way as to minimize the integral of a cost function, which
takes account of both adjustment costs and costs associated
153
with the deviation of x(t) from f (z(t)). These costs can

be either monetary or psychological (loss of utility). We
assume that the cost function can be approximated by a
quadratic function, so that, if the future path of f (z(t))
were known, the integral to be minimized would be

2
2
0 [{x(t + r ) f (z(t + r ))} + a{d x(t + r )/dr } ]dr, where a
is a parameter reflecting the importance of adjustment
costs.
Since the future path of z(t) is unknown, it is necessary to
make some assumption about expectations. We shall assume
that z(t) is expected to change in such a way that f (z(t))
increases at a constant linear rate . Then, the expected
value, at time t, of f (z(t + r )) is given by
E[ f (z(t + r ))] = f (z(t)) + r.
(3.A1)
This assumption is consistent with the formulation of the

model in Section 3.2 of this chapter. There, nearly all of the
variables are measured in logarithms, and the formulation
of the adjustment equations, implicitly, assumes that their
partial equilibrium values are expected to increase at constant linear rates equal to their steady state growth rates. In
particular, if, at time t, a variable is at its partial equilibrium
level, it will not be held constant, but will be increased at a
154
Model Specification
rate equal to its steady state growth rate. This is consistent

with the assumption of long-run rational expectations.
Taking account of the above assumption with respect to
expectations and the assumption of a quadratic cost function, the optimal planned path, at time t, of x(t + r ) (r > 0)
is that which minimizes the integral

C {x(t + r ), d x(t + r )/dr }dr, where
0
C {x(t + r ), d x(t + r )/dr } = {x(t + r ) f (z(t)) r }2

+ a{d x(t + r )/dr }2 .
(3.A2)
A necessary condition for x(t + r )(r 0) to be the path of x

that minimizes 0 C {x(t + r ), d x(t + r )/dr }dr, is that it satisfies Eulers differential equation
d
Cx
C x = 0 (r 0),
dr
(3.A3)
where C x and C x are the partial derivatives of the function

C with respect to x(t + r ) and d x(t + r )/dr , respectively (see
Courant [1936, Vol. II, Ch. 7]).
From (3.A2) we obtain
C x = 2{x(t + r ) f (z(t)) r },
(3.A4)
C x = 2a{d x(t + r )/dr }.
(3.A5)
155
Substituting from (3.A4) and (3.A5) into (3.A3) and dividing

by 2a, we obtain the differential equation
d 2 x(t + r ) 1
{x(t + r ) f (z(t)) r } = 0
dr 2
a
(r 0)
(3.A6)
which must be satisfied by the optimal planned path of

x(t + r ) (r 0). Now let y(t + r ) be defined by
y(t + r ) = x(t + r ) f (z(t)) r,
(3.A7)
so that y(t + r ) is the deviation of x from its partial equilibrium level at time t + r. Then, from (3.A6) and (3.A7), we
obtain
d 2 y(t + r ) 1
y(t + r ) = 0
dr 2
a
(r 0),
(3.A8)
which can be written in the factorized form

(D )(D + )y(t + r ) = 0
(r 0),
(3.A9)
where = 1/ a and D = d/dr.

The general solution of the differential equation (3.A9)
is
y(t + r ) = c 1 e r + c 2 e r
(r 0),
(3.A10)
where c 1 and c 2 are arbitrary constants. However, since the

optimal planned path of y(t + r ) must converge to zero as
156
Model Specification
r , we must put c 2 = 0 and c 1 = y(t), so that the optimal planned path of y(t + r ) is given by
y(t + r ) = y(t)e r
(r 0).
(3.A11)
The solution given by (3.A11) satisfies the differential equation

dy(t + r )
= y(t + r ) (r 0),
(3.A12)
dr
and, from (3.A7) and (3.A12), we obtain
d x(t + r )
= + { f (z(t)) + r x(t + r )} (r 0),
dr
(3.A13)
which is the differential equation satisfied by the optimal

planned path of x(t + r ).
The actual path of x(t) will not, of course, follow the
optimal planned path, since z(t) is changing continuously.
However, d x(t)/dt will change continuously in response to
changes in z(t), in such a way that, at each point of time t,
x(t) will be starting to move along the optimal planned path
corresponding to the value of z(t) at that point of time. That
is

d x(t)
d x(t + r )
,
=
dt
dr
r =0
(3.A14)
where the derivative on the left-hand side of (3.A14) is

that of the actual path of x, while the derivative on the
157
right-hand side is that of the optimal planned path. It follows from (3.A13) and (3.A14) that the actual path of x
satisfies the differential equation
d x(t)
= + { f (z(t)) x(t)}.
dt
(3.A15)
Nearly all of the first-order differential equations in the

model formulated in Section 3.2 of this chapter have the
form of equation (3.A15). It says that the dependent variable x increases at a rate equal to the expected rate of
increase in its partial equilibrium level plus a speed of adjustment parameter times the current excess of its partial equilibrium level over its actual level. It follows from the definition of below (3.A9) that the speed of adjustment is
inversely related to the adjustment cost, as we should expect.
We turn now to the case in which the adjustment cost
depends on both the rate of change and acceleration of x(t);
that is on both d x(t)/dt and d 2 x(t)/dt 2 . Making the same
assumption about expectations as was made in the simpler case and assuming, again, that the cost function can be
approximated by a quadratic function, the optimal planned
path, at time t, of x(t + r ) is that which minimizes the
integral
158
Model Specification
C {x(t + r ), d x(t + r )/dr, d 2 x(t + r )/dr 2 }dr
where
C {x(t + r ), d x(t + r )/dr, d 2 x(t + r )/dr 2 }
= {x(t + r ) f (z(t) r )}2 + a{d x(t + r )/dr }2
+ b{d 2 x(t + r )/dr 2 }2
(3.A16)
and a and b are parameters measuring the importance of

first- and second-order adjustment costs.
A necessary condition for x(t + r ) (r 0) to be the path of

x that minimizes 0 C {x(t + r ), d x(t + r )/dr, d 2 x(t + r )/dr 2 }
dr , is that it satisfies Eulers differential equation
d
d2
C x + 2 C x = 0 (r 0),
dr
dr
Cx
(3.A17)
where C x , C x and C x are the partial derivatives of the function C with respect to x(t + r ), d x(t + r )/dr and d 2 x(t + r )/
dr 2 , respectively (see Courant [1936, Vol. II, Ch. 7]). From
(3.A16) we obtain
C x = 2{x(t + r ) f (z(t)) r },
(3.A18)
C x = 2a{d x(t + r )/dr },
(3.A19)
C x = 2a{d 2 x(t + r )/dr }.
(3.A20)
159
Substituting from (3.A18), (3.A19) and (3.A20) into (3.A17)

and dividing by 2b, we obtain the differential equation

d 4 x(t + r ) a d 2 x(t + r )
dr 4
b
dr 2
1
+ {x(t + r ) f (z(t)) r } = 0 (r 0), (3.A21)
b
which must be satisfied by the optimal planned path of
x(t + r ) (r 0). Then, from (3.A7) and (3.A21) we obtain

d 4 y(t + r ) a d 2 y(t + r )
1
+ y(t + r ) = 0 (r 0),
dr 4
b
dr 2
b
(3.A22)
which is the differential equation that must be satisfied by
the optimal planned path of y(t + r ) the deviation of x(t + r )
from its expected partial equilibrium level.
The solution of (3.A22) depends on the roots of the
equation
u4
a 2 1
u + = 0.
b
b
(3.A23)
We shall assume that all four roots of (3.A23) are real, in

which case the convergence of the optimal planned path of
x(t + r ) to its expected partial equilibrium path f (z(t)) + r,
as r , will be monotonic rather than oscillatory. The
assumption that the roots of (3.A23) are real, obviously,
implies that they have the form , , , , where and
160
Model Specification
are positive numbers, and the differential equation (3.A22)

can be written in the factorized form
(D + )(D )(D + )(D )y(t + r ) = 0
(r 0), (3.A24)
where D = d/dr.
The general solution of the differential equation (3.A24)
is
y(t + r ) = C 1 e r + C 2 e r + C 3 e r + C 4 e r,
(3.A25)
where C 1 , C 2 , C 3 and C 4 are arbitrary constants. However,

since the optimal planned path of y(t + r ) must converge to
zero as r , we must have C 3 = C 4 = 0. Moreover, C 1
and C 2 must satisfy the initial conditions:
[C 1 e r + C 2 e r ]r =0 = y(t),

d
{C 1 e r + C 2 e r }
dr
r =0
dy(t)
,
dt
(3.A26)
(3.A27)
or
C 1 + C 2 = y(t),
C 1 C 2 =
dy(t)
.
dt
(3.A28)
(3.A29)
Solving (3.A28) and (3.A29) for C 1 and C 2 , we obtain

C1 =
{y(t) + dy(t)/dt}
,
( )
161
(3.A30)
C2 =
{y(t) + dy(t)/dt}
.
( )
(3.A31)
The optimal planned path of y(t + r ) is given, therefore, by

y(t + r ) = C 1 e r + C 2 e r
(r 0),
(3.A32)
where C 1 and C 2 are given by (3.A30) and (3.A31).

The path of y(t + r ) given by (3.A32) satisfies the differential equation
(D + )(D + )y(t + r ) = 0
or
(r 0),
(3.A33)

d 2 y(t + r )
dy(t + r )
= 1
2 y(t + r ), (3.A34)
dr 2
dr
where 1 = + and 2 = . From (3.A7) and (3.A4) we

obtain

d 2 x(t + r )
d x(t + r )
1
dr 2
dr
+ 2 { f (z(t)) + r x(t + r )}
(r 0),
(3.A35)
which is the differential equation satisfied by the optimal

planned path of x(t + r ).
As in the case of the first-order adjustment process discussed earlier in this Appendix, the actual path of x(t)
will not follow the optimal path, because z(t) is changing
162
Model Specification
continuously. However, d 2 x(t)/dt 2 will change continuously

in response to changes in z(t), in such a way that, at each
point in time, x(t) will be starting to move along the optimal
planned path corresponding to the value of z(t) at that time.
That is

2
d 2 x(t)
d x(t + r )
=
dt 2
dr 2
r =0,
(3.A36)
where the second derivative on the left-hand side of (3.A36)

is that of the actual path of x and that on the right-hand side
is that of the optimal planned path. It follows from (3.A35)
and (3.A36) that the actual path of x satisfies the differential
equation

d 2 x(t + r )
d x(t + r )
+ 2 { f (z(t)) x(t)}.
= 1
dr 2
dr
(3.A37)
Nearly all of the second-order differential equations in the

model formulated in Section 3.2 of this chapter have the
form of (3.A37). It says that the dependent variable x accelerates at a rate equal to the speed of adjustment parameter
1 times the excess of the expected rate of increase of the
partial equilibrium level of x over its actual rate of increase
plus the speed of adjustment parameter 2 times the excess
its partial equilibrium level over its actual level.
163
Appendix B: Distributed Lag Relations

The first- and second-order adjustment equations (3.A15)
and (3.A37) derived in Appendix A each imply that the
value of the dependent variable at time t depends on all past
values of the explanatory variables; that is they depend on
z(t r ) (r 0). In this Appendix we investigate the precise
form of these distributed lag relations. See also Bergstrom
[1967], Bergstrom and Wymer [1976], Bergstrom, Nowan and Wymer [1992] and Gandolfo [1981] for these distributed lag relations.
Let the trend path of the partial equilibrium f (z(t)) in
(3.A15) be + t, so that, over a long period, f (z(t)) and
x(t) fluctuate about + t . We now define y(t) and g(z(t))
by
y(t) = x(t) ( + t),
(3.B1)
g(z(t)) = f (z(t)) ( + t).
(3.B2)
Then, (3.A15) implies that

dy(t)
= {g(z(t)) y(t)}.
dt
(3.B3)
We shall now show that the solution of (3.B3) is given by

y(t) =
e r g(z(t r ))dr
(3.B4)
0
164
Model Specification
Substituting s = t r, we obtain from (3.B4)

t
y(t) =
e (ts ) g(z(s ))ds
t
t
e s g(z(s ))ds.
= e
(3.B5)
Then, differentiating (3.B5), we obtain

t
t
dy(t)
2 t
s
t d
s
e g(z(s ))ds + e
e g(z(s ))ds
= e
dt
dt
t
e (ts ) g(z(s ))ds + e t e t g(z(t))
= 2
= {g(z(t)) y(t)},
which shows that the expression for y(t) given by (3.B4) is

the solution (3.B3).
A verbal interpretation of (3.B4) is that y(t) depends with
an exponentially distributed time lag on the past values of
g(z(t)). The density function of the time lag distribution is

e r , and the mean of the distribution is 0 r e r dr = 1 .
We can derive the distributed lag relation implied by the
second-order adjustment equation in a similar way.
From (3.A37), (3.B1) and (3.B2), we obtain
d 2 y(t)
dy(t)
= 1
+ 2 {g(z(t)) y(t)}.
dt 2
dt
(3.B6)
The solution of (3.B6) depends on the roots of the equation

w 2 + 1 w + 2 = 0.
165
(3.B7)
We shall assume that the roots of (3.B7) are real and negative, in which case y(t) converges monotonically to g(z(t))
as t . Letting and denote the two roots, we have
+ = 1 and = 2 , so that (3.B6) can be written
dy(t)
d 2 y(t)
+ ( + )
+ y(t) = g(z(t)).
dt 2
dt
(3.B8)
The form of the time lag distribution with which y(t)

responds to variations in g(z(t)) depends on whether the
roots and are equal or different. We shall deal, first,
with the case in which they are different, and assume that
> > 0. We shall now show that, in this case, the solution
of (3.B8) is given by
y(t) =
(r )g(z(t r ))dr,
(3.B9)
where
(r ) =
{e r e r }.

(3.B10)
Substituting s = (t r ), we obtain from (3.B9) and (3.B10)

y(t) =
e (ts ) g(z(s ))ds

e (ts ) g(z(s ))ds .
(3.B11)
166
Model Specification

dy(t)
dt
=

t
d t (ts )
d
e (ts ) g(z(s ))ds
e
g(z(s ))ds .
dt
dt
(3.B12)
But,
d
dt

t
d
e (ts ) g(z(s ))ds =
e s g(z(s ))ds
e t
dt
t
t
d
e s g(z(s ))ds + e t
e s g(z(s ))ds
= e t
dt
t
=
e (ts ) g(z(s ))ds + e t e t g(z(t))
t
e (ts ) g(z(s ))ds + g(z(t)),
(3.B13)
=
and, similarly,
t
t
d
e (ts ) g(z(s ))ds =
e (ts ) g(z(s ))ds + g(z(t)).
dt
(3.B14)
From (3.B12), (3.B13) and (3.B14), we obtain
t
dy(t)
e (ts ) g(z(s ))ds

=
dt

t
(ts )
e
g(z(s ))ds .
167
(3.B15)
Differentiating (3.B15) and using (3.B13) and (3.B14), we

obtain
t
d 2 y(t)
=
e (ts ) g(z(s ))ds
2
dt 2

t
e (ts ) g(z(s ))ds + g(z(t)).
2
(3.B16)

d 2 y(t)
dy(t)
+ y(t) = g(z(t)),
+ ( + )
dt 2
dt
which shows that y(t) as defined by (3.B9) and (3.B10) is
the solution of (3.B8) and, hence, of (3.B6). It follows that
y(t) depends on z(t) through the distributed time lag relation
(3.B9), the density function (r ) of the lag distribution being
given by (3.B10).
The mean of this distribution (the mean time lag) is given
by
r (r )dr =
=
=
=
=

1
r
r
r e dr
re dr ,

0
0

1
2

2
1
,

+
,
1
1
+ .
(3.B17)
168
Model Specification
The mode of the lag distribution (the modal time lag) is

given by
arg max(r ) =
log log
.

(3.B18)
To prove (3.B18), we note that argmax(r ) is the value of r

that satisfies the condition
d(r )
= 0.
dr
(3.B19)
It follows from (3.B10) that (3.B19) will be satisfied if and

only if
e r = e r ,
(3.B20)
log r = log r.
(3.B21)
and hence,
Solving (3.B21) for r, we obtain the expression on the righthand side of (3.B18).
We shall deal, finally, with the case in which the roots of
(3.B7) are equal, each being equal to . In this case, (3.B6)
can be written
d 2 y(t)
dy(t)
+ 2
+ 2 y(t) = 2 g(z(t)).
dt 2
dt
169
(3.B22)
We shall now show that the solution of (3.B22) is given by

y(t) =
(r )g(z(t r ))dr,
(3.B23)
0
where
(r ) = 2 r e r .
(3.B24)
Substituting s = (t r ), we obtain from (3.B23) and

(3.B24)
y(t) =
2 (t s )e (ts ) g(z(s ))ds
t
te t e s g(z(s ))ds 2
s e t e s g(z(s ))ds
t
t
e s g(z(s ))ds 2 e t
s e s g(z(s ))ds.
= 2 te t
= 2
(3.B25)

t
dy(t)
e s g(z(s ))ds { 2 e t 3 te t }
=
dt
t
s e s g(z(s ))ds { 3 e t }
+
+ tg(z(t)) 2 tg(z(t))
t
e s g(z(s ))ds { 2 e t 3 te t }
=
t
s e s g(z(s ))ds { 3 e t }.
(3.B26)
+
170
Model Specification

t
d 2 y(t)
=
e s g(z(s ))ds {2 3 e t + 4 te t }
dt 2
t
s e s g(z(s ))ds { 4 e t }
+
+ 2 g(z(t)) 3 tg(z(t)) + 3 tg(z(t))

t
e s g(z(s ))ds {2 3 e t + 4 te t }
=
t
s e s g(z(s ))ds { 4 e t } + 2 g(z(t)).
+
(3.B27)

dy(t)
d 2 y(t)
+ 2 y(t) = 2 g(z(t)),
+ 2
dt 2
dt
which shows that y(t), as defined (3.B23) and (3.B24), is the
solution of (3.B22) and, hence, of (3.B6). It follows that y(t)
depends on z(t) through the distributed time lag relation
(3.B23), the density function (r ) of the lag distribution
being given by (3.B24).
The distribution with density function given by (3.B24) is
a special case of the Gamma distribution. The density function f (x) of the Gamma distribution, in the general case, is
given by
f (x) =
1 x
x e ,
()
171
where and are parameters and () is the Gamma function defined by

() =
x 1 e x d x.
In the special case where = 2, we obtain the density function defined by (3.B24). The mean of the Gamma distribution is /, and hence the mean of the distributed lag defined
by (3.B24) is 2/. Moreover, it is easily verified that the mode
of the latter distribution is 1/.
172
CHAPTER FOUR
Steady State and

Stability Analysis
4.1 Introduction
n important feature of the continuous time macroeconometric model specified in this monograph is that,
like the model of Bergstrom and Wymer [1976] and the

many models for which that model has served as a prototype, it is designed in such a way as to permit a rigorous
mathematical analysis of its steady state and stability properties, thus providing an important check on its capacity to
generate plausible long-run behaviour. Such a steady state
and stability analysis is carried out in this chapter. In Section 4.2, we show that, when the exogenous variables satisfy
certain conditions and the stochastic innovations have their
expected value zero, the model has a steady state solution
in which all variables grow at constant exponential rates
173
(in some cases zero). Closed formulae from which the exact
steady state paths of the variables can be computed are
derived in Appendix A. In Section 4.3 and Appendix 4.3,
we derive, from the original model, an explicit differential
equation system in the logarithms of the ratios of the variables to their steady state paths and show that the nonlinear
part of this system satisfies the PoincareLiapounovPerron

conditions (see Bellman [1953, p. 93] and Coddington and
Levinson [1955, p. 314]). The latter conditions ensure that
the asymptotic stability of the steady state solution derived
in Section 4.2 depends on the eigenvalues of the linear
part of this system; that is it depends on the eigenvalues
of the matrix of coefficients of the linear part of this system when it is written in the form of a first-order system of higher dimension. Section 4.4 reviews some previous work on fiscal and monetary policy in the Bergstrom,
Nowman and Wymer [1992] model and results concerning bifurcation analysis. Section 4.5 concludes. The results
obtained in this chapter will be applied to the empirical estimates obtained from UK data in Chapter 5. There
we shall derive the steady state solution of the estimated
model and investigate the stability and cyclical properties
of the estimated model in the neighbourhood of the steady
state.
174
4.2 The Steady State

The steady state solution of the model is derived under the
assumptions represented by (4.1)(4.17).
B = B e 4 t ,
(4.1)
dx = 0,
(4.2)
E 0 = E 0 e (1 +2 )t ,
(4.3)
Gc = g (Q + P ),
(4.4)
K p = K p e (1 +2 )t ,
(4.5)
M = M e 4 t ,
(4.6)
pf = pf e 5 t ,
(4.7)
pi = pi e 5 t = pf e 5 t ,
(4.8)
rf = rf ,
(4.9)
T1 = T1 ,
(4.10)
T2 = T2 ,
(4.11)
Yf =
Y f e
1 +2
23
175
"
t
(4.12)
32 =
q p
pf
(4.13)
33 = r ,
(4.14)
1 (0) = 0,
(4.15)
2 (0) = 0,
(4.16)
3 (0) = 0,
(4.17)
where B , E 0 , g , K p , M , pf , pi , rf , T1 , T2 , Y f , 4 and 5 are

given constants, while q , p and r are functions of the
parameters of the model given by the long-run rational
expectations steady state solution. The assumptions concerning 32 and 33 can be interpreted as long-run rational
expectations in the foreign exchange markets.
It can easily be verified that, under the above assumptions,
the system of (3.1)(3.18) has a steady state solution of the
form:
C = C e (1 +2 )t ,
(4.18)
E n = E n e (1 +2 )t ,
(4.19)
F = F e (1 +2 )t ,
(4.20)
I = I e (1 +2 )t ,
(4.21)
176
K = K e (1 +2 )t ,
(4.22)
K a = K a e (1 +2 )t ,
(4.23)
K h = K h e (1 +2 )t ,
(4.24)
L = L e 2 t ,
(4.25)
p = p e (3 +4 1 2 )t ,
(4.26)
P = P e (1 +2 )t ,
(4.27)
q = q e (1 +2 +5 3 4 )t ,
(4.28)
Q = Q e (1 +2 )t ,
(4.29)
r = r ,
(4.30)
S = S e (1 +2 )t ,
(4.31)
w = w e (3 +4 2 )t ,
(4.32)
1 = 1 t,
(4.33)
2 = 2 t,
(4.34)
3 = 3 t,
(4.35)
177
where C , E n , F , I , K , K a , K h , L , p, P , q , Q , r , S and
w are constants (hereafter referred to as steady state level
parameters) satisfying the nonlinear simultaneous equation
system (4.36)(4.50).
C = 1 e

Q + P
,
T1
1/6
5 (K )6
,
{(2 3 )(3 +4 1 2 )2 r }

L = 4 (Q )6
(4.36)
(4.37)

Q + P
, (4.38)
= 7 e
T1
1+6
Q
, (4.39)
r = 10 (3 + 4 1 2 ) 11 + 5
K

1 12 q p /pi 13 {1 + 14 (1 + 2 )}
C +G +(1 +2 ) K + K + K + E + E
p
n
o
c
h
8 log
K h
+ 9 log
{(8 9 )(3 +4 1 2 )8 r }

14 C + Gc + (1 + 2 ) K + K h + K p
+E n + E o
= 0,
p = 15 4 T2 w {1 5 (Q /K )6 }(1+6 )/6 ,
6
(Q )
5 (K )

=
178
16
4
(4.40)
(4.41)
6
,
(4.42)
r = 18 rf + 20 { p (Q + P )/M } + 21 (B /M )
+ 17 + 19 (3 + 4 1 2 5 ),
(4.43)

{12 q p /pi 13 }{1 + 14 (1 + 2 )}

C +G +(1 +2 ) K + K + K + E + E
p
n
o
c

h
17 log
pi /q p I

14 C + Gc + (1 + 2 ) K + K h + K p

+E + E
n
o
+ 18 log
= 0,
(4.44)
E n

25

22 Y f 23
Q + P
F

=
pf
q p
24
,
20 + 1 + 2
,
20
!
" K
21 + 1 + 2
a
=
,
26 + 27 rf 5
P
21
(4.45)
(4.46)
(4.47)
!
"
Q + P
28 + 29 rf r + 30 (3 + 4 1 2 5 )
K a
2
23 + (1 + 2 ) + 22 (1 + 2 )
=
,
(4.48)
23
179
(The last equation is obtained by using the relation

D 2K a /K a = D 2 logK a + (DlogK a )2
= (DlogK a )2 when K a = K a e (1 +2 )t ).
pi I
= E n + E o + P F (1 + 2 )K a ,
q p
(4.49)
p I
(1 + 2 )S = Q + i C + Gc + (1 + 2 )
q p

K + K h + K p + E n + E o .
(4.50)
A procedure for solving the equation system (4.36)(4.50)

is derived in Appendix A. The first step is to derive from
(4.36)(4.50) a pair of simultaneous equations (4.A12) and
(4.A19) in the steady state interest rate r and the steady
state real exchange rate x =
ing
pi
p f ).
q p
p f
q p
pi
(implicitly, assum-
These must be solved, numerically, for r and
x . Then, the remaining steady state level parameters can be

computed, recursively, from a sequence of closed formulae
set out in Appendix A.
4.3 Stability Analysis

The stability analysis is based on a transformed model in the
variables y1 (t), y2 (t), . . . , y15 (t), defined as follows.
180
y1 (t) =
y3 (t) =
y5 (t) =
y7 (t) =
y9 (t) =
y11 (t) =
y13 (t) =
y15 (t) =

C (t)
L(t)
log
(t)
=
log
,
y
,
2
C e (1 +2 )t
L e 2 t

K h (t)
K (t)
,
y4 (t) = log
,
log
K h e (1 +2 )t
K e (1 +2 )t

Q (t)
p(t)
,
y
,
log
(t)
=
log
6
Q e (1 +2 )t
p e (3 +4 1 2 )t

w(t)
, y8 (t) = r (t) r ,
log
w e (3 +4 2 )t

I (t)
E n (t)
log ( + )t ,
y10 (t) = log
,
I e 1 2
E n e (1 +2 )t

F (t)
P (t)
,
y
,
log
(t)
=
log
12
F e (1 +2 )t
P e (1 +2 )t

K a (t)
q(t)
,
y
,
log
(t)
=
log
14
K a e (1 +2 )t
q e (1 +2 +5 3 4 )t

S(t)
log ( + )t .
S e 1 2
The transformed model is derived from the original model

(3.1)(3.18) in Appendix B and comprises the system of
nonlinear differential equations (4.51)(4.65).
D y1 = 1 {log(Q e 5 + P e y12 ) log(Q + P ) 2 y8
+ (2 3 )D y6 y1 },
(4.51)

1
log{(Q )6 5 (K )6 }
D 2y2 = 2 D y2 + 3
6

1
y5 6
y4 6
log{(Q e )
5 (K e ) } y2 , (4.52)
6
181
D 2y3 = 4 D y3 + 5 {log(Q e 5 + P e y12 ) log(Q + P )

8 y8 + (8 9 )D y6 y3 },
(4.53)
D 2y4 = 6 D y4 + 7 [(1 + 6 ) (y5 y4 )

+ log{r 10 (3 + 4 1 2 ) + 11 }
log{y8 + r 10 (D6 + 3 + 4 1 2 ) + 11 }],
y6 +y14
(4.54)
13
q p e
pi

13
q p
log 1 12
pi

+ log C e y1 + Gc + K e y4 (D y4 + 1 + 2 )
D y5 = 8 log 1 12

+ K h e y3 (D y3 + y1 + y2 ) +K p (1 + 2 ) + E n e y10 + E o

log C +Gc +(1 +2 ) K + K h + K p + E n + E o y5

+ 9 log C e y1 + Gc + K e y4 (D y4 + 1 + 2 )

+ K h e y3 (D y3 + y1 + y2 )+ K p (1 + 2 )+ E n e y10 + E o

log C +Gc +(1 +2 ) K + K h + K p + E n + E o y15 ,
(4.55)

1
+
6
D 2y6 = 10 (D y7 D y6 ) + 11 y7 y6
6

y5 6

1 + 6
Q e
log 1 5
+
K e y4
6

6
Q
log 1 5
,
(4.56)
K

182
D 2y7 = 12 (D y6 D y7 ) 13 (D y6 + D y14 )

1
6
log (Q )6 5 (K )
+ 14
6

1
y5 6
y4 6
log{(Q e )
5 (K e ) } , (4.57)
6

p e y6 (Q e y5 + P e y12 )
D y8 = 15 D y8 + 16 20
M

p (Q + P )
20
(4.58)
D
y
y
19
14
8 ,
M
2

D y9 = 17 (1 + 13 )(y6 + y14 ) y9

+ log C e y1 + Gc + K e y4 (D y4 + 1 + 2 )

+ K h e y3 (D y3 + y1 + y2 ) + K p (1 + 2 ) + E n e y10 + E o

log C + Gc + (1 + 2 ) K + K h + K p + E n + E o

+ 18 log C e y1 + Gc + K e y4 (D y4 + 1 + 2 )

+ K h e y3 (D y3 + y1 + y2 ) + K p (1 + 2 ) + E n e y10 + E o

log C + Gc + (1 + 2 ) K + K h + K p

(4.59)
+ E n + E o y15 ,
D y10 = 19 [24 (y6 + y14 ) + y10 ],

D y11 = 20 25
Q e y5 + P e y12
F e y11
183
Q + P
F
(4.60)

,
(4.61)

!
" K e y13
K a
a
D y12 = 21 26 + 27 rf 5
,
P e y12
P
(4.62)
D 2y13 = {22 + 2(1 + 2 )}D y13 (D y13 )2

!
"
+ 23 28 + 29 rf r y8

30 (D y14 + 1 + 2 + 5 3 4 )

y5
!
"
Q e + P e y12
28 + 29 rf r
y
K a e 13

Q + P
30 (1 + 2 + 5 3 4 )
,
K a
(4.63)
D 2y14 = 24 (D y6 + D y14 ) 25 (y6 + y14 ) + 26 y8

+ 27 log E n e y10 + E o + P e y12 F e y11

K e y13 (D y13 + 1 + 2 )

log E n + E o + P F (1 + 2 )K a

(4.64)
+ y6 + y14 y9 ,

D y15 = Q e y5 + pi /q p I e y9 y6 y14 C e y1
Gc K e y4 (D y4 + y1 + y2 ) K h e y3 (D y3 + y1 + y2 )

K p (1 +2 ) E n e y10 E o /S e y15 Q + pi /q p I

C Gc (1 + 2 ) K + K h + K p

(4.65)
E n E o /S .
184
It can, easily, be verified that the system (4.51)(4.65) has

a steady state solution
yi (t) = 0
(i = 1, . . . , 15).
(4.66)
Moreover, it is clear from the definitions of the variables

yi (t) = 0(i = 1, . . . , 15) that the solution (4.66) corresponds
to the steady state solution of the original model. For, it
implies that the variables C(t), E n (t), . . . , w(t) satisfy (4.18)
(4.32).
Next, for each of equations (4.51)(4.65), we shall express
the function on the right-hand side of the equation as the
sum of two terms, the first being a linear function of the
variables y1 , y2 , . . . , y15 and the second a nonlinear function of lower order of smallness as yi 0(i = 1, . . . , 15).
For this purpose, we define y = [y1 , y2 , . . . , y15 ] and |y| =
15
i=1 |yi |. Then, by a Taylor series expansion to the first
order, about the origin, of the functions on the righthand side of (4.51)(4.65), we obtain the system (4.67)
(4.81), in which the functions f i (y)(i = 1, . . . , 15) satisfy
the PoincareLiapounovPerron
conditions f i (y)/|y| 0 as
|y| 0 (i = 1, . . . , 15).
D y1 = 1 y1 1 2 y8 1 (3 2 )D y6 +

1 P
+
y12 + f 1 (y),
Q + P
185
1 Q
Q + P

y5
(4.67)

3 (Q )6
y5
D 2y2 = 2 D y2 3 y2 +
(Q )6 5 (K )6

3 5 (K )6
y4 + f 2 (y),
(4.68)
(Q )6 5 (K )6
D 2y3 = 4 D y3 + 5 (8 9 )D y6 5 y3 5 8 y8

5 Q
5 P
y
y12 + f 3 (y), (4.69)
+
+
5
Q + P
Q + P

D 2y4 = 6 D y4 +
7 10
r 10 (3 + 4 1 2 ) + 11

D y6
7 (1 + 6 )y4 + 7 (1 + 6 )y5

7
y8 + f 4 (y),
r 10 (3 + 4 1 2 ) + 11
(4.70)

8 12 13 (x )13
D y5 = 5 y5 9 y15
(y6 + y14 ) + (8 + 9 )
1 12 (x )13

C y1 + (1 + 2 )K y4 + K D y4 + (1 + 2 )K h y3
+ K h D y3 + E n y10

C + Gc + (1 + 2 ) K + K h + E n + E o
+ (1 + 2 )K p
+ f 5 (y),
(4.71)
D 2y6 = 10 (D y6 D y7 ) 11 (y6 y7 )
! "6
11 5 (1 + 6 ) KQ
+
(y5 y4 ) + f 6 (y),

6
1 5 KQ
186
(4.72)
D 2y7 = 12 (D y6 D y7 ) 13 (D y6 + D y14 )
6

(Q ) y5 5 (K )6 y4
+ 14
+ f 7 (y),
(Q )6 5 (K )6
(4.73)
D 2y8 = 15 D y8 16 19 D y14 16 y8

p (Q + P )
p Q
y6 + 16 20
y5
+ 16 20
M
M

p Q
y12 + f 8 (y),
+ 16 20
(4.74)
M
D y9 = 17 y9 18 y15 + 17 (1 + 13 )(y6 + y14 ) + (17 + 18 )

C y1 + (1 + 2 )K y4 + K D y4 + (1 + 2 )K h y3
+ K h D y3 + E n y10

C + Gc + (1 + 2 ) K + K h + K p + E n + E o
+ f 9 (y),
(4.75)
D y10 = 19 y10 19 24 (y6 + y14 ),

D y11

Q
P
y5 + 20 25
y12
= 20 25
F
F

Q + P
y11 + f 11 (y),
20 25
F
(4.76)
(4.77)

!
" K
a
(y13 y12 ) + f 12 (y),
D y12 = 21 26 + 27 rf 5
P
(4.78)
187
D y13

Q + P
y8
= {22 + 2(1 + 2 )}D y13 23 29
K a

!
"
Q + P
D
y
r
23 30
+
r
14
23
28
29
f
K a

Q
+ 30 (3 + 4 1 2 5 )}
y5
K a

!
"
+ 23 28 + 29 rf r + 30 (3 + 4 1

!
"
P
r
2 5 )
r
y
12
23
28
29
f
K a

+ 30 (3 + 4 1 2 5 )

Q + P
y13 + f 13 (y),
(4.79)
K a
D 2y14 = 24 (D y6 + D y14 ) 25 (y6 + y14 ) + 26 y8

+ 27 (y6 + y14 y9 )

E n y10 + P y12 F y11 (1 + 2 )K a y13
K D y13
a
+ 27
E n + E o + P F (1 + 2 )K a
+ f 14 (y),
(4.80)

pi I
Q
C
y5 +
(y9 y6 y14 )
y1
=
S
q p S
S

(1 + 2 )K h
(1 + 2 )K
K
y4
D y4
y3
S
S
S

D y15
188
K h
S

D y3
E n
S

y10
p I
Q + q i p C Gc (1 + 2 )
K + K + K E E
p
n
o
h
y15 + f 15 (y).
(4.81)
Finally, we represent the mixed-order system (4.67)(4.81)

as a first-order system
Du = C u + (u),
(4.82)
where C is a 23 23 matrix, whose elements are functions of the parameters of the model (3.1)(3.18) and u =
[u1 , u2 , . . . , u23 ], = [1 , 2 , . . . , 23 ], ui (i = 1, . . . , 23) and
i (i = 1, . . . , 23) are defined as follows.
u1 (t) = y1 (t), u2 (t) = y5 (t), u3 (t) = y9 (t), u4 (t) = y10 (t),
u13 (t) = y8 (t), u14 (t) = y13 (t), u15 (t) = y14 (t), u16 (t) = D y2 (t),
u17 (t) = D y3 (t), u18 (t) = D y4 (t), u19 (t) = D y6 (t), u20 (t)
= D y7 (t),
u21 (t) = D y8 (t), u22 (t) = D y13 (t), u23 (t) = D y14 (t),
1 (u) = f 1 (y), 2 (u) = f 5 (y), 3 (u) = f 9 (y), 4 (u) = f 10 (y),
189
5 (u) = f 11 (y), 6 (u) = f 12 (y), 7 (u) = f 15 (y), i (u)

= 0(i = 8, . . . , 15),
16 (u) = f 2 (y), 17 (u) = f 3 (y), 18 (u) = f 4 (y), 19 (u)
= f6 (y),
20 (u) = f 7 (y), 21 (u) = f 8 (y), 22 (u) = f 13 (y), 23 (u)
= f 14 (y).
It should be noted that the ordering of the variables

ui (i = 1, . . . , 23) differs from that of the variables yi to
which they correspond. For example, u2 = y5 (not y2 ). The
variables u1 , u2 , . . . , u7 are transformations of the variables
in the original model (3.1)(3.18) that adjust through a
first-order differential equation, the variables u8 , u9 , . . . , u15
are transformations of the variables that adjust through
a second-order differential equation, and the variables
u16 , u17 , . . . , u23 are the rates of change of the variables
u8 , u9 , . . . , u15 . Similarly, the ordering of the functions
i (i = 1, . . . , 23) differs from that of the functions f i to
which they correspond.
It should be noted, also, that each of the rows 815 of the
matrix C has one element equal to 1 and all other elements
equal to 0. This is because rows 815 contain the coefficients
of the identity relations
Dui = ui+8
(i = 8, . . . , 15).
190
It is for the same reason that i (u) = 0 (i = 8, . . . , 15). Since

ui = 0 (i = 1, . . . , 23) implies that yi = 0 (i = 1, . . . , 15) and
the system (4.51)(4.65) has a zero solution, the system
(4.82) has a zero solution u(t) = [0, . . . , 0] . The latter solution is said to be asymptotically stable if there exists a > 0
such that, for any solution u(t) of the system (4.82) satisfying the condition |u(0)| < , |u(0)| 0 as t (see
Coddington and Levinson [1955, p. 314]).
As was mentioned earlier, the properties of the higher
order terms in the Taylor series expansion imply that
the functions f i (y)(i = 1, . . . , 15) satisfy the Poincare

LiapounovPerron
conditions
f i (y)/|y| 0
as
|y|
0 (i = 1, . . . , 15). Hence, the functions i (u) (i = 1, . . . , 23)

must, also, satisfy these conditions: that is i (u)/|u| 0
as |u| 0 (i = 1, . . . , 23). (The conditions are, trivially,
satisfied from the functions 8 (u), . . . , 15 (u), which are
equal to zero.) It follows, therefore, by a theorem of Perron
(see Coddington and Levinson [1955, p. 314]), and by
earlier results of Poincare and Liapounov (see Bellman
[1953, p. 93]), that the zero solution of the system (4.82)
is asymptotically stable if the eigenvalues of C all have
negative real parts.
If the eigenvalues of C all have negative real parts, we
shall, also, describe the steady state solution (4.18)(4.35)
191
of the original model (3.1)(3.18) as asymptotically stable. For, convergence of the variables ui (t) (i = 1, . . . , 23) to
zero implies that the proportional deviations of the variables
C(t), E n (t), . . . , w(t) from their steady state paths converge
to zero. The steady state level parameters are of interest in
themselves, and, assuming they have, already, been computed, the additional cost of computing the unknown elements of C and the eigenvalues of this matrix is comparatively light. Finally, it should be noted that the eigenvalues
of C provide information not only about the stability of the
steady state solution but also about the cyclical properties of
a solution of the model in the neighbourhood of the steady
state.
4.4 Stability and Bifurcations

In this section we discuss some work on stability and bifurcation analysis, using the Bergstrom, Nowman and Wymer
[1992] model. As the Bergstrom, Nowman and Wymer
[1992] models steady state computed from the estimated
parameters was slightly unstable (although, statistically they
could not reject the hypothesis that it is stable) the investigation of methods to reduce or eliminate the instability
192
was undertaken by Bergstrom, Nowman and Wandasiewicz

[1994]. They investigated the use of a simple fiscal policy
feedback and the use of complicated control theory considered in Bergstrom [1987] that could reduce or eliminate
the instability. The simple fiscal policy feedback was introduced by Bergstrom [1984b] in which tax rates are reduced
and (or) transfer payments, such as social security benefits, increased during a recession and vice versa during a
boom. Simple Keynesian analysis suggests that such a policy is likely to have a stabilizing influence.
The policy feedback used was specified as follows:

DlogT1 = log
Q
Q e (1 +2 )t

log
T1
T1

.
It allows at each point in time the value of the fiscal policy

variable T1 to adjust towards a partial equilibrium level that
is an increasing function of the ratio of output to its steady
state level. The parameters can be interpreted as follows:
measures the strength of the feedback, the speed of
adjustment parameter that reflects delays in obtaining new
information about the level of output (and administrative
delays in changing tax and benefit rates) and T1 is the steady
state level of the fiscal policy variable. They also considered
193
the limiting case of in which case the feedback relation becomes instantaneous. Bergstrom, Nowman and Wandasiewicz [1994] found that this simple fiscal policy feedback
in cases of either instantaneously or with an exponential
time lag had no definite stabilizing influence on the dynamic
behaviour of the model.
In Bergstrom, Nowman and Wandasiewicz [1994] they
also investigated stability analysis, using the more advanced
approach of optimal control theory. They proved the existence of stabilizing feedbacks involving the fiscal policy variable and the money supply as combined policy instruments.
They also derived optimal feedbacks by minimizing an infinite horizon quadratic cost function involving deviations of
output, the exchange rate and the monetary and fiscal policy instruments from their steady state paths (and also, the
rates of change of the policy instruments). One of the findings was that fiscal policy had an important role to play
in the stabilization of the model than does monetary policy (see Bergstrom, Nowman and Wandasiewicz [1994] for
full details). In a more recent study, Bergstrom [2001] also
investigated the stability of the Bergstrom, Nowman and
Wymer [1992] model, using the estimates of the model
in Nowman [1996] by looking at the wage-acceleration
parameter that determines the effect of labour demand
194
on wage acceleration in the model. It was shown that

there was a bounded range of positive values of the wage
acceleration parameter in which the steady state was asymptotically stable.
We finally discuss a series of important papers by Barnett
and He [1998, 1999, 2002] (see references in these articles for related literature). In Barnett and He [1999] they
investigated bifurcations in the Bergstrom, Nowman and
Wymer [1992] model. Using the estimates of the Bergstrom,
Nowman and Wymer [1992] model, they used the gradient method to find a set of parameters where all the
eigenvalues had strictly negative real parts and the system is stable. They then developed a numerical algorithm
to obtain stability boundaries and applied it to particular
cases of parameters of the model. Two types of boundaries were considered, which could occur according to the
way unstable eigenvalues are created. They found that the
Bergstrom, Nowman and Wymer [1992] model displayed
an interesting set of bifurcations that included transcritical bifurcations, Hopf bifurcations and the important class
of co-dimension two bifurcations was confirmed. In Barnett and He [1998], which is a continuation of the research
in Barnett and He [1999], they investigated the dynamic
behaviour on bifurcation boundaries in the Bergstrom,
195
Nowman and Wymer [1992] model and found that the

model in certain cases is unstable on bifurcation boundaries. They also derived a new formula for determining the
analytical forms of the transcritical bifurcation boundaries
and investigated the statistical significance of the instability inference. Barnett and He [1998, 2002] also investigated the use of the simple ad hoc fiscal policy feedback
rule used in Bergstrom, Nowman and Wandasiewicz [1994]
and found that the models stable region became smaller
and argued that the policy may be counter-productive. See
Barnett and He [1998, 1999, 2002] for further discussions
on the issues of the Lucas critique, time inconsistency (see
Kydland and Prescott [1977]), use of control theory and a
number of interesting numerical examples and figures illustrating bifurcation boundaries. Barnett and He [2001] try
to resolve some controversies with respect to nonlinearity,
chaos and bifurcation in macroeconomic models and also
used the Bergstrom, Nowman and Wymer [1992] model as
part of their analysis. Finally a more recent series of interesting papers by Barnett and He [2004, 2005a, b, 2006]
investigates singularity-induced bifurcations that they introduce into economics and its relevance to macroeconometric
models.
196
4.5 Conclusion
We have shown, in this chapter, that, under certain assumptions concerning the paths of the exogenous variables,
the continuous time macroeconometric model specified in
Chapter 3 has a steady state solution in which all variables
grow at constant proportional rates (in some cases, zero). We
have, also, derived sufficient conditions for the asymptotic
stability of the steady state solution in terms of the coefficients of the linear part of a transformed model. We now
turn to some empirical results from UK data in the next
chapter.
Appendix A: Steady State Level Parameters

In this Appendix, we derive a computational procedure for
solving the simultaneous equation system (4.36)(4.50) for
the steady state level parameters. As was mentioned in Section 4.2, the first step in this computational procedure is to
solve a pair of simultaneous equations in the steady state
interest rate r and the steady state real exchange rate x .
This pair of simultaneous equations, (4.A12) and (4.A19),
is derived as follows.
197
From (4.47) and (4.48) we obtain

P
= h(r )
Q + P
where
h(r ) =

!
"
!
"
21 23 26 + 27 rf 5 {28 + 29 rf r
+ 30 (3 + 4 1 2 5 )}
(21 + 1 + 2 ){23 + (1 + 2 )2 + 22 (1 + 2 )}
Therefore,
Q + P
1
=
.
Q
1 h(r )
From (4.36) and (4.A1), we obtain

C
1
1 e {(2 3 )(3 +4 1 2 )2 r }
,
=
Q
1 h(r )
T1
(4.A1)
(4.A2)
From (4.4) and (4.A1),

Gc
g
,
=
Q
1 h(r )
and from (4.38) and (4.A1),

K h
1
7 e {(8 9 )(3 +4 1 2 )8 r }
.
=
Q
1 h(r )
T1
From (4.39), we obtain

r 10 (3 + 4 1 2 ) + 11 1(1+6 )
K
=
.
Q
5
198
(4.A3)
(4.A4)
(4.A5)
Equations (4.40), (4.44) and (4.50) can be regarded as

a set of simultaneous equations in Q , pi I /q p and S ,
which can be solved to express each of these variables as a
function of C , Gc , K , K h , K p , E n , E o and q p /pi . It can,
easily, be verified, by substitution into (4.40), (4.44) and
(4.50), that the resulting expressions for Q , pi I /.q p and
S are

13

p
q
{1 + 14 (1 + 2 )} C + Gc
Q = 1 12
pi

+ (1 + 2 ) K + K h + K p + E n + E o ,
(4.A6)

13

{1 + 14 (1 + 2 )} C + Gc + (1 + 2 )

(4.A7)
(K + K h + K p ) + E n + E o ,

S = 14 C + Gc + (1 + 2 ) K + K h + K p + E n + E o .
pi I
= 12
q p
q p
pi
(4.A8)
From (4.42), we obtain

6

K
16 6
6
(Q )
,
1 5
=
Q
4
and hence
Q =
16
4

1 5
199
K
Q
6 1/6
(4.A9)
From (4.A5) and (4.A9), we obtain

6 /(1+6 ) 16
(
+
)
r
10
3
4
1
2
+ 11
16
,
1 5
Q =
4
5
(4.A10)
and from (4.A6)

C + Gc + (1 + 2 ) K + K h + K p + E n + E o
Q
1
=
.
! "13
1 12 q pp
{1 + 14 (1 + 2 )}
(4.A11)
Then, from (4.45), (4.A2), (4.A3), (4.A4), (4.A5), (4.A10)

and (4.A11), we obtain
(r , x ) = 0,
(4.A12)
where
x =
and
q p
q p
=
,
pf
pi
1 e (2 3 )(3 +4 1 2 ) e 2 r
(r , x ) =
g +
T1

(1 + 2 )7 e (8 9 )(3 +4 1 2 ) e 8 r
+
T1
1
1 h(r )
200
1/1+6

r 10 (3 + 4 1 2 )
+
11
+ (1 + 2 )

! "23
(1 + 2 )K p + 22 Y f
(x )24 + E o
+

6 /(1+6 ) 16
r 10 (3 +4 1 2 )+11
16
1 5
4
5
1
.
{112 (x )13 }{1 + 14 (1 + 2 )}
From (4.A6) and (4.A7), we obtain

pi I
12 (x )13
=
,
q p Q
1 12 (x )13
(4.A13)
P
h(r )
=
,
Q
1 h(r )
(4.A14)
from (4.A1),
from (4.47) and (4.A14),
h(r )
21 + 1 + 2
K a

!
"
=
,
Q
1 h(r )
21 26 + 27 r 5
(4.A15)
from (4.46),
F
25 20
=
,
Q +P
20 + 1 + 2
201
(4.A16)
from (4.A1) and (4.A16),

F
1
25 20
,
=
Q
1 h(r )
20 + 1 + 2
(4.A17)
and from (4.49),

pi I
E n E o
F
P
(1 + 2 )K a
+
= 0.
Q Q q p Q
Q
Q
Q
(4.A18)
Then, from (4.45), (4.A10), (4.A13), (4.A14), (4.A15),

(4.A17) and (4.A18), we obtain
(r , x ) = 0
(4.A19)
where

! "23
24
22 Y f
(x )
+ Eo
(r , x ) =
16
4

1 5
r 10 (3 +4 1 2 )+11
5
6 /(1+6 ) 16

12 (x)13
1
25 20
112 (x )13
1 h(r )
20 + 1 + 2

(1 + 2 )(21 + 1 + 2 )
h(r )

!
" .
1
+
1 h(r )
+
r
21
26
27
We, also, need a separate equation for p . This is derived

as follows. From (4.43), we obtain
202
p =

r 18rf 21 (B /M ) 17 19 (3 + 4 1
2 5 )
20

M
,
P + Q
(4.A20)
and from (4.A1) and (4.A20),
{1 h(r )}{r 18 rf 21 (B /M ) 17
19 (3 + 4 1 2 5 )}
p =
20
M
.
(4.A21)
The above equations provide the basis for the computation

of the steady state levels of the 15 endogenous variables.
Appendix B: Transformed Model

In this Appendix, we derive the transformed model (4.51)
(4.65) in the variables y1 , . . . , y15 defined in Section 4.3. For
this purpose, we assume, as in the derivation of the steady
state solution, that the exogenous variables and trends satisfy (4.1), . . . ,(4.17). Equations (4.51), . . . , (4.65) are then
derived as follows.
203
From (3.1),

DlogC = 1 + 2 + 1 log1 2 (r Dlog p)

3 Dlog p + log(Q + P ) logT1 logC ,
and, hence, using the relations

DlogC = D y1 + 1 + 2 ,
C = C e (1 +2 )t e y1 ,
Q = Q e (1 +2 )t e y5 ,
P = P e (1 +2 )t e y12 ,
r = r + y8 ,
Dlog p = D y6 + 3 + 4 1 2 ,
we obtain

D y1 = 1 log1 2 (y8 + r ) + (2 3 )(D y6 + 3 + 4

1 2 )+log(Q e y5 + P e y12 ) logT1 log(C e y1 ) .
(4.B1)
From (4.36),
logC = log1 2r + (2 3 )(3 + 4 1 2 )
+ log(Q + P ) logT1 ,
and, from (4.B1) and (4.B2), we obtain (4.51).
204
(4.B2)
From (3.2) and (4.33)

6
1
2
D y2 = 2 D y2 + 3 log4 1 t
log Q e (1 +2 )t e y5
6

( + )t y 6
2 t y2
1
2
4
log(L e e ) ,
5 K e
e

1
= 2 D y2 + 3 log4
log{(Q e y5 )6
6

y4 6
y2
,
(4.B3)
5 (K e ) } logL e
from (4.37),
log4 = logL +
1
log{(Q )6 5 (K )6 },
6
(4.B4)

From (3.3),
D 2 logK h = 4 (1 + 2 DlogK h ) + 5 {log7 8 (r Dlog p)
9 Dlog p + log(Q + P ) logT1 logK h },

= 4 D y3 + 5 log7 8 (y8 + r )
+ (8 9 )(D y6 + 3 + 4 1 2 )

+ log(Q e y5 + P e y12 ) logT1 log(K h e y3 ) ,
(4.B5)
From (4.38),
logK h = log7 + (8 9 )(3 + 4 1 2 ) 8 r
+ log(Q + P ) logT1 ,
205
(4.B6)
From (3.4),
D y4 = 6 D y4 + 7 log5 + (1 + 6 ) log
2
Q e y5
K e y4

log{y8 + r 10 (D y6 + 3 + 4 1 2 ) + 11 }
= 6 D y4 +7 [log5 +(1+6 )(logQ logK + y5 y4 )
log{y8 + r 10 (D y6 + 3 + 4 1 2 ) + 11 }],
(4.B7)
from (4.39),
log5 + (1 + 6 )(logQ logK )
= log{r 10 (3 + 4 1 2 ) + 11 },
(4.B8)

From (3.5),

y6+ y14 13

q p e
+ log{1+14 (1 +2 )}
D y5 = 8 log 112
pi

+ log C e y1 + Gc + K e y4 (D y4 + 1 + 2 )

+ K h e y3 (D y3 + 1 + 2 ) + K p (1 + 2 ) +E n e y10 + E o

logQ y5 + 9 log14 + log C e y1 + Gc + K e y4

(D y4 + 1 + 2 ) + K h e y3 (D y3 + 1 + 2 )

+ K p (1 + 2 ) + E n e y10 + E o logS y15 ,
206
(4.B9)
from (4.40),

8 log 1 12
q p
pi
13
+ log{1 + 14 (1 + 2 )}

+ log C + Gc + (1 + 2 ) K + K h + K p +E n + E o

logQ + 9 log14 + log {C + Gc + (1 + 2 )

K + K h + K p + E n + E o logS = 0,
(4.B10)

From (3.6),
D y6 = 10 (D y7 D y6 ) + 11 log15 + log4 + logT2

2

+ logw + y7

1 + 6
6

log 1 5
Q e y5
K e y4
6
log p y6 ,
(4.B11)
from (4.41),
log p = log15 + log4 + logT2 + logw

6

Q
1 + 6
, (4.B12)
log 1 5
6
K
207
From (3.7),
D 2y7 = 12 (D y6 D y7 ) 13 (D y6 + D y14 )

1
log{(Q e y5 )6
+ 14 log4 log16
6

y4 6
5 (K e ) } ,
(4.B13)
from (4.42),
log4 log16 =
1
log{(Q )6 5 (K )6 },
6
(4.B14)

From (3.8),

D 2y8 = 15 D y8 + 16 17 + 18 rf
19 (D y14 + 1 + 2 + 5 3 4 )
y6 y5

p e (Q e + P e y12 )
B
+12
y8 r ,
+ 20
M
M
(4.B15)
from (4.43),
r = 17 19 (1 + 2 + 5 3 4 ) + 8rf

p (Q + P )
B
+ 20
+ 21
,
(4.B16)
M
M
208
From (3.9),

q p
D y9 = 17 log12 + (1 + 13 ) log
pi
+ (1 + 13 )(y6 + y14 ) + log{1 + 14 (1 + 2 )}

+ log C e y1 + Gc + K e y4 (D y4 + 1 + 2 )

+ K e y3 (D y3 + 1 + 2 ) + K p (1 + 2 ) + E n e y10 + E o

logI y9 + 18 log14 + log C e y1 + Gc + K e y4
(D y4 + 1 + 2 ) + K h e y3 (D y3 + 1 + 2 )

+ K p (1 + 2 ) + E n e y10 + E o logS y15 , (4.B17)
from (4.44),

q p
+ log{1 + 14 (1 + 2 )}
17 log12 + (1 + 13 ) log
pi

+ log C + Gc + (1 + 2 ) K + K h + K p + E n + E o

logI +18 log14 + log C + Gc + (1 + 2 )

(4.B18)
K + K h + K p + E n + E o logS = 0,
From (3.10),
D y10 = 19 log22 +
23 logY f
q p
24 log
pf

24 (y6 + y14 ) logE n y10 ,
209
(4.B19)
from (4.45),
q p
logE n = log22 + 23 logY f 24 log
pf
+
,
(4.B20)

From (3.11),

Q+P
DlogF = 20 25
1 ,
(4.B21)
F
from (4.B21),

y5

Q e + P e y12
1
,
D y11 = (1 + 2 ) + 20 25
F e y11
(4.B22)
from (4.46),

Q + P
= 20 + 1 + 2 ,
F
20 25
(4.B23)
From (3.12),

!
" K e y13
a
1 ,
Dlog p = 21 26 + 27 rf 5
P e y12
(4.B24)
from (4.B24),
D y12 = (1 + 2 ) + 21

!
" K e y13
a
26 + 27 rf 5
1
,
P e y12
(4.B25)
from (4.47),

!
" K
a
21 26 + 27 rf 5
= 21 + 1 + 2 , (4.B26)
P
210

From (3.13),
D 2 logK a + (DlogK a )2 = 22 DlogK a + 23

Q+P
{28 + 29 (rf r )30 Dlogq 31 dx }
1 ,
Ka
(4.B27)
from (4.B27),
D 2y13 = 22 (D y13 + 1 + 2 ) (D y13 + 1 + 2 )2

!
"
+ 23 28 + 29 rf r y8
30 (D y14 + 1 + 2 + 5 3 4 )}

y5
Q e + P e y12
1 ,
(4.B28)
K a e y13
from (4.48),

!
"
22 (1 + 2 ) + (1 + 2 )2 + 23 = 23 28 + 29 rf r
Q + P
+ 30 (3 + 4 1 2 5 )
,
(4.B29)
K a

From (3.14), (4.13) and (4.14),
D y14 = 24 (D y6 + D y14 ) 25 (y6 + y14 ) + 26 y8

+ 27 log E n e y10 + E o + P e y12 F e y11

pi I
y9 y14 y6
e
,
K a e y13 (D y13 + 1 + 2 ) log
q p
(4.B30)
2
211
and, from (4.B30) and (4.49) we obtain (4.64).

From (3.15),

Q + ( pi /qp)I C DK DKh DK p E n
E o Gc
DlogS =
,
S
(4.B31)
from (4.B31)
D y15 = (1 + 2 )
y5
y y y
y
Q e + ( pi /q p )I e 9 6 14 C e 1
K e y4 (D y4 + 1 + 2 ) K h e y3 (D y3 +1 +2 )
K p (1 + 2 ) E n e y10 E o Gc
,
+
S e y15
(4.B32)
from (4.A6), (4.A7) and (4.A8),

Q + ( pi /q p ) I C Gc (1 + 2 ) K + K h + K p
E n E o = (1 + 2 )S ,
212
(4.B33)
CHAPTER FIVE
Empirical Estimation of the

Model and Derived Results
5.1 Introduction
n this chapter we now turn our attention to the empirical

analysis of the new continuous time macroeconometric
model, using UK data and the exact Gaussian method of

Bergstrom [1997]. Since the model of Bergstrom [1997] is
linear in the variables, although nonlinear in the parameters, while the continuous time macroeconometric model
is non-linear in both the variables and the parameters, it
was necessary, in the estimation, to use a linear (in the variables) approximation to the latter model. This approximation is discussed in Section 5.2, where we also present and
discuss the parameter estimates. In the subsequent sections
we derive and discuss the properties of the estimated model,
213
using the results obtained in earlier chapters. The time lag

distributions with which the variables each depend on the
past values of other variables on which they are directly
dependent have been derived from the estimated parameters using the formulae derived in Appendix B of Chapter 3. They are presented (graphically and through a table
of their means and modes) and discussed in Section 5.3.
The steady state and stability properties of the estimated
model have been derived using the results obtained in
Chapter 4 and are discussed in Section 5.4. Finally, in Section 5.5, we discuss the post-sample forecasting performance of the estimated model and compare it with that of
an unrestricted second-order VARX model (vector autoregressive model with exogenous variables) and Section 5.6
concludes.
5.2 Estimation from United Kingdom Data

As was mentioned earlier, it was necessary to use a linear (in
the variables) approximation to the model specified in Chapter 3, in order to obtain estimates of the parameters by the
exact Gaussian method described in Chapter 2 (Section 2.4).
We have used, for this purpose, a linear approximation about
214
Empirical Estimation of the Model and Derived Results
the sample means of the variables. Alternatively, we could

have used the linear approximation about the steady state
derived in Chapter 4 (Section 4.3 and Appendix B). But, if
the variables are not very close to their steady state paths
during the sample period, better estimates are likely to be
obtained by using the linear approximation about the sample means as is common in most of the literature on continuous time macroeconomic models. It should be emphasized,
however, that nothing can be, rigorously, deduced about
the stability properties of the model from the linear approximation about the sample means. Indeed, it is possible that,
even if all the eigenvalues of the linear approximation about
the sample means have negative real parts, the solution of
the nonlinear model, starting at the sample means, will be
explosive. The stability analysis must, therefore, be based on
the linear approximation about the steady state.
To put the model into the form used in Chapter 2 (equations (2.1)(2.3)), we define a 7 1 vector x1 (t) of endogenous variables that adjust through first-order differential
equations, an 8 1 vector x2 (t) of endogenous variables that
adjust through second-order differential equations, a 15 1
vector z(t) of exogenous variables and a 3 1 vector (t) of
stochastic trends. These vectors are defined as follows.
215
x1 (t) = [logC(t), logE n (t), F (t), logI (t), P(t), logQ (t),
logS(t)] ,
x2 (t) = [logK(t), K a (t), logKh (t), logL(t), log p(t),
logq(t), r (t), logw(t)] ,
z(t) = [logB(t), logK p (t), logM(t), DlogK p (t), dx , logE 0 (t),
logGc (t), log pf (t), log pi (t), r f (t), logT1 (t), logT2 (t),
logY f (t), Dlog pf (t), Dlog pi (t)] ,
(t) = [1 (t), 2 (t), 3 (t)] .
By replacing the expression on the right-hand side of each

of the equations (3.1), (3.2), . . . , (3.15) of Chapter 3 by a
linear approximation (the first-order term of a Taylor series
expansion) about the sample means of the variables, and
adding a white noise innovation to each of these equations
and to each of equations (3.16)(3.18), we obtain the mixedorder linear stochastic differential equation system
d x 1 (t) = [A1 x1 (t) + A2 x2 (t) + A3 D x2 (t) + B1 z(t)
+ C 1 (t) + b1 ] dt + 1 (dt)
(t 0),
(5.1)
d[D x2 (t)] = [A4 x1 (t) + A5 x2 (t) + A6 D x2 (t) + B2 z(t)

+ C 2 (t) + b2 ] dt + 2 (dt)
d(t) = dt + 3 (dt)
(t 0),
(5.2)
(5.3)
where 1 , 2 and 3 are white noise innovation vectors,

A1 , A2 , A3 , A4 , A5 , A6 , B1 , B2 , C 1 and C 2 are matrices whose
216
elements are functions of the sample means of the variables

and the parameter vector = [, ], and b1 and b2 are vectors whose elements are functions of the sample means of
the variables and the parameter vector [, ]. The precise
forms of the functions defining the elements of these matrices and vectors are given in Appendix A to this chapter.
The system (5.1)(5.3) has the same form as the model
(2.1)(2.3) used in Chapter 2. The estimation algorithm
described in that chapter is, therefore, directly applicable
in the estimation of the parameters of (5.1)(5.3), which
are also the parameters of the model (3.1)(3.18) specified
in Chapter 3. The estimates are derived by minimizing a
function L, which is, formally, equal to minus twice the
logarithm of the Gaussian likelihood (less a constant) and
is derived from the system (5.1)(5.3) in the same way as
the function L defined by (2.47) is derived from (2.1)(2.3).
(See Bergstrom [1997], for a complete list of steps in the
computation of the value of L, for a given set of values of
the parameters of the continuous time model, from the discrete observations of the variables.) The estimates obtained
by minimizing L with respect to the parameters of the continuous time model would be exact maximum likelihood
estimates if the system (5.1)(5.3) were the true model and
1 , 2 and 3 were Brownian motion vectors, and they will
217
be good estimates under much more general conditions. As

in Bergstrom, Nowman and Wymer [1992], we placed prior
bounds on all 63 of the structural parameters and minimized
L subject to the constraint that the parameter vector belongs
to a compact rectangular set defined by these bounds.
The parameters were estimated, in the above way, from
UK quarterly data for the period 197594 inclusive, and
the post-sample predictive performance of the model, which
will be discussed in Section 5.4, was tested against quarterly
data for the 2-year period 199596. The data are defined in
Appendix B of this chapter. Before being used in the estimation procedure, they were seasonally corrected by the
method of Durbin [1963].
The parameter estimates are presented in Table 5.1. They
all have plausible values, and only one of them is on a bound.
The mark-up parameter 15 is on its lower bound, which
implies perfect competition. The column headed Standard
Error contains the square roots of the diagonal elements
of minus the inverse of the Hessian of the logarithm of the
Gaussian likelihood function (which is twice the inverse of
the Hessian of L), evaluated at the estimated values of the
parameters. They cannot, however, be interpreted in the
same way as under classical assumptions, since, because
of the stochastic trends, the parameter estimates are not
218
Table 5.1. Parameter Estimates from United Kingdom Quarterly Data

197594
Parameter
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Prior Lower
Bound
Prior Upper
Bound
0.8000
0.0000
0.0000
0.1000
0.0100
0.0100
1.0000
0.0000
0.0000
0.0000
0.0000
0.1000
0.5000
0.0000
1.0000
20000
1.0000
0.5000
0.0000
0.0050
0.0050
10000
0.1000
0.5000
1.0000
4.0000
4.0000
10.000
1.0000
1.0000
10.000
4.0000
4.0000
1.0000
0.0500
0.3000
0.5000
1.0000
1.1000
40000
0.0100
1.0000
1.0000
0.0500
0.0500
35000
1.0000
1.5000
Estimate
0.9351
0.2014
1.6176
0.1494
0.2630
0.2965
3.5000
0.1815
3.0574
0.1535
0.0077
0.2169
0.4033
0.7556
1.0000
23150
0.0832
0.9312
0.1593
0.0073
0.0060
29500
0.5000
1.0000
Standard
Error
0.0073
0.0138
0.0203
0.0298
0.0001
0.0043
0.1401
0.0042
0.0945
0.0365
0.0005
0.0102
0.0598
0.0258
44.2813
0.1087
1.6399
0.0224
0.0007
0.0142
451.7023
0.3562
0.0614
(continued)
219
Table 5.1 (continued)
Parameter
25
26
27
28
29
30
31
32
33
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Prior Lower
Bound
0.0000
0.0000
0.0000
2.0000
0.0000
0.0000
0.0000
0.5000
0.0000
0.0010
0.1000
0.0010
0.1000
0.0010
0.1000
0.0001
0.0010
0.0010
0.1000
0.0010
0.1000
0.0010
0.0010
0.1000
0.0010
0.0010
Prior Upper
Bound
0.0100
0.0200
1.0000
2.0000
100.00
10.000
0.1000
1.5000
10.000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
220
Estimate
0.0083
0.0081
0.2520
0.1303
57.0548
0.1577
0.0007
0.7059
0.0089
0.5294
0.2200
0.0100
4.9999
0.2760
0.0976
0.0009
0.1029
0.0944
0.2384
0.0015
4.3549
0.6494
0.0220
1.4070
0.0292
0.1039
Standard
Error
0.0021
0.0003
0.0315
0.0057
12.9646
0.0305
0.0001
0.1997
0.4913
0.0735
0.1650
0.0145
0.0336
0.0001
0.0040
0.00001
0.0028
0.0015
0.0195
0.000001
0.0749
0.0056
0.0016
0.5086
0.0039
0.0062
Parameter
18
19
20
21
22
23
24
25
26
27
1
2
3
Prior Lower
Bound
Prior Upper
Bound
Estimate
0.0010
0.0010
0.0010
0.0010
0.1000
0.0010
0.1000
0.0010
0.0010
0.0000
0.0000
0.0100
0.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
0.0100
0.0100
0.0500
0.0917
0.1019
2.8565
0.7942
0.2033
0.0075
3.9345
0.1808
0.1244
0.0009
0.0028
0.0061
0.0098
Standard
Error
0.0010
0.0002
0.4912
0.2061
0.0811
0.0014
0.1045
0.0010
0.0047
0.00003
0.0019
0.0025
0.0017
asymptotically normally distributed. They can, however, be

given a simple Bayesian interpretation, as was pointed out in
Chapter 2. Assuming that the innovations in the continuous
time model are Brownian motion and a uniform prior distribution of the parameters, minus the inverse of the Hessian
of the logarithm of the likelihood function, is the covariance matrix of a normal approximation to the posterior
distribution of the parameters in the neighbourhood of its
mode.
221
Turning to the stochastic trend parameters 1 , 2 and 3 ,

the estimate of 1 implies that the amount of labour required
to produce a given output with a given amount of capital decreases, through technical progress, at a mean instantaneous rate of 1.12 per cent per annum, while the estimate of 2 implies that the non-accelerating inflation level
of employment increases at a mean instantaneous rate of
2.44 per cent per annum. The estimate of 3 implies that
the use of cards (plastic money) increases at a rate equivalent (in its effect) to an increase in the money supply at a
mean instantaneous rate of 3.92 per cent per annum.
Among the long-run structural parameters 1 , 2 , . . . , 33 ,
the estimates imply that the long-run marginal propensity
to consume 1 is 0.9351. The parameter 6 determines the
elasticity of substitution between labour and capital, which,
as was mentioned in Chapter 3, equals 1/(1 + 6 ). Substituting the estimated value of 6 into this expression, we
obtain an estimate 0.7713 for the elasticity of substitution
between labour and capital. Another important elasticity is
the price elasticity of demand for imports, which, as was
mentioned in Chapter 3, equals (1 + 13 ). Substituting
the estimated value of 13 into this expression, we obtain
an estimate 0.5967 for the price elasticity of demand for
imports. The direct interpretation of the estimates of the
222
speed of adjustment parameters 1 , 2 . . . , 27 is fairly obvious. Less obvious and, perhaps, of more interest are the
time lag distributions with which the variables adjust, each
in response to changes in the other variables on which
it is directly dependent. These time lag distributions are
presented and discussed in Section 5.3, to which we now
turn.
5.3 Time Lag Distributions

Nearly all of the equations of the model are in the form of
one of the continuous adjustment equations (3.A15) and
(3.A37) derived in Appendix A of Chapter 3. It is shown
in Appendix B of Chapter 3 that each of equations (3.A15)
and (3.A37) implies that the dependent variable depends,
with a distributed time lag, on all past values of a subset of
the other variables in the model. More precisely, it depends,
with a distributed time lag, on all past values of its partial
equilibrium level, which is a function of these other variables. Appendix B of Chapter 3 includes the derivation of the
forms of the time lag distributions in these relations and their
means and modes (also presented in Bergstrom, Nowman
and Wymer [1992] for the second-order equations). When
the adjustment equation is a first-order differential equation
223
of the form (3.A15), the time lag distribution is an exponential distribution given by (3.B4). The density is a decreasing
function of the time lag, and the mean of the distribution is
the reciprocal of the speed of adjustment parameter .
When the adjustment equation is a second-order differential equation of the form (3.A37), the time lag distribution
is of a more complicated form, depending on the roots of
the quadratic equation (3.B7). In the normal case, where
these roots are real and negative, it will be a unimodal distribution with a positive mode. But, its exact form depends
on whether the roots are equal or unequal. When they are
unequal, the density function of the distribution is given
by (3.B10), its mean by (3.B17) and its mode by (3.B18),
while when they are equal, the density function is given
by (3.B24), the mean is 2/ and the mode is 1/. Estimates
of the means and modes of the lag distributions have been
derived from the estimated values of the parameters, for all
variables in the model that adjust through differential equations of the form (3.A15) or (3.A37) and are presented in
Table 5.2.
As can be seen, the private non-residential fixed capital
stock variable adjusts with the longest time lag and the financial flow variables with the shortest time lags. In particular,
the private non-residential fixed capital variable logK
224
Table 5.2. Estimated Time Lag Parameters
Variable
logC
logEn
F
logI
logKh
logK
Ka
logL
P
logQ
r
Mean Time Lag

(Quarters)
1.8889
9.8135
0.3501
9.6246
18.1156
108.4444
27.1067
22.0000
1.2591
9.7182
48.1849
Modal Time Lag

(Quarters)
0.9172
27.7487
10.9215
9.6795
3.0672
adjusts with a mean time lag of approximately 27 years and

a modal time lag of approximately 7 years; interest rate variable r adjusts with a mean time lag of 12 years and a modal
time lag of under 1 year; cumulative net real investment
abroad variable K a adjusts with a mean time lag of approximately 7 years and a modal time lag of approximately
3 years; employment variable logL adjusts with a mean time
lag of 51/2 years and a modal time lag of approximately 21/2
years and the residential fixed capital variable logKh adjusts
with a mean time lag of 41/2 years and a modal time lag of
approximately 3 months.
225
0.6
Density
0.5
0.4
0.3
0.2
0.1
0
0
10
Quarters
Figure 5.1 Lag Distribution Private Consumption: log C
Turning to the first-order equations, we find that the nonoil exports variable logE n , output variable logQ and the
import variable logI adjust with mean time lags of approximately 21/2 years; the consumption variable logC and profits,
interest and dividends from abroad variable P adjust with
mean time lags of approximately 1/2 a year, and finally the
international transfers variable F adjusts with a mean time
lag of approximately 1 month. The graphs of the density
functions of the lag distributions are presented in Figures 5.1
to 5.11.
226
0.04
0.035
Density
0.03
0.025
0.02
0.015
0.01
0.005
0
0
10
20
30
40
Quarters
Figure 5.2 Lag Distribution Employment: log L

0.06
Density
0.05
0.04
0.03
0.02
0.01
0
0
10
15
Quarters
Figure 5.3 Lag Distribution Residential Fixed Capital: log Kh
227
20
0.009
0.008
0.007
Density
0.006
0.005
0.004
0.003
0.002
0.001
0
0
20
40
60
80
100
120
140
Quarters
Figure 5.4 Lag Distribution Private Non-Residential Fixed Capital:

log K
0.12
0.1
Density
0.08
0.06
0.04
0.02
0
0
10
15
Quarters
Figure 5.5 Lag Distribution Output: log Q
228
20
0.025
Density
0.02
0.015
0.01
0.005
0
0
10
20
30
40
Quarters
Figure 5.6 Lag Distribution Interest Rate: r

0.12
0.1
Density
0.08
0.06
0.04
0.02
0
0
10
15
Quarters
Figure 5.7 Lag Distribution Imports: log I
229
20
0.12
0.1
Density
0.08
0.06
0.04
0.02
0
0
10
15
20
Quarters
Figure 5.8 Lag Distribution Non-Oil Exports: log En

3
Density
2.5
2
1.5
1
0.5
0
0
0.2
0.4
0.6
0.8
1.2
1.4
1.6
Quarters
Figure 5.9 Lag Distribution Current Transfers Abroad: F
230
1.8
0.9
0.8
0.7
Density
0.6
0.5
0.4
0.3
0.2
0.1
0
0
Quarters
Figure 5.10 Lag Distribution Profits Interest and Dividends From

Abroad: P
0.007
0.006
Density
0.005
0.004
0.003
0.002
0.001
0
0
10
20
30
40
50
60
Quarters
Figure 5.11 Lag Distribution Investment Abroad: Ka
231
70
5.4 Steady State and Stability Properties

In this section we discuss the steady state and stability properties of the estimated model, which have been derived
using the results obtained in Chapter 4. For the purpose
of this derivation, we have adjusted 4 parameters from their
estimated values to values that seem more realistic for the
future, bearing in mind that the purpose of the steady state
analysis is to provide a check on the capacity of the model
to generate plausible long-run behaviour. The parameters
adjusted are 13 , which determines the price elasticity of
demand for imports, 14 the optimal (partial equilibrium)
ratio of stocks to sales, 2 the mean rate of growth of the
labour supply and 3 the mean rate of growth of the use of
plastic money. The parameter 13 has been set at the value
0.2, which implies a price elasticity of demand for imports
of 1.2. In view of the greater liberalization of trade, which
is now taking place, this seems more realistic for the future
than the price elasticity of 0.60 implied by the estimated
value of 13 .
In view of the growing relative importance of the service
sector of the economy, we should expect the value of the
parameter 14 to be lower in the future than it was over
the sample period. Moreover, the fitting of a trend to the
232
actual ratio of stocks to sales S/(C + Gc + DK + DKh + DK p +

E n + E o ) provides strong evidence that there was a significant fall in 14 over the sample period. Since the adjusted
value of 14 is used in the generation of post-sample forecasts, to be discussed in Section 5.5, as well as in the
steady state derivation, it has been chosen in an objective
way, which makes use of no data beyond the end of the
sample period. We have fitted a quadratic trend to the ratio
S/(C + Gc + DK + DKh + DK p + E n + E o ), using data up to
the last quarter of the sample period, and set 14 equal to
the trend value in the last quarter of the forecast period,
i.e. 2 years beyond the end of the sample period. The
value of 14 obtained in this way is 0.5804, which is significantly lower than the estimated value 0.7556 shown in
Table 5.1.
The parameter 2 has been set at the value 0.002. This
implies a rate of growth of the labour supply of 0.8 per cent
per annum, which seems more realistic for the future than
the growth rate of 2.4 per cent per annum implied by the
estimated value of 2 . The parameter 3 has, for simplicity,
been set a value 0.0 and the assumed rate of growth of the
money supply adjusted accordingly.
To obtain the steady state solution, we must, also, assume
values of the parameters in (4.1)(4.17) of Chapter 4, which
233
determine the assumed paths of the exogenous variables.

The parameter 4 , representing the rate of growth of the
money supply, has been set at the value 0.0048, which,
together with the estimated value 0.0028 of 1 and the
assumed value 0.0020 of 2 , implies a steady state rate of
increase in UK prices of zero. The parameter 5 , representing the rate of increase in foreign prices, has been set
at the value 0.0, which, in conjunction with the zero steady
state rate of increase in UK prices, implies a constant
steady state exchange rate. The level parameters B , E 0 ,
Kp , M , pf , pi and Y f have been set at values that ensure
that the levels of the steady state paths of the corresponding variables are equal to the actual levels of the variables
in the base period. The parameters T1 , T2 and g are set at
the average values of T1 , T2 and Gc /(Q + P ) over the sample period. The numerical values of the parameters of the
assumed steady state paths of the exogenous variables are
shown in Table 5.3.
The steady state growth rates, under the above assumptions, are shown in Table 5.4 and the steady state level
parameters in Table 5.5. The actual and steady state levels
of the endogenous variables in the last quarters of 1996 are,
also, shown in Table 5.5.
234
Table 5.3. Values of Exogenous Variable

Time Path Parameters Assumed in Steady
State Analysis
Parameter
Assumed Value
B
E 0
g
Kp
M
pf
pi
r f
T1
T2
Y f
4
5
178786
1580
0.2400
172434
17143
1.0000
1.0000
0.0098
1.3103
1.1526
0.5514
0.0048
0.0000
Table 5.4. Derived Steady State Growth Rates

Instantaneous Growth Rate
(per cent per annum)
Variable
C , En , F , I , K , Kh , K a , P , Q, S
L
p
w
r
q
1.92
0.80
0.00
1.12
0.00
0.00
235
Table 5.5. Derived Steady State Levels
Variable
C
En
F
I
Kh
K
Ka
L
P
p
Q
q
r
S
w
Steady State
Steady State Level in Actual Level in Last
Level Parameter Last Quarter of 1996 Quarter of 1996
65,787
26,456
765
20,218
246,283
741,387
11,365
23,151
120
0.5457
92,488
1.5174
0.0100
70,553
1.6230
100,365
40,362
1,168
30,844
375,733
1,131,074
17,339
27,606
182
0.5457
141,102
1.5174
0.0100
107,637
2.0764
100,619
42,999
1,096
47,103
579,636
713,621
18,154
26,244
1,570
1.2449
142,808
0.9139
0.0184
116,833
4.5106
The values of the parameters 2 and 4 assumed for the

purpose of the steady state analysis are reflected in the large
differences between the actual and steady state levels of
the variables K and p in the last quarter of 1996. The large
excess of the steady state level of the capital variable K over
the actual level is a result of the downward adjustment of
the parameter 2 (the rate of growth of the labour supply)
from the estimated value 0.0061 to the value 0.002. This
236
increases the steady state ratio of capital to output, because

the proportion of output that must be invested to maintain
a given ratio of capital to output is smaller the lower is the
growth rate.
The low steady state price level in the last quarter of 1996,
compared with the actual price level in that quarter, is a
consequence of setting 4 (the rate of increase in the money
supply) at a level that keeps the steady state price level constant. This reduces the real cost of holding wealth in the
form of money and, hence, increases the demand for and
steady level of real money balances. But, since M has been
set at the level that makes the steady state level of M equal to
the actual level of M in the last quarter of 1996, an increase
in the steady state level of real money balances implies a
decrease in the steady state price level. We conclude this section with a discussion of the stability and other properties of
the solution of the estimated model in the neighbourhood
of the steady state. Here, we rely on the results obtained in
Chapter 4, Section 4.3 and Appendix B.
It is shown in Chapter 4 that the solution of the model
is asymptotically stable if the eigenvalues of the matrix C
in (4.82) all have negative real parts. If this condition is
satisfied, then, any solution of the system (3.1)(3.18) in
which the initial values of the variables are sufficiently close
237
to their steady state paths will converge to the steady state

solution if undisturbed. Formulae expressing the elements
of the matrix C as functions of the parameters of the model
(3.1)(3.18) are easily implied from Chapter 4. These elements have been evaluated at the estimated values of the
parameters, after adjusting the parameter value of 13 to the
value 0.2 and 14 to the value 0.5804, as in the computation
of the steady state level parameters. (The other parameters
adjusted for the computation of the steady state levels have
no effect on the asymptotic stability of the solution.) The
eigenvalues of the estimated matrix C obtained in this way
are presented in Table 5.6.
All except 5 of these eigenvalues have negative real parts.
The eigenvalues with positive real parts include one real
eigenvalue 0.0066 and two complex pairs 0.0012 0.0202i
and 0.1695 0.0389i. The long-term solution of the system
(4.82) will be dominated by the cycle associated with the
complex pair 0.1695 0.0389i, whose real part is greater
than that of any of the other eigenvalues in Table 6. If unconstrained, this would be an explosive cycle with a period of,
approximately, 40 years. A realistic model must, however,
take account of a full employment ceiling or some similar
constraint on the productive capacity of the economy. Taking
account of such a constraint, the model, in its deterministic
238
Table 5.6. Eigenvalues for Linear

Approximation about Steady State
Real Part
Imaginary Part
4.9633
2.8613
1.4068
0.7990
0.5419
0.1597
0.1558
0.1033
0.1016
0.0642
0.0533
0.0469
0.0295
0.0017
0.0066
4.2409
0.0699
0.0012
0.1695
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.1792
0.0094
0.0202
0.0389
form, can be expected to generate a perpetual long cycle

with a period of 3540 years and, approximately, constant
amplitude. This could be the well-known Kondratieff cycle
observed and discussed by Kondratieff [1926, 1935].
239
The model also generates a cycle with a period of approximately 9 years associated with the pair of eigenvalues
4.2409 0.1792. This could be the well-known economic
cycle commonly referred to as the trade cycle or business
cycle and generally recognized as having a period of 711
years. Although the pair of eigenvalues with which this cycle
is associated has a negative real part, the cycle will be prevented from dying out by the stochastic innovations in the
model, and its average amplitude will depend on the variances of these innovations. We may conclude, therefore,
that the model (3.1)(3.18) generates plausible long-run
behaviour, with fluctuations about the steady state neoclassical growth path including both a trade cycle with a period
of about 9 years and a longer (Kondratieff) cycle with a
period of about 40 years.
5.5 Post-Sample Forecasting Performance

To assess the post-sample forecasting of the model, we
have computed the optimal post-sample forecasts of the
discrete observations of the endogenous variables for the
8 post-sample quarters 1995Q11996Q4, using the method
described in Chapter 2, Section 2.4. The method is a recursive procedure, in which the computation of the forecast for
240
each post-sample quarter makes use of the forecasts for the

proceeding post-sample quarters. It should be emphasized,
however, that the procedure makes no use, either directly
or indirectly, of the actual value of any endogenous variable after 1994 quarter 4, the last quarter of the sample
period. For the computation of the post-sample forecasts, all
parameters of the model except 14 were set equal to their
estimated values. For reasons discussed in the preceding section, 14 the optimal (partial equilibrium) ratio of the stocks
to sales has been adjusted from its estimated value 0.7556 to
the value 0.5804, which was obtained, as explained in the
preceding section, by fitting a quadratic trend to the ratio of
stocks to sales over the sample period. The latter procedure
makes no use of any data relating to quarters after 1994
quarter 4, and it could, therefore, have been used in actual
forecasting at the end of the sample period.
The root-mean-square errors (RMSE) of the forecasts
from the continuous time model for the 8 post-sample quarters 1995Q11996Q4 are compared, in Table 5.7, with the
RMSE of the corresponding forecasts from a second-order
VARX model (vector autoregressive model with exogenous
variables). The comparison is confined to the 12 endogenous variables that can assume positive values only (i.e. all
except F, P and K a ). These variables can be represented in
241
Table 5.7. Root Mean Square Errors of Post-Sample MultiPeriod Forecasts 1995Q11996Q4
Variable
logC
logEn
logI
logKh
logK
logL
log p
logQ
logq
r
logS
logw
RMSE of Forecasts from

Continuous Time Model
RMSE of Forecasts
from VARX Model
0.0477
0.0797
0.0450
0.0350
0.0024
0.0137
0.0121
0.0395
0.0472
0.0012
0.1396
0.0130
0.0148
0.1131
0.1329
0.0024
0.0077
0.0057
0.0218
0.0181
0.0453
0.0007
0.0096
0.0227
logarithmic form and the forecast errors (when multiplied

by 100) interpreted as percentage errors. The forecasts, from
the continuous time model, of the sequence of values of
each of these variables for the quarters 1995Q11996Q4 are
shown in Table 5.8 together with the actual values of the
variables in the same quarters.
Comparing the RMSE of the forecasts from the continuous time and VARX models shown in Table 5.7, we
see that for the interest rate variable r the RMSE of the
242
Table 5.8. Post-Sample Multi-Period Forecasts 1995Q1

1996Q4
Variable
Quarter
Actual Value
Forecast
logC
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
11.4364
11.4306
11.4545
11.4424
11.4579
11.4630
11.4862
11.4761
11.4436
11.4609
11.4739
11.4881
11.4990
11.5165
11.5424
11.5599
logEn
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
10.5943
10.5545
10.6041
10.6157
10.6551
10.6283
10.6432
10.6472
10.5315
10.5325
10.5348
10.5379
10.5418
10.5466
10.5523
10.5586
l ogI
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
10.6232
10.6450
10.6737
10.6913
10.7258
10.7203
10.7384
10.7746
10.6627
10.6566
10.6562
10.6598
10.6671
10.6761
10.6888
10.7010
(continued)
243

Variable
Quarter
Actual Value
Forecast
logKh
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
13.2508
13.2540
13.2564
13.2593
13.2616
13.2644
13.2667
13.2700
13.2566
13.2659
13.2753
13.2849
13.2941
13.3053
13.3166
13.3280
logK
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
13.4279
13.4345
13.4415
13.4476
13.4558
13.4627
13.4697
13.4771
13.4273
13.4337
13.4407
13.4481
13.4561
13.4644
13.4732
13.4824
logL
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
10.1576
10.1594
10.1558
10.1639
10.1593
10.1611
10.1680
10.1721
10.1629
10.1675
10.1713
10.1744
10.1772
10.1798
10.1824
10.1852
244

Variable
Quarter
Actual Value
Forecast
log p
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
0.1714
0.1817
0.1810
0.1870
0.2023
0.2058
0.2109
0.2199
0.1650
0.1710
0.1766
0.1819
0.1868
0.1917
0.1966
0.2017
logQ
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
11.7785
11.7749
11.7855
11.7857
11.7927
11.8043
11.8117
11.8270
11.7669
11.7767
11.7907
11.8079
11.8275
11.8486
11.8730
11.8967
logq
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
0.1332
0.1735
0.1769
0.1759
0.1761
0.1678
0.1624
0.0854
0.1064
0.1172
0.1284
0.1391
0.1497
0.1613
0.1722
0.1824
(continued)
245

Variable
Quarter
Actual Value
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
0.0207
0.0195
0.0195
0.0185
0.0203
0.0199
0.0196
0.0184
0.0192
0.0189
0.0187
0.0186
0.0184
0.0183
0.0182
0.0181
logS
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
11.6071
11.6243
11.6358
11.6506
11.6474
11.6565
11.6566
11.6709
11.6085
11.5754
11.5404
11.5114
11.4908
11.4813
11.4712
11.4822
logw
1995 Q1
Q2
Q3
Q4
1996 Q1
Q2
Q3
Q4
1.4552
1.4489
1.4535
1.4613
1.4983
1.4884
1.4856
1.5049
1.4382
1.4464
1.4542
1.4615
1.4686
1.4776
1.4862
1.4968
246
Forecast
forecasts from the two models differ by only 0.0005. We

shall treat the forecasts of this variable from the two models as equally good, therefore. With regard to the remaining 11 variables in Table 5.7, the continuous time model
yields forecasts with the lower RMSE for the 5 variables
logE n , logI, logK , log p and logw, while the VARX model
yields forecasts with the lower RMSE for the six variables,
logC , logKh , logL , logQ , logq and logS.
The forecasting performance of the VARX model is a little better, therefore, than that of the continuous time model.
But, such a small difference could, easily, be a random property of the stochastic innovations in the particular 8 quarters that we have chosen to forecast, even if the continuous
time model is correctly specified. We may conclude, therefore, that the forecasting performance of the continuous
time model is satisfactory. It should be emphasized, moreover, that our principal aim, in producing this model, was to
produce a prototype mixed-order continuous time macroeconometric model with stochastic trends and we have
done no experimentation with alternative specifications. For
the purpose of practical forecasting, we could use a somewhat less parsimonious specification with a more refined
specification of some of the partial equilibrium levels (for
example, distinguishing between the effects on the partial
247
equilibrium consumption level of income from employment

and other income) and, perhaps, replacing some of the firstorder adjustment equations with second-order adjustment
equations.
5.6 Conclusion
In this final chapter, we have presented and discussed the
estimates of the parameters of the model in Chapter 3, which
were obtained from UK data by the exact Gaussian estimation procedure described in Chapter 2. Since the latter procedure is applicable only to a model that is linear in the
variables, although, generally, nonlinear in the parameters,
it was necessary, for the purpose of estimation, to use a linear (in the variables) approximation to the model specified
in Chapter 3. The parameter estimates are, on the whole,
very plausible, and all of the time lag distributions derived
from the speed of adjustment parameters are very plausible.
We have examined the steady state and stability properties of
the estimated model, using the results obtained in Chapter 4,
and shown that it generates plausible long-run behaviour.
More specifically, it generates fluctuations about the steady
state neoclassical growth path, including both a trade cycle
with a period of about 9 years and a longer (Kondratieff)
248
cycle with a period of about 40 years. Finally, we have examined the post-sample forecasting performance of the estimated model and compared it with that of a second-order
VARX model (vector autoregressive model with exogenous
variables). We conclude that the forecasting performance is
satisfactory, particularly considering the very parsimonious
specification of the model.
Appendix A: Linear Approximation

about Sample Means
In this Appendix we give the formulae for the non-zero elements of the matrices A1 , A2 , . . . , A6 , B1 , B2 , C 1 and C 2 and
the vectors b1 and b2 in the system (5.1)(5.3). Letting [Ah ]i j
denote the element in row i and column j of the matrix
Ah (h = 1, . . . , 6), using a similar notation for the elements
of the matrices Bh and C h (h = 1, 2), and letting [bh ]i denote
the ith element of the vector bh (h = 1, 2), the formulae are
as follows, the sample mean of any variable x being denoted
by x.
[A1 ]11 = 1 ,

1
[A1 ]15 = 1
,
e logQ + P
249
e logQ
[A1 ]16 = 1
[A1 ]22
e logQ + P
= 19 ,

,
[A1 ]33 = 20 ,
[A1 ]35 = 20 25 ,
[A1 ]36 = 20 25 e logQ ,
[A1 ]41 =
(17 + 18 )e logC
e logC + e logGc + e logK DlogK + e logK h DlogK h
,
+ e logK p DlogK p + e logEn + e logEo

[A1 ]42 =
(17 + 18 )e logEn
,
[A1 ]44 = 17 ,
[A1 ]47 = 18 ,
[A1 ]55 = 21 ,
[A1 ]61 =
(8 + 9 )e logC
,

[A1 ]62 =
[A1 ]66
(8 + 9 )e logEn

= 8 ,
[A1 ]67 = 9 ,
250
,
[A1 ]71 = e [logC logS] ,

[A1 ]74 = e [log pi +logI logqlog plogS] ,
[A1 ]76 = e [logQ logS] ,
[A1 ]77 = e [logQ logS] e [log pi +logI logqlog plogS]
+ e [logC logS] + e [logEn logS] + e [logEo logS]
+ e [logGc logS] + DlogK e [logK logS]
+ DlogK h e [logK h logS] + DlogK p e [logK p logS] ,
[A2 ]17 = 1 2 ,
[A2 ]25 = 19 24 ,
[A2 ]26 = 19 24 ,
[A2 ]41 =
[A2 ]43 =
(17 + 18 )e logK DlogK

(17 + 18 )e logK h DlogK h
,
,
[A2 ]45 = 17 (1 + 13 ),
[A2 ]46 = 17 (1 + 13 ),
[A2 ]52 = 21 {26 + 27 (r f Dlog pf )},
[A2 ]61 =
(8 + 9 )e logK DlogK
251
,
(8 + 9 )e logK h DlogK h
[A2 ]63 =
[A2 ]65
[A2 ]66


8 12 13 e 13 (logq+log plog pi )
,
=
1 12 e 13 (logq+log plog pi )

=
,
,
[A2 ]71 = DlogK e (logK logS) ,

[A2 ]73 = DlogK h e (logK h logS) ,
[A2 ]75 = e (log pi +logI logqlog plogS) ,
[A2 ]76 = e (log pi +logI logqlog plogS) ,
[A3 ]15 = 1 (2 3 ),
[A3 ]41 =
(17 + 18 )e logK
,

[A3 ]43 =
(17 + 18 )e logK h
,

[A3 ]61 =
(8 + 9 )e logK
252
,
[A3 ]63 =
(8 + 9 )e logK h
,
+ e logK p DlogK p + e logE n + e logEo

[A3 ]71 = e {logK logS} ,
[A3 ]73 = e {logK h logS} ,
[A4 ]16 = 7 (1 + 6 ),

[A4 ]25 = 23 28 + 29 (r f r ) 30 Dlogq 31 dx ,

[A4 ]26 = 23 28 + 29 (r f r ) 30 Dlogq 31 dx e logQ ,

1
,
[A4 ]35 = 5
e logQ + P

e logQ
[A4 ]36 = 5
,
e logQ + P
[A4 ]46 =
[A4 ]56 =
[A4 ]62 =
[A4 ]63
[A4 ]64
[A4 ]65
3 e 6 logQ
e 6 logQ 5 e 6 logK
11 5 (1 + 6 )e 6 (logQ logK )
1 5 e 6 (logQ logK )
27 e
logE n
,
e logEn + e logEo + P F DKa
27
=
,
logE
logE
n + e
o + P F DK
e
a
= 27 ,
27
=
,
logE
logE
n
o
e
+e
+ P F DKa
253
[A4 ]75 = 16 20 e (log plogM3 ) ,
where 3 =
3 T
2
[A4 ]76 = 16 20 e (log p+logQ logM3 ) ,

[A4 ]86 =
[A5 ]11 =
[A5 ]17 =
[A5 ]22 =
[A5 ]27 =
14 e 6 logQ
,
e 6 logQ 5 e 6 logK
7 (1 + 6 ),

7
,
r 10 Dlog p + 11
23 ,
!
"
23 29 e logQ + P ,
[A5 ]33 = 5 ,
[A5 ]37 = 5 8 ,

[A5 ]41 =
[A5 ]44
[A5 ]51
[A5 ]55
3 5 e 6 logK
,
e 6 logQ 5 e 6 logK
= 3 ,

11 5 (1 + 6 )e 6 (logQ logK )
,
=
= 11 ,
[A5 ]58 = 11 ,
[A5 ]65 = 27 25 ,
[A5 ]66 = 27 25 ,
[A5 ]67 = 26 ,

[A5 ]75 = 16 20 e (log p+logQ logM3 ) + P e (log plogM3 ) ,
254
[A5 ]77 = 16 ,

[A5 ]81 =
14 5 e 6 logK
e 6 logQ 5 e 6 logK

,
[A6 ]11 = 6 ,
[A6 ]15 =
[A6 ]22
[A6 ]26
7 10
,
r 10 Dlog p + 11
= 22 ,
!
"
= 23 30 e logQ + P ,
[A6 ]33 = 4 ,
[A6 ]35 = 5 (8 9 ),
[A6 ]44 = 2 ,
[A6 ]55 = 10 ,
[A6 ]58 = 10 ,

[A6 ]62 =
[A6 ]65
27
e logEn + e logE o + P F DKa

= 24 ,
[A6 ]66 = 24 ,
[A6 ]76 = 16 19 ,
[A6 ]77 = 15 ,
[A6 ]85 = 12 13 ,
[A6 ]86 = 13 ,
[A6 ]88 = 12 ,
[B1 ]1,11 = 1 ,
[B1 ]1,28 = 19 24 ,
255
[B1 ]2,13 = 19 23 ,
[B1 ]42 =
(17 + 18 )e logK p DlogK p

,

[B1 ]44 =
(17 + 18 )e logK p
,

[B1 ]46 =
(17 + 18 )e logE o
,

[B1 ]47 =
[B1 ]49
(17 + 18 )e logGc
,

= 17 (1 + 12 ),
[B1 ]5,10 = 21 27 K a ,
[B1 ]5,14 = 21 27 K a ,
[B1 ]62 =
(8 + 9 )e logK p DlogK p
,

[B1 ]64 =
(8 + 9 )e logK p
256
,
[B1 ]66 =
(8 + 9 )e logEo
,

[B1 ]67 =
(8 + 9 )e logGc
[B1 ]69 =
[B1 ]72 = DlogK p e (logK p logS) ,

[B1 ]74 = e (logK p logS) ,
[B1 ]76 = e (logEo logS) ,
[B1 ]77 = e (logGc logS) ,
[B1 ]79 = e (log pi +logI logqlog plogS) ,

[B2 ]25 = 23 31 e logQ + P ,

[B2 ]2,10 = 23 29 e logQ + P ,
[B2 ]3,11 = 5 ,
[B2 ]5,12 = 11 ,
[B2 ]66 =
[B2 ]68
27 e logEo
e logEn + e logE o + P F DKa

= 25 ,
[B2 ]69 = 27 ,
[B2 ]6,14 = 24 ,
257
,

[B2 ]71 = 16 21 e (logBlogM) ,

[B2 ]73 = 16 20 e (log p+logQ logM3 ) + 20 P e (log plogM3 )

+ 21 e (logBlogM) ,
where 3 = 32T ,
[B2 ]7,10 = 16 18 ,
[B2 ]8,15 = 13 ,
[C 2 ]41 = 3 ,
[C 2 ]51 = 11 ,

[C 2 ]73 = 16 20 e (log p+logQ logM3 ) + P e (log plogM3 ) ,
[C 2 ]81 = 14 ,
[C 2 ]82 = 14 ,
!
"
[b1 ]1 = 1 + 2 + 1 log1 + 1 log e logQ + P

logQ e logQ
P
1
,
1
e logQ + P
e logQ + P
[b1 ]2 = 1 + 2 + 19 log22 ,
[b1 ]3 = 20 25 e logQ {1 logQ },
[b1 ]4 = 1 + 2 + 17 log12 + 18 log14
+ 17 log{1 + 14 (1 + 2 )} + (17 + 18 )

log e logC +e logGc +e logK DlogK +e logK h DlogK h

+ e logK p DlogK p +e logEn +e logEo (17 + 18 )
258
e logC logC +e logGc logGc +e logK DlogK logK
+ e logK h DlogK logK + e logK p DlogK logK

h
+ e logK DlogK +e logK h DlogK h +e logK p DlogK p
+ e logEn logE +e logEo logE

n
o

,
e logC +e logGc + e logK DlogK +e logK h DlogK h
+ e logK p DlogK p +e logEn +e logE o
[b1 ]5 = 21 27 (r f Dlog pf )K a ,
[b1 ]6 = 1 + 2 + 8 log{1 + 14 (1 + 2 )} + 9 log14

+ 8 log 1 12 e 13 (logq+log plog pi )

+
(logq + log p log pi ) + (8 +9 )

log e logC +e logGc + e logK DlogK +e logK h DlogK h

+ e logK p DlogK p +e logEn +e logEo (8 +9 )
e logC logC +e logGc logGc +e logK DlogK logK
+ e logK h DlogK h logK h +e logK p DlogK p logK p
+ e logK DlogK +e logK h DlogK h +e logK p DlogK p
logE n
logE o
+e
logE n +e
logE o

,
e logC +e logGc +e logK DlogK +e logK h DlogK h
+ e logK p DlogK p + e logEn + e logE o
259
[b1 ]7 = e {logQ logS} {1 logQ + logS}

+ e {log pi +logI logqlog plogS}
{1 log pi logI + logq + log p + logS}
e {logC logS} {1 logC + logS}
e {logEn logS} {1 logE n + logS}
e {logEn logS} {1 logE o + logS}
e {logGc logS} {1 logGc + logS}
+ DlogK e {logK logS} {logK logS}
+ DlogK h e {logK h logS} {logK h logS}
+ DlogK p e {logK p logS} {logK p logS},
[b2 ]1 = 6 (1 + 2 )+7 log5 7 log(r 10 Dlog p + 11 )

r 10 Dlog p
+ 7
,
r 10 Dlog p + 11
[b2 ]2 = 23 28 e logQ (1 logQ )
23 {29 (r f r ) 30 Dlogq 31 dx }

e logQ logQ + P ,
!
"
[b2 ]3 = 4 (1 + 2 ) + 5 log7 + 5 log e logQ + P

logQ e logQ
P
5
,
5
e logQ + P
e logQ + P
260
[b2 ]4 = 2 2 + 3 log4

+

3
log e 6 logQ 5 e 6 logK
6

3 e 6 logQ
e 6 logQ 5 e 6 logK
3 5 e 6 logK
logQ

e 6 logQ 5 e 6 logK
logK ,
[b2 ]5 = 11 (log4 + log15 ) 10 1

11 (1 + 6 )
log 1 5 e 6 (logQ logK )

6

11 5 (1 + 6 ) e 6 (logQ logK )
(logQ logK ),
[b2 ]6 = 25 log32 26 33

+ 27 log e logEn + e logE o + P F DKa

27 e logEn logE n + e logE o logE 0 + P F DKa

,
e logEn + e logEo + P F DKa

[b2 ]7 = 16 17 + 16 20 e (log p+logQ logM3 )
{1 log p logQ + logM + 3 }
!
"
log plogM3
+ 16 20 P e
{logM log p + 3 }

+ 16 21 e (logBlogM) {1 + logM logB},
261
[b2 ]8 = 12 1 + 14 (log4 log16 )

14
log e 6 logQ 5 e 6 logK

6

14 e 6 logQ
logQ
e 6 logQ 5 e 6 logK

14 5 e 6 logK
logK .
+
e 6 logQ 5 e 6 logK
Appendix B: Data
Sources of data:
BLUE United Kingdom National Accounts: The Blue
Book, 1997, ONS
BOE
Bank of England
ET
Economic Trends
ETAS
Economic Trends Annual Supplement, 1997, ONS
FS
Financial Statistics
IFS
International Financial Statistics
MD
Monthly Digest of Statistics
UKBP United Kingdom Balance of Payments

Definitions of series
262
The series used in this study consisted of quarterly observation for the period 197496 and were defined as follows:
(C ) Real Private Consumption
Consumers expenditure at current prices deflated by gross
domestic product implicit price deflator, p, as defined below
(source: ETAS).
(E n ) Real Non-Oil Exports
Exports of goods and services less oil exports at current prices
deflated by p (sources: ETAS and ET).
(F ) Real Current Transfers Abroad
Net government and private transfers abroad deflated by p
(source: ETAS).
(I ) Volume of Imports
Imports of goods and services at 1990 market prices (source:
ETAS).
(Kh ) Residential Fixed Capital
Cumulative net real residential fixed capital formulation
since the beginning of 1975 plus a constant derived from
the official estimate of the capital stock for dwellings at the
end of 1996 adjusted to 1990 prices. Net real residential fixed
capital formulation equals gross residential fixed capital formulation at current prices deflated by p minus depreciation
at current replacement for residential dwellings deflated by
p (sources: ETAS and BLUE).
263
(K ) Private Non-Residential Fixed Capital

Cumulative net real private non-residential fixed capital formulation since the beginning of 1975 plus a constant derived
from the official estimate of the capital stock for private nonresidential fixed capital at the end of 1996 adjusted to 1990
prices. Net real private non-residential fixed capital formulation equals gross private non-residential fixed capital formulation at current prices deflated by p minus depreciation at
current replacement for private non-residential investment
deflated by p (sources: ETAS and BLUE).
(K a ) Cumulative Net Real Investment Abroad
(excluding change in official reserves)
Cumulative net real foreign investment (excluding changes
in official reserves) since the beginning of 1975 plus a constant derived from the official estimate of net external assets
(excluding official reserves) at the end of 1996 adjusted to
1990 prices. Net real foreign investment equals net foreign
investment (excluding increase in official reserves) at current prices deflated by p (sources: ETAS and UKBP).
(L) Employment
Employed Labour Force (source: ETAS).
(P ) Real Profits, Interest and Dividends from Abroad
Net interest, profits and dividends from abroad deflated by
p (source: ETAS).
264
( p) Price Level
Gross domestic product at market prices divided by
gross domestic product at 1990 market prices (source:
ETAS).
(Q ) Real Net Output
Net domestic product at current prices deflated by p (source:
ETAS).
(q) Exchange Rate (price of sterling in foreign
currency)
Sterling effective exchange rate average (1990 = 1.0000)
(source: BOE).
(r ) Interest Rate
Yield on long-dated UK government securities adjusted to
instantaneous rate per quarter (source: ETAS).
(S) Inventories
Value of real physical increase in stocks at current prices
since the beginning of 1975 plus a constant derived from
the official estimate of the book value of stocks and work
in progress held at the end of 1996 adjusted to 1990 prices.
(sources: ETAS and BLUE).
(w) Wage Rate
Income from employment (including pay of forces) divided
by sum of employees in employment and HM forces
(sources: ETAS and MD).
265
(B) Bonds
Sterling debt (British government stock) held by the
domestic private sector, million, stock outstanding (source:
BOE).
(E o ) Real Oil Exports
Oil exports at current prices deflated by p (source: ET).
(Gc ) Real Government Consumption
Central government consumption deflated by p plus local
authorities consumption deflated by p (source: ETAS).
(K p ) Public Non-Residential Fixed Capital
Cumulative net real public non-residential fixed capital formulation since the beginning of 1975 plus a constant derived
from the official estimate of the capital stock for public nonresidential fixed capital at the end of 1996 adjusted to 1990
prices. Net real public non-residential fixed capital formulation equals gross public non-residential fixed capital formulation at current prices deflated by p minus depreciation at
current replacement for public non-residential investment
deflated by p (sources: ETAS and BLUE).
( pf ) Price Level in Leading Foreign Industrial
Countries
Average of implicit price deflator of US GDP and implicit
price deflator of German GDP (source: IFS).
266
( pi ) Price of Imports (in foreign currency)

Implicit price deflator of imports multiplied by q (source:
ETAS).
(M) Money Supply
Money aggregate M0 (source: BOE)
(r f ) Foreign Interest Rate
Average of US and German long-term government bond
yields adjusted to instantaneous quarterly rate (source: IFS).
(T1 ) Total Taxation Policy Variable
Defined so that (Q + P)/T1 = real private disposable income.
Private disposable income equals net national income at
market prices plus transfers less direct and indirect taxes
(souces: ETAS).
(T2 ) Indirect Taxation Policy Variable
Gross domestic product at market prices divided by gross
domestic product at factor cost (source: ETAS).
(Y f ) Real Income in Leading Foreign Industrial
Countries
Average of index of US real GDP (1990 = 1.0000) and index
of German real GDP (1990 = 1.0000) (source: IFS).
267
References
Adkins, L. C. and Krehbiel, T. [1999]: Mean Reversion and Volatility of Short-Term London Interbank Offer Rates. International
Review of Economics and Finance 8, 4554.
Agbeyegbe, T. D. [1984]: The Exact Discrete Analog to a Closed
Linear Mixed-Order System. Journal of Economics Dynamics and
Control 7, 363375.
Agbeyegbe, T. D. [1987]: The Exact Discrete Analog to a Closed
Linear First-Order System with Mixed Sample. Econometric
Theory 3, 142149.
Agbeyegbe, T. D. [1988]: An Exact Discrete Analog of an Open
Order Linear Non-Stationary First-Order Continuous-Time
System with Mixed Sample. Journal of Econometrics 39, 237
250.
At-Sahalia, Y. [1996]: Testing Continuous-Time Models of the
Spot Interest Rate. Review of Financial Studies 9, 385342.
Andersen, T. G. and Lund, J. [1997]: Estimating Continuous-Time
Stochastic Volatility Models of the Short-Term Interest Rate.
Journal of Econometrics 77, 343377.
269
References
Babbs, S. H. and Nowman, K. B. [1999]: Kalman Filtering of Generalized Vasicek Term Structure Models. Journal of Financial and
Quantitative Analysis 34, 115130.
Bachelier, L. [1900]: Theorie

De La Speculation. Ann. Sci. Ecole
Norm 3, Sup., 2186.
Bailey, P. W., Hall, V. B. and Phillips, P. C. B. [1987]: A Model of
Output, Employment, Capital Formulation and Inflation. In G.
Gandolfo and F. Marzano (Eds.) Keynesian Theory, Planning Models
and Quantitative Economics: Essays in Memory of Vittorio Marrama,
` Milano.
Vol. 2, Giuffre,
Bandi, F. M. and Phillips, P. C. B. [2003]: Fully Nonparametric
Estimation of Scalar Diffusion Models. Econometrica 71, 241283.
Barnett, W. A. and He, Y. [1998]: Center Manifold, Stability, and Bifurcations in Continuous Time Macroeconometric Systems. Unpublished
Paper.
Barnett, W. A. and He, Y. [1999]: Stability Analysis of ContinuousTime Macroeconometric Systems. Studies in Nonlinear Dynamics
and Econometrics 3, 169188.
Barnett, W. A. and He, Y. [2001]: Unsolved Econometric Problems
in Nonlinearity, Chaos and Bifurcation. Central European Journal
of Operations Research 9, 147182.
Barnett, W. A. and He, Y. [2002]: Stabilization Policy as Bifurcation Selection: Would Stabilization Policy Work If the Economy
Really Were Unstable? Macroeconomic Dynamics 6, 713747.
Barnett, W. A. and He, Y. [2004]: Bifurcations in Macroeconomic
Models. In S. Dowrick, R. Pitchford and S. Turnovsky (Eds),
Economic Growth and Macroeconomic Dynamics: Recent Developments
in Economic Theory. Cambridge: Cambridge University Press, 95
112.
Barnett, W. A. and He, Y. [2005a]: Robustness of Inferences to
Singularity Bifurcations. Proceedings of the Joint Statistical Meetings
of the American Statistical Association. Forthcoming.
270
References
Barnett, W. A. and He, Y. [2005b]: New Phenomena Identified in a

Stochastic Dynamic Macroeconometric Model: A Bifurcation Perspective.
Unpublished Paper, University of Kansas.
Barnett, W. A. and He, Y. [2006]: Singularity Bifurcation. Journal
of Macroeconomics 28, 522.
Bartlett, M. S. [1946]: On the Theoretical Specification and Sampling Properties of Autocorrelated Time-Series. Journal of the
Royal Statistical Society Supplement 8, 2741.
Beaglehole, D. R. and Tenney, M. S. [1991]: General Solutions of
Some Interest Rate-Contingent Claim Pricing Equations. Journal
of Fixed Income 1, 6983.
Bellman, R. [1953]: Stability Theory of Differential Equations. New
York.
Bergstrom, A. R. [1966a]: Non-Recursive Models as Discrete
Approximations to Systems of Stochastic Differential Equations.
Econometrica 34, 173182.
Bergstrom, A. R. [1966b]: Monetary Phenomena and Economic
Growth: A Synthesis of Neoclassical and Keynesian Theories.
Economic Studies Quarterly 17, 18.
Bergstrom, A. R. [1967]: The Construction and Use of Economic Models.
London: The English Universities Press Ltd.
Bergstrom, A. R. and Wymer, C. R. [1976]: A Model of Disequilibrium Neoclassical Growth and Its Application to the United
Kingdom. In A. R. Bergstrom (Ed.), Statistical Inference in Continuous Time Economic Models. Amsterdam: North-Holland.
Bergstrom, A. R. [1978]: Monetary Policy in a Model of the United
Kingdom. In A. R. Bergstrom, A. J. L. Catt, M. H. Peston and
B. D. J. Silverstone (Eds.), Stability and Inflation. New York:
Wiley.
Bergstrom, A. R. [1983]: Gaussian Estimation of Structural Parameters in Higher-Order Continuous Time Dynamic Models. Econometrica 51, 117152.
271
References
Bergstrom, A. R. [1984a]: Continuous Time Stochastic Models

and Issues of Aggregation Over Time. In Z. Griliches and M. D.
Intriligator (Eds.) Handbook of Econometrics. Amsterdam: NorthHolland, vol. 2, 11461212.
Bergstrom, A. R. [1984b]: Monetary, Fiscal and Exchange Rate
Policy in a Continuous Time Model of the United Kingdom.
In P. Malgrange and P. Muet (Eds.), Contemporary Macroeconomic
Modelling. Oxford: Blackwell.
Bergstrom, A. R. [1985]: The Estimation of Parameters in Nonstationary Higher-Order Continuous Time Dynamic Models. Econometric Theory 1, 369385.
Bergstrom, A. R. [1986]: The Estimation of Open Higher-Order
Continuous Time Dynamic Models with Mixed Stock and Flow
Data. Econometric Theory 2, 350373.
Bergstrom, A. R. [1987]: Optimal Control in Wide-Sense Stationary Continuous Time Stochastic Models. Journal of Economic
Dynamics and Control 11, 425443.
Bergstrom, A. R. [1988]: The History of Continuous-Time Econometric Models. Econometric Theory 4, 350373.
Bergstrom, A. R. [1989]: Optimal Forecasting of Discrete Stock
and Flow Data Generated by a Higher Order Continuous Time
System. Computers and Mathematics with Applications 17, 1203
1214.
Bergstrom, A. R. [1990]: Continuous Time Econometric Modelling.
Oxford: Oxford University Press.
Bergstrom, A. R., Nowman, K. B. and Wymer, C. R. [1992]: Gaussian Estimation of a Second Order Continuous Time Macroeconometric Model of the United Kingdom. Economic Modelling
9, 313351.
Bergstrom, A. R., Nowman, K. B. and Wandasiewicz, S. [1994]:
Monetary and Fiscal Policy in a Second Order Continuous Time
272
References
Macroeconometric Model of the United Kingdom. Journal of

Economics Dynamics and Control 18, 731761.
Bergstrom, A. R. [1996]: Survey of Continuous Time Econometrics. In W. A. Barnett, G. Gandolfo and C. Hillinger (Eds.),
Dynamic Disequilibrium Modelling. Cambridge: Cambridge University Press.
Bergstrom, A. R. [1997]: Gaussian Estimation of Mixed Order Continuous Time Dynamic Models with Unobservable Stochastic
Trends from Mixed Stock and Flow Data. Econometric Theory 13,
467505.
Bergstrom, A. R. and Nowman, K. B. [1999]: Gaussian Estimation of a Two-Factor Continuous Time Model of the Short-Term
Interest Rate. Economic Notes 28, 2541.
Bergstrom, A. R. [2001]: Stability and Wage Acceleration in
Macroeconomic Models of Cyclical Growth. Journal of Applied
Econometrics 16, 327340.
Black, F. and Scholes, M. [1973]: The Pricing of Options and Corporate Liabilities. Journal of Political Economy 81, 637654.
Brennan, M. J. and Schwartz, E. S. [1979]: A Continuous Time
Approach to the Pricing of Bonds. Journal of Banking and Finance
3, 133155.
Brennan, M. J. and Schwartz, E. S. [1980]: Analyzing Convertible
Bonds. Journal of Financial and Quantitative Analysis 15, 907929.
Chan, K. C., Karolyi, G. A., Longstaff, F. A. and Sanders, A. B.
[1992]: An Empirical Comparison of Alternative Models of the
Short-Term Interest Rate. Journal of Finance 47, 12091227.
Chamberlin, E. H. [1946]: The Theory of Monopolistic Competition.
Harvard: Harvard University Press.
Chambers, M. J. [1991]: Discrete Models for Estimating General Linear Continuous Time Systems. Econometric Theory 7,
531542.
273
References
Chambers, M. J. [1999]: Discrete Time Representation of Stationary and Non-Stationary Continuous Time Systems. Journal of
Economic Dynamics and Control 23, 619639.
Chambers, M. J. [2001]: Temporal Aggregation and the Finite
Sample Performance of Spectral Regression Estimators in Cointegrated Systems: A Simulation Study. Econometric Theory 17,
591607.
Chambers, M. J. [2003]: The Asymptotic Efficiency of Cointegration Estimators under Temporal Aggregation. Econometric Theory
19, 4977.
Chen, L. [1996]: Interest Rate Dynamics, Derivatives Pricing, and Risk
Management. Springer, Lecture Notes in Economics and Mathematical Systems.
Chen, R. R. and Scott, L. [1992]: Pricing Interest Rate Options in
a Two-Factor CoxIngersollRoss Model of the Term Structure.
Review of Financial Studies 5, 613636.
Coddington, E. A. and Levinson, N. [1955]: Theory of Ordinary Differential Equations. New York: McGraw Hill.
Corradi, V. [1997]: Comovements between Diffusion Processes:
Characterization, Estimation, and Testing. Econometric Theory 13,
646666.
Courant, R. [1936]: Differential and Integral Calculus. London:
Blackie and Sons Ltd., vol. II.
Cox, J. C. [1975]: Notes on Option Pricing 1: Constant Elasticity of
Variance Diffusion. Working Paper, Stanford University.
Cox, J. C. and Ross, S. [1976]: The Valuation of Options for Alternative Stochastic Processes. Journal of Financial Economics 3, 145
166.
Cox, J. C., Ingersoll, J. E. and Ross, S. A. [1985]: A Theory
of the Term Structure of Interest Rates. Econometrica 53, 385
407.
274
References
Cox, J. C., Ingersoll, J. E. and Ross, S. [1980]: An Analysis of Variable Rate Loan Contracts. Journal of Finance 35, 389403.
Dahlquist, M. [1996]: On Alternative Interest Rate Processes. Journal of Banking and Finance 20, 10931119.
Dothan, U. L. [1978]: On the Term Structure of Interest Rates.
Journal of Financial Economics 6, 5969.
Donaghy, K. P. [1993]: A Continuous-Time Model of the United
States Economy. In G. Gandolfo (Ed.), Continuous Time Econometrics. London: Chapman and Hall.
Duffie, D. and Kan, R. [1996]: A Yield-Factor Model of Interest
Rates. Mathematical Finance 6, 379406.
Durbin, J. [1961]: Efficient Fitting of Linear Models for Continuous
Time Stationary Time Series from Discrete Data. Bulletin of the
International Statistical Institute 38, 273282.
Durbin, J. [1963]: Trend Elimination for the Purpose of Estimating Seasonal and Periodic Components of Time Series. In M.
Rosenblatt (Ed.), Time Series Analysis. New York: John Wiley.
Einstein, A. [1906]: Zur Theorie Der Brownschen Bewegung.
Annalen Der Physik 19, 371381.
Episcopos, A. [2000]: Further Evidence on Alternative Continuous
Time Models of the Short-Term Interest Rate. Journal of International Financial Markets, Institutions and Money 10, 199212.
Engle, R. F. and Granger, C. W. J. [1987]: Cointegration and Error
Correction: Representation Estimation and Testing. Econometrica
55, 251276.
Franses, P. H. [1996]: Periodicity and Stochastic Trends in Economic
Time Series. Oxford: Oxford University Press.
Gandolfo, G. [1981]: Quantitative Analysis and Econometric Estimation
of Continuous Time Dynamic Models. Amsterdam: North-Holland.
Gandolfo, G. and Padoan, P. C. [1982]: Policy Simulations
with a Continuous Time Macrodynamic Model of the Italian
275
References
Economy: A Preliminary Analysis. Journal of Economic Dynamics

and Control 4, 205224.
Gandolfo, G. and Padoan, P. C. [1984]: A Disequilibrium Model of Real
and Financial Accumulation in an Open Economy. Berlin: SpringerVerlag.
Gandolfo, G. and Padoan, P. C. [1987a]: The Mark V Version of the
Italian Continuous Time Model. Siena: Instituto Di Economia Della
Facolta Di Scienze Economiche E Bancarie.
Gandolfo, G. and Petit, M. L. [1987b]: Dynamic Optimization in
Continuous Time and Optimal Policy Design in the Italian Economy. Annales D Economie Et De Statistique 5/6, 311333.
Gandolfo, G. and Padoan, P. C. [1990]: The Italian Continuous
Time Model, Theory and Empirical Results. Economic Modelling
7, 91132.
Gandolfo, G. (Ed.) [1993]: Continuous Time Econometrics: Theory and
Applications. London: Chapman-Hall.
Hansen, L. P. [1982]: Large Sample Properties of Generalized
Method of Moments Estimators. Econometrica 50, 10291054.
Hansen, L. P. and Sargent, T. J. [1983]: The Dimensionality of the
Aliasing Problem in Models with Rational Spectral Densitites.
Harvey, A. C. and Stock, J. H. [1985]: The Estimation of HigherOrder Continuous Time Autoregressive Models. Econometric
Theory 1, 97117.
Harvey, A. C., Henry, S. G. B., Peters, S. and Wren-Lewis, S. [1986]:
Stochastic Trends in Dynamic Regression Models: An Application to the EmploymentOutput Equation. Economic Journal 96,
975985.
Harvey, A. C. and Stock, J. H. [1988]: Continuous Time Autoregressive Models with Common Stochastic Trends. Journal of
Economic Dynamics and Control 12, 365384.
276
References
Harvey, A. C. [1989]: Forecasting, Structural Time Series Models and

the Kalman Filter. Cambridge: Cambridge University Press.
Hiraki, T. and Takezawa, N. [1997]: How Sensitive Is Short-Term
Japanese Interest Rate Volatility to the Level of the Interest Rate?
Economic Letters 56, 325332.
Houthakker, H. S. and Taylor, L. D. [1966]: Consumer Demand in
the United States, 19291970, Analysis and Projections. Cambridge:
Harvard University Press.
Hull, J. and White, A. [1994]: Numerical Procedures for Implementing Term Structure Models II: Two-Factor Models. Journal
of Derivatives, Winter, 3748.
Ito, K. [1946]: On a Stochastic Integral Equation. Proceedings of the
Japanese Academy 1, 3235.
Ito, K. [1951]: On Stochastic Differential Equations. Memoir of the
American Mathematical Society 4, 51.
Jiang, G. and Knight, J. [1997]: A Nonparametric Approach to
the Estimation of Diffusion Processes with an Application to a
Short-Term Interest Rate Model. Econometric Theory 12, 615645.
Jonson, P. D., Moses, E. R. and Wymer, C. R. [1977]: The RBA
76 Model of the Australian Economy. In Proceedings Conference in
Applied Economic Research, Reserve Bank of Australia, Sydney, 936.
Jonson, P. D. and Trevor, R. G. [1981]: Monetary Rules: A Preliminary Analysis. Economic Record 57, 150167.
Jonson, P. D., Mckibbin, W. J. and Trevor, R. G. [1982]: Exchange
Rates and Capital Flows: A Sensitivity Analysis. Canadian Journal
of Economics 15, 669692.
Johansen, S. [1988]: Statistical Analysis of Cointegrating Vectors.
Journal of Economic Dynamics and Control 12, 231254.
Johansen, S. [1991]: Estimation and Hypothesis Testing of Cointegration Vectors in Gaussian Vector Autoregressive Models.
277
References
Kawai, K. and Maekawa, K. [2003]: A Note on Yu-Phillips Estimation for a Continuous Time Model of a Diffusion Process. Unpublished
Paper, Hiroshima University.
Kirkpatrick, G. [1987]: Employment, Growth and Economic Policy: An
Econometric Model of Germany. Tubingen:

Mohr.
Knight, M. D. and Mathieson, D. J. [1979]: Model of an Industrial
Country under Fixed and Flexible Exchange Rates. In J. Martin
and A. Smith (Eds.), Trade and Payments Adjustment under Flexible
Rates. London: Macmillan.
Knight, M. D. and Wymer, C. R. [1978]: A Macroeconomic Model
of the United Kingdom. IMF Staff Papers 25, 742778.
Kondratieff, N. D. [1926]: Die Langen Wellen Der Konjunk Sozialwissenschaft Und Sozialpolitik LVI, 573
tur. Archiv Fur
609.
Kondratieff, N. D. [1935]: The Long Waves in Economic Life.
Review of Economic Statistics XVII, 105115. (Translation of Kondratieff 1926)
Koopmans, T. C. [1950]: Models Involving a Continuous Time
Variable. In T. C. Koopmans (Ed.), Statistical Inference in Dynamic
Economic Models. New York: Wiley.
Kydland, F. E. and Prescott, E. C. [1977]: Rules Rather than
Discretion: The Inconsistency of Optimal Plans. Journal of Political
Economy 85, 473491.
Langetieg, T. C. [1980]: A Multivariate Model of the Term Structure. Journal of Finance 35, 7197.
Lo, K. M. [2005]: An Evaluation of MLE in a Model of the Nonlinear
Continuous-Time Short-Term Interest Rate. Working Paper Number
45, Bank of Canada.
Longstaff, F. and Schwartz, E. S. [1992]: Interest Rate Volatility and
the Term Structure: A Two-Factor General Equilibrium Model.
Journal of Finance 47, 12591282.
278
References
Lucas, R. E. [1976]: Econometric Policy Evaluation: A Critique. Carnegie-Rochester Conference Series on Public Policy 1,
1946.
Malinvaud, E. [1966]: Statistical Methods of Econometrics. NorthHolland, Amsterdam.
Melino, A. [1994]: Estimation of Continuous-Time Models in
Finance. In C. A. Sims (Ed.), Advances in Econometrics: Sixth World
Congress. Cambridge: Cambridge University Press, vol. II.
Merton, R. C. [1969]: Lifetime Portfolio Selection under Uncertainty: The Continuous Time Case. Review of Economics and
Statistics 51, 247257.
Merton, R. C. [1971]: Optimum Consumption and Portfolio Rules
in a Continuous Time Model. Journal of Economic Theory 3, 373
413.
Merton, R. C. [1973a]: Theory of Rational Option Pricing. Bell
Journal of Economics 4, 141183.
Merton, R. C. [1973b]: An Intertemporal Capital Asset Pricing
Model. Econometrica 41, 867887.
Merton, R. C. [1990]: Continuous Time Finance. New York: Oxford
University Press.
McCrorie, J. R. [2000]: Deriving the Exact Discrete Analog of a
Continuous Time System. Econometric Theory 16, 9981015.
McCrorie, J. R. [2001]: Interpolating Exogenous Variables in Open
Continuous Time Dynamic Models. Journal of Economic Dynamics
and Control 25, 13991427.
McCrorie, J. R. [2002]: The Likelihood of the Parameters of a Continuous Time Vector Autoregressive Model. Statistical Inference for
Stochastic Processes 5, 273286.
McGarry J. S. [2003]: The Exact Discrete Time Representation of
a System of Fourth-Order Differential Equation. Computers and
Mathematics with Applications 46, 213230.
279
References
Niewenhuis, H. J. [1995]: Continuous Time Macroeconometric Modelling with An Application to the Dutch Economy. Den Haag:
Labyrinth.
Nowman, K. B. [1991]: Open Higher-Order Continuous Time
Dynamic Model with Mixed Stock and Flow Data and Derivatives of Exogenous Variables. Econometric Theory 7, 404
408.
Nowman, K. B. [1993]: Finite-Sample Properties of the Gaussian
Estimation of an Open Higher-Order Continuous-Time Dynamic
Model with Mixed Stock and Flow Data. In G. Gandolfo (Ed.),
Continuous Time Econometrics: Theory and Applications. London:
Chapman-Hall.
Nowman, K. B. [1996]: A Note on Continuous Time Dynamic
Disequilibrium Macroeconometric Modelling of The United
Kingdom. In W. A. Barnett, G. Gandolfo and C. Hillinger
(Eds.), Dynamic Disequilibrium Modelling. Cambridge: Cambridge
University Press.
Nowman, K. B. [1997]: Gaussian Estimation of Single-Factor Continuous Time Models of the Term Structure of Interest Rates.
Journal of Finance 52, 16951706.
Nowman, K. B. [1998a]: Continuous-Time Short Term Interest
Rate Models. Applied Financial Economics 8, 401407.
Nowman, K. B. [1998b]: Econometric Estimation of a Continuous Time Macroeconomic Model of the United Kingdom with
Segmented Trends. Computational Economics 12, 243254.
Nowman, K. B. and Sorwar, G. [1999]. An Evaluation of Contingent Claims Using the CKLS Interest Rate Model: An Analysis
of Australia, Japan, and United Kingdom. Asia Pacific Financial
Markets 6, 205219.
Nowman, K. B. [2001]: Gaussian Estimation and Forecasting of
Multi-Factor Term Structure Models with an Application to
280
References
Japan and the United Kingdom. Asia Pacific Financial Markets 8,

2334.
Nowman, K. B. [2002]: The Volatility of Japanese Interest Rates:
Evidence for Certificate of Deposit and Gensaki Rates. International Review of Financial Analysis 11, 2938.
Nowman, K. B. [2003]: A Note on Gaussian Estimation of the
CKLS and CIR Models with Feedback Effects for Japan. Asia
Pacific Financial Markets 10, 275279.
guez, T. M. [2005a]: Estimating the DynamNowman, K. B. and N
ics of Interest Rates in the Japanese Economy. Unpublished Paper,
Westminster Business School.
Nowman, K. B. and Sorwar, S. [2005b]: Derivative Prices from
Interest Rate Models: Results for Canada, Hong Kong and
United States. International Review of Financial Analysis 14, 428
438.
Nowman, K. B. [2006]: Continuous Time Interest Rate Models in
Japanese Fixed Income Markets. In J. A. Batten, T. A. Fetherston
and P. G. Szilagyi (Eds.), Japanese Fixed Income Markets: Money,
Bond and Interest Rate Derivatives. Elsevier.
Park, J. Y. and Phillips, P. C. B. [1988]: Statistical Inference
in Regressions with Integrated Processes: Part 1. Econometric
Theory 4, 468497.
Park, J. Y. and Phillips, P. C. B. [1989]: Statistical Inference
in Regressions with Integrated Processes: Part 2. Econometric
Theory 5, 95131.
Phillips, A. W. [1954]: Stabilization Policy in a Closed Economy.
Economic Journal 64, 290323.
Phillips, A. W. [1959]: The Estimation of Parameters in Systems of
Stochastic Differential Equations. Biometrika 46, 6776.
Phillips, P. C. B. [1972]: The Structural Estimation of a Stochastic
Differential Equation System. Econometrica 40, 10211041.
281
References
Phillips, P. C. B. [1973]: The Problems of Identification in Finite

Parameter Continuous Time Models. Journal of Econometrics 1,
351362.
Phillips, P. C. B. [1974a]: The Estimation of Some Continuous Time
Models. Econometrica, 42, 803824.
Phillips, P. C. B. [1974b]: Problems in the Estimation of Continuous
Time Models, Ph.D. Thesis, University of London.
Phillips, P. C. B. [1976]: The Estimation of Linear Stochastic Differential Equations with Exogenous Variables. In A. R. Bergstrom
(Ed.), Statistical Inference in Continuous Time Economic Models.
Amsterdam: North-Holland.
Phillips, P. C. B. and Durlauf, S. N. [1986]: Multiple Time Series
with Integrated Variables. Review of Economic Studies 53, 473496.
Phillips, P. C. B. [1987]: Time Series Regression with a Unit Root.
Phillips, P. C. B. [1991a]: Optimal Inference in Cointegrated Systems. Econometrica 59, 283306.
Phillips, P. C. B. [1991b]: Error Correction and Long-Run Equilibrium in Continuous Time. Econometrica 59, 967980.
Phillips, P. C. B. [1995]: Fully Modified Least Squares and Vector
Autoregression. Econometrica 63, 10231079.
Phillips, P. C. B. and Yu, J. [2005]: Jackknifing Bond Option Prices.
Review of Financial Studies 18, 707742.
Revuz, D. and Yor, M. [1999]: Continuous Martingales and Brownian
Motion. New York: Springer.
Saltoglu,
B. [2003]: Comparing Forecasting Ability of Parametric and Nonparametric Methods: Application with Canadian
Monthly Interest Rates. Applied Financial Economics 13, 169176.
Sargan, J. D. [1974]: Some Discrete Approximations to Continuous Time Stochastic Models. Journal of the Royal Statistical Society,
Series B, 36, 7490.
282
References
Sargan, J. D. [1976]: Some Discrete Approximations to Continuous Time Stochastic Models. In A. R. Bergstrom (Ed.), Statistical
Inference in Continuous Time Economic Models. Amsterdam: NorthHolland.
Sassanpour, C. and Sheen, J. [1984]: An Empirical Analysis of the
Effect of Monetary Disequilibrium in Open Economies. Journal
of Monetary Economics 13, 127163.
Simos, T. [1996]: Gaussian Estimation of a Continuous Time
Dynamic Model with Common Stochastic Trends. Econometric
Theory 12, 361373.
B. [1993]: CONTIMOS A Continuous Time Econometric
Sjo o,
Model for Sweden Based on Monthly Data. In G. Gandolfo (Ed.),
Continuous Time Econometrics. London: Chapman and Hall.
Stanton, R. [1997]: A Nonparametric Model of Term Structure
Dynamics and the Market Prices of Interest Rate Risk. Journal of
Finance 52, 19732002.
Stavrev, E. [2001]: A Small Continuous Time Macro-Econometric
Model of the Czech Republic. Empirical Economics 26, 673705.
Stefansson, S. B. [1981]: Inflation and Economic Policy in a Small Open
Economy: Iceland in the Post-War Period. Ph.D. Thesis, University
of Essex.
Sundaresan, S. M. [2000]: Continuous-Time Methods in Finance:
A Review and an Assessment. Journal of Finance 55, 15691622.
Tse, Y. K. [1995]: Some International Evidence on the Stochastic Behavior of Interest Rates. Journal of International Money and
Finance 14, 721738.
Tullio, G. [1981]: Demand Management and Exchange Rate Policy:
The Italian Experience. IMF Staff Papers 28, 80117.
Van Loan C. F. [1978]: Computng Integrals Involving the Matrix
Exponential. IEEE Transactions on Automatic Control AC-23, 396
404.
283
References
Vasicek, O. [1977]: An Equilibrium Characterization of the Term

Structure. Journal of Financial Economics 5, 177188.
Wiener, N. [1923]: Differential Space. Journal of Mathematical
Physics 131174.
Wymer, C. R. [1972]: Econometric Estimation of Stochastic Differential Equation Systems. Econometrica, 40, 565577.
Yu, J. and Phillips, P. C. B. [2001]: A Gaussian Approach for Continuous Time Models of the Short-Term Interest Rate. Econometrics
Journal 4, 210224.
Zadrozny, P. A. [1988]: Gaussian Likelihood of Continuous Time
ARMAX Models When Data Are Stock and Flow at Different
Frequencies. Econometric Theory 4, 108124.
284
Author Index
Adkins, L. C., 21
Agbeyegbe, T. D., 14, 51
At-Sahalia, Y., 28
Andersen, T. G., 29
Courant, R., 155, 159

Cox, J. C., 19
Dahlquist, M., 21
Donaghy, K. P., 31, 115
Dothan, U. L., 19
Duffie, D., 29
Durbin, J., 3, 218
Durlauf, S. N., 46, 66, 72
Babbs, S. H., 29, 30

Bachelier, L., 2
Bailey, P. W., 31, 115
Bandi, F. M., 28
Barnett, W. A., 43, 195, 196
Bartlett, M. S., 2
Bellman, R., 35, 174, 191
Beaglehole, D. R., 29
Black, F., 2, 7, 18
Brennan, M. J., 19, 29, 31
Engle, R. F., 46, 66

Einstein, A., 2
Episcopos, A., 26
Franses, P. H., 46, 66
Chamberlin, E. H., 141

Chambers, M. J., 14, 46, 70
Chan, K. C., 19
Chen, L., 29
Chen, R. R., 29
Coddington, E. A., 35, 174, 191
Corradi, V., 46
Granger, C. W. J., 46, 66

Gandolfo, G., 7, 31, 42, 115,
164
Hall, V. B., 31, 115
Hansen, L. P., 20, 41
Harvey, A. C., 18, 30, 46, 56, 57
285
Author Index
Henry, S. G. B., 56
He, Y., 43, 195, 196
Hiraki, T., 21
Houthakker, H. S., 10
Hull, J., 29
Ingersoll, J. E., 19
Ito, K., 2
Jiang, G., 28
Johansen, S., 46, 66
Jonson, P. D., 31, 42, 115
Kan, R., 29
Karolyi, G. A., 19
Kawai, K., 27
Kirkpatrick, G., 31, 42, 115
Knight, M. D., 31, 114, 115
Knight, J., 28
Kondratieff, N. D., 239
Koopmans, T. C., 3
Krehbiel, T., 21
Kydland, F. E., 196
Langetieg, T. C., 29
Levinson, N., 35, 174, 191
Longstaff, F. A., 19, 29
Lo, K. M., 27
Lucas, R. E., 43
Lund, J., 29
Malinvaud, E., 29
Maekawa, K., 27
Mathieson, D. J., 31, 115
McCrorie, J. R., 14
McGarry J. S., 14
Mckibbin, W. J., 42
Melino, A., 20
Merton, R. C., 2, 7, 18, 19
Moses, E. R., 31, 115
Niewenhuis, H. J., 31, 115
guez, T. M., 27
N
Padoan, P.C., 31, 42, 115
Park, J. Y., 46, 66, 72
Petit, M. L., 42
Peters, S., 56
Phillips, A. W., 3, 17, 42
Phillips, P. C. B., 11, 12, 16, 17,
24, 26, 27, 28, 29, 31, 35, 41,
46, 66, 72, 115
Prescott, E. C., 196
Revuz, D., 27
Ross, S., 19
Saltoglu,
B., 31
Sanders, A. R., 19
Sargan, J. D., 10
Sargent, T. J., 41
Sassanpour, C., 42
Scholes, M. 2, 7, 18
Schwartz, E. S., 19, 29, 31
Scott, L., 29
Sheen, J., 42
Simos, T., 46
B., 31, 115
Sjo o,
Sorwar, G., 28
Stanton, R., 28
Stavrev, E., 32, 115
Stefansson, S. B., 42
Stock, J. H., 18, 46, 56, 57
Sundaresan, S. M., 7, 18, 19
286
Author Index
Takezawa, N., 21
Taylor, L. D., 10
Tenney, M. S., 29
Trevor, R. G., 42
Tse, Y. K., 21
Tullio, G., 42
Wiener, N., 2
Wren-Lewis, S., 56
Wymer, C. R., 10, 31, 32, 34,
35, 36, 41, 42, 44, 45, 48,
58, 114, 115, 121, 123, 124,
125, 126, 132, 152, 164, 173,
174, 192, 194, 195, 196, 218,
223
Van Loan C. F., 113

Vasicek, O., 19
Wandasiewicz, S., 42, 193, 194,
196
White, A., 29
Yor, M., 27
Yu, J., 26, 27, 28
Zadrozny, P. A., 18
287
Subject Index
aliasing, 41
approximate discrete model, 25,
27
asymptotic stability, 35, 174,
197, 238
autocovariance, 48, 85
Bank of England, 30, 262
Bank of Japan, 30
bifurcation, 43, 174, 192, 195,
196
bond, 2, 19, 2829, 117, 127,
135, 144145, 148149,
266267
Brownian motion, 19, 27, 30,
56, 69, 72, 74, 76, 80, 217,
221
(BRSC), 26, 28
capital, 4, 18, 3334, 39, 58,
115118, 128136, 142,
149150, 222, 224225,
236237, 263264, 266
charge cards, 118, 144

cointegrated, 46
consumption, 4, 1718, 33,
3940, 58, 115117,
125129, 136, 226, 248,
263, 266
control theory, 42, 43, 193194,
196
cycle, 17, 41, 238240,
248249
deep parameters, 43
discrete time, 26, 89, 20,
26, 46, 66, 80, 82, 85,
87
dividends, 39, 115, 117,
147148, 226, 264
drift, 1920, 27, 56, 58, 67, 71,
116, 142
dynamic, 67, 1314, 44,
4647, 79, 115, 123124,
128129, 144, 150, 152,
194195
288
Subject Index
dummy variable, 40, 117, 149

eigenvalue, 174, 191192, 195,
215, 237238, 240
employment, 8, 37, 39, 45, 48,
56, 58, 71, 115, 117118,
130134, 141, 143, 222,
225, 238, 248, 265
exact discrete model, 1012, 16,
2627, 2930, 36, 48,
5152, 58, 62
exchange control, 40, 117, 149
exchange rate, 40, 43, 116117,
141142, 144145,
149151, 180, 194, 197,
234, 265
expectations, 4344, 57, 7577,
80, 123124, 128130, 140,
151, 154155, 158, 176
exports, 4, 32, 3940, 58,
115117, 136, 146, 226,
263, 266
fiscal policy, 4243, 193194
forecasting, 5, 8, 41, 4547,
7879, 125, 214, 240241,
247, 249
foreign interest, 40, 117, 145,
148149
gaussian estimation, 14, 17, 24,
26, 28, 3031, 46, 4851,
66, 248
generalized vasicek, 30
government, 3, 56, 8, 1819,
40, 117, 127, 263, 265,
266267
identification, 4142, 44
imports, 4, 34, 3940, 58,
115117, 136137,
145146, 150, 222, 232,
263, 267
income, 1719, 28, 33, 40,
117118, 126127,
144149, 248, 265, 267
indirect taxation, 40, 118, 138,
267
inflation, 37, 45, 48, 118, 127,
134135, 141, 143,
148149, 222
interest rate, 1834, 40, 115,
117, 127, 131, 135136,
144145, 148151, 180,
197, 225, 242
investment, 39, 116117,
135, 147148, 225, 264,
266
kalman filter, 18, 30
labour, 32, 37, 4445, 118,
123, 131134, 139143,
194, 222, 232233, 236,
264
Lag distributions, 13, 58, 214,
223224, 226, 248
Lebesgue measure, 54
Lucas critique, 196
macroeconometric, 31, 3436,
43, 4647, 51, 71, 114,
124, 173, 196197, 213,
247
mean reversion, 1920, 27
289
Subject Index
mean square, 13, 55, 241242

monetary policy, 30, 43, 174,
194
money supply, 18, 37, 39,
45, 117, 140, 143, 151,
194, 222, 233234,
237
non-residential, 115117, 130,
134, 224, 264, 266
official reserves, 39, 117, 150,
264
oil, 3940, 115117, 146, 226,
263, 266
output, 4, 17, 33, 40, 43, 56, 58,
71, 115, 117118, 130133,
136140, 142, 146,
193194, 222, 226, 237
plastic money, 118, 144, 222
price level, 34, 40, 115, 117,
130131, 138, 140142,
146, 237
productivity, 44, 118, 123, 132
policy, 2, 3, 6, 30, 40, 4243,
117118, 125126,
193194, 196
Poincare-Liapounov-Perron,
35,
174, 185, 191
post-sample, 41, 49, 52, 66, 74,

7980, 214, 218, 233,
240241, 249
residential, 116, 128129, 225,
263
sales, 6, 131, 136137, 146,
232233, 241
stability analysis, 49, 173, 180,
194, 215
stochastic trends, 3, 4551, 54,
5657, 66, 74, 79, 114, 121,
144, 152, 215, 218, 247
taxation, 40, 117, 126, 247
technical progress, 32, 37, 45,
48, 51, 5657, 71, 138,
141142, 222
time inconsistency, 196
transfers abroad, 39, 115116,
147, 263
treasury bill, 26
VARMAX, 52, 63, 6566, 68, 70,
74, 7980, 87
wage, 34, 40, 44, 59, 71, 115,
117, 138, 141142, 150,
194195
290

Albert Rex Bergstrom, Khalid Ben Nowman-A Continuous Time Econometric Model of The United Kingdom With Stochastic Trends (2007)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Albert Rex Bergstrom, Khalid Ben Nowman-A Continuous Time Econometric Model of The United Kingdom With Stochastic Trends (2007)

Uploaded by

Copyright:

Available Formats

A Continuous Time Econometric Model of the United

Kingdom with Stochastic Trends

Albert Rex Bergstrom

Khalid Ben Nowman

CAMBRIDGE UNIVERSITY PRESS

This publication is in copyright. Subject to statutory exception

List of Figures and Tables

Foreword by Peter C. B. Phillips

Introduction to Continuous Time Modelling

Continuous Time Econometrics with

Steady State and Stability Analysis

Empirical Estimation of the Model

List of Figures and Tables

UK and US Interbank Interest Rates

Japan and Singapore Interbank Interest Rates

Lag Distribution Private Consumption

Lag Distribution Employment

Lag Distribution Residential Fixed Capital

Lag Distribution Private Non-Residential Fixed Capital

Lag Distribution Output

Lag Distribution Interest Rate

Lag Distribution Imports

Lag Distribution Non-Oil Exports

Lag Distribution Current Transfers Abroad

List of Figures and Tables

5.10 Lag Distribution Profits Interest and Dividends from

5.11 Lag Distribution Investment Abroad

Single Factor Continuous Time Interest Rate Models

Parameter Estimates from United Kingdom Quarterly

Estimated Time Lag Parameters

Values of Exogenous Variable Time Path Parameters

Derived Steady State Growth Rates

Derived Steady State Levels

Eigenvalues for Linear Approximation about Steady

Root Mean Errors of Post-Sample Multi-Period

Post-Sample Multi-Period Forecasts

The daunting complexity of aggregate economic behaviour

exogenous variables that involves 18 endogenous variables,

This book carries the indelible signature of Bergstroms

and aspirations for productive research in econometrics. In

incorporating unobservable stochastic trends. The book also

writing of this book. Royalties from this volume will be used

he worlds economies and many of its financial markets

models to represent the economy and these markets is one of

Continuous Time Econometric Model of UK with Stochastic Trends

continuous time record of observations and most economic

Introduction to Continuous Time Modelling

1.2 Why Model in Continuous Time

Continuous Time Econometric Model of UK with Stochastic Trends

Introduction to Continuous Time Modelling

the unit observation period. Continuous time econometric

Continuous Time Econometric Model of UK with Stochastic Trends

time path forecasts of manufacturing production and gross

Introduction to Continuous Time Modelling

main reason why economic variables adjust gradually, rather

Continuous Time Econometric Model of UK with Stochastic Trends

increase in the number of Ph.D. educated graduates in

Introduction to Continuous Time Modelling

1.3 Introduction to General Continuous

Continuous Time Econometric Model of UK with Stochastic Trends

[1984a] for a precise definition and interpretation of this

In using this model it was shown in Bergstrom [1966a] that

Introduction to Continuous Time Modelling