Business Cycle Fluctuations

BUSINESS CYCLE FLUCTUATIONS
AND ECONOMIC POLICY
No part of this digital document may be reproduced, stored in a retrieval system or transmitted in any form or
by any means. The publisher has taken reasonable care in the preparation of this digital document, but makes no
expressed or implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of information
contained herein. This digital document is sold with the clear understanding that the publisher is not engaged in
rendering legal, medical or any other professional services.
BUSINESS CYCLE FLUCTUATIONS
AND ECONOMIC POLICY
KHURSHID M. KIANI
Nova Science Publishers, Inc.

New York
Copyright © 2009 by Nova Science Publishers, Inc.
All rights reserved. No part of this book may be reproduced, stored in a retrieval system or
transmitted in any form or by any means: electronic, electrostatic, magnetic, tape, mechanical
photocopying, recording or otherwise without the written permission of the Publisher.
For permission to use material from this book please contact us:
Telephone 631-231-7269; Fax 631-231-8175
Web Site: http://www.novapublishers.com
NOTICE TO THE READER

The Publisher has taken reasonable care in the preparation of this book, but makes no expressed
or implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of
information contained in this book. The Publisher shall not be liable for any special,
consequential, or exemplary damages resulting, in whole or in part, from the readers’ use of, or
reliance upon, this material. Any parts of this book based on government reports are so
indicated and copyright is claimed for those parts to the extent applicable to compilations of
such works.
Independent verification should be sought for any data, advice or recommendations contained
in this book. In addition, no responsibility is assumed by the publisher for any injury and/or
damage to persons or property arising from any methods, products, instructions, ideas or
otherwise contained in this publication.
This publication is designed to provide accurate and authoritative information with regard to the
subject matter covered herein. It is sold with the clear understanding that the Publisher is not
engaged in rendering legal or any other professional services. If legal or any other expert
assistance is required, the services of a competent person should be sought. FROM A
DECLARATION OF PARTICIPANTS JOINTLY ADOPTED BY A COMMITTEE OF THE
AMERICAN BAR ASSOCIATION AND A COMMITTEE OF PUBLISHERS.
LIBRARY OF CONGRESS CATALOGING-IN-PUBLICATION DATA
Kiani, Khurshid M.
Business cycle fluctuations and economic policy / Khurshid M. Kiani.
p. cm.
ISBN 978-1-60741-406-3 (E-Book)
1. Business cycles--Kazakhstan. 2. Kazakhstan--Economic policy. I. Title.
HB3711.K484 2009
338.5'42--dc22
2008046176
  New York
Published by Nova Science Publishers, Inc.
CONTENTS
Preface ...................................................................................................vii
Chapter 1 Asymmetric Business Cycle Fluctuations via Time
Series Models and Neural Network Linearity Tests ................... 1
Chapter 2 Neural Network Forecasts Evaluations and Business
Cycle Fluctuations ..................................................................... 27
Chapter 3 A Robust Evidence of Business Cycle Asymmetries in
G7 Countries.............................................................................. 45
Chapter 4 Asymmetric Business Cycle Fluctuations and Contagion
Effects in G7 Countries ............................................................. 63
Chapter 5 Business Cycle Asymmetries in Asian Economies via
Nonlinear Time Series Models and Neural Networks ............... 85
Index ................................................................................................. 109
PREFACE
The primary purpose of this book entitled Business Cycle Fluctuations and
Economic Policy is to disseminate knowledge and understanding of the techniques
that can be employed for testing possible existence of similarity and asymmetries
in business cycle fluctuations in macroeconomic time series. In doing so, the book
assembles a number of papers that employ various parametric and non-parametric
time series techniques for modeling macroeconomic time series in addition to
modeling these series using the other non-parametric techniques such as artificial
neural networks to study whether business cycle fluctuations in all the countries
studied are alike.
The question whether business cycle fluctuations in G7 countries are alike
was unsuccessfully explored by previous research which poses a serious challenge
to macroeconomic theorist for building new theories on business cycle
fluctuations unless empirical evidence warrants that business cycle fluctuations in
all the countries are alike. Therefore, to answer this question of prime empirical
importance the present work employs a number of nonlinear augmented and
switching time series models and artificial neural networks that are considered to
be highly flexible functional form of nonlinear models to investigate whether
business cycle fluctuations in G7 countries are alike. Additionally, forecast
performance of the neural network vis-à-vis traditional time series models are also
compared to conclude that in-sample forecast performance of neural networks is
superior to the in-sample forecasts from linear and other nonlinear models but
inferior to the out-of-sample performance from linear models.
After establishing that business cycle fluctuations are asymmetrical in G7
countries, thus alike, the analysis is extended to a number of additional selected
countries that include, developed, developing, transition and other economics
particularly from the Asian region to study whether business cycle fluctuations in
viii Khurshid M. Kiani
this heterogeneous group of countries of the world are alike. Finally, based on the
results obtained from all the macroeconomic time series employed in different
spells of time using various nonlinear and switching time series models, and
artificial neural networks it can be concluded that business cycles in all the
countries studied are asymmetrical, thus alike.
The target audience of this book is graduate students, researchers, and
policymakers who would particularly be interested to know the impact of a unit
shock on the output in any of the countries studied that would destabilize the
economy. In such situations, the policymakers would be interested to know the
impact of a unit monetary shock on output so that appropriate stabilization
policies e.g. monetary policy is put into action to stabilize the economy. The
analyses of the book reveals that business cycles in all the countries studied (G7
plus 10 Asian economies) are asymmetrical, therefore, given all this I end up
concluding that linear model forecasts including those derived from linear vector
auto-regressions cannot be employed for forecasting the impact of any shock on
output when the underlying data generating process is nonlinear. Indeed, the
policymaker would need to employ appropriate nonlinear forecasting models to
anticipate the impact of one unit monetary policy or any other shocks on the
output in any of the economies studied. I hope you, the reader; find benefit in my
having done so.
Chapter 1
ASYMMETRIC BUSINESS CYCLE

FLUCTUATIONS VIA TIME SERIES MODELS
AND
NEURAL NETWORK LINEARITY TESTS
ABSTRACT
The present work employs highly flexible functional form of nonlinear
models that are also known as artificial neural network (ANN), and a few
nonlinear time series models for constructing linearity tests that are
employed for testing possible existence of business cycle asymmetries in
Canada, France, Japan, UK, and USA real GDP growth rates. While results
from each of the nonlinear time series models employed are used to construct
one type of linearity test, linearity tests constructed from artificial neural
networks approximations are of two types. One is neural network test for
neglected nonlinearities, and the other type is neural network test for possible
existence of nonlinearities in Canada, France, Japan, UK, and USA real GDP
growth rates.
Neural network tests for neglected nonlinearities show statistically
significant evidence of nonlinearities in all the series. Likewise, neural
network tests for possible existence of nonlinearities also show statistically
significant evidence of business cycle asymmetries in all the series.
Similarly, most linearity tests constructed from time series models also show
statistically significant evidence of business cycle asymmetries in all of the
series. Therefore, policymakers in these countries would not be able to
anticipate the impact of one unit monetary policy or any other shock on the
2 Khurshid M. Kiani
output based on forecasts obtained from linear models including those

derived from linear vector autoregressions.
Keywords: asymmetries; business cycles; neural network; nonlinearities;

principal components; real GDP
1. INTRODUCTION
People were not familiar with business cycles until nineteenth century when
economists thought of using new economic tools to learn the consequences of
shifts in aggregate supply and demand. However, starting early twentieth century,
economists started working on business cycle fluctuations and the factors
affecting business cycles. However, the quest to understand business cycles gotten
strengthened after the great depression of 1930s, whereupon the Keynesian
macroeconomics remained in tact for some time until 70s when Lucas (1976)
pioneered rational expectation hypothesis, and thereafter Kydland and Prescott
(1982), and Long and Plosser (1983) showed that business cycle models do not
encompass monetary factors and money management.
Business cycles were classified into a number of types according to their
peak-to-peak and trough-to-trough duration. Schumpeter (1939) divided business
cycles into four phases i.e. recession, depression, recovery and boom. While many
macroeconomic variables follow these phases, the span and the amplitude of a
business cycle would depend on its type, and in general a business cycle lasts for
about ten years (Sims, 1980).
Possible existence of business cycle asymmetries is imperative to detect for
several reasons. For instance nonlinearities imply that the effects of various
monetary policy shocks on the output are not symmetric; therefore, any
asymmetry would invalidate the measures of the persistence of monetary policy or
any other shock on output that is based on linear models as well as linear vector
autoregressions. Therefore, to validate the theories of business cycles such as real
business cycles (RBC) one would need to go beyond merely matching the first
and second moment of data with the moments implied by these theories.
Most of the empirical research on business cycles focuses on univeriate
econometric modeling techniques. For example, Neftci (1984), Brunner (1992,
1997), Beaudry and Koop (1993), Potter (1995), and Ramsey and Rothman
(1996) demonstrated business cycle asymmetries in macroeconomic time series.
Contrary to that Falk (1986), Sichel (1989), Delong and Summers (1986), and
Asymmetric Business Cycle Fluctuations via Time Series Models … 3
Diebold and Rudebusch (1990) have either failed to conclude that business cycles
are asymmetric or found weak evidences in this prospect.
Granger (1995) recommended testing linearity using heteroskedasticity-robust
test. French and Sichel (1993), and Brunner (1992, 1997) demonstrated the
existence of the conditional heteroskedasticity in real GNP data. Similarly,
Schienkman and Lebaron (1989) reported a weakening evidence of linearity after
accounting for conditional heteroskedasticity in GNP data.
Blanchard and Watson (1986) demonstrated presence of outliers in GNP data.
Tsay (1988) concluded that linearity could be rejected by the presence of outliers
in the data. Balke and Fomby (1994), and Scheinkman and LeBaron (1989)
reported weakened evidence against linearity in US real GNP data once outliers
were taken into account. Thus, there is a growing perception that the evidence of
nonlinearity in macroeconomic data reported in many studies so far could be
because of the presence of outliers.
Beaudry and Koop (1993) used an ad hoc nonlinear term in autoregressive
moving average framework to find nonlinearities in macroeconomic time series,
Terasvirta and Anderson (1992) used smooth transition autoregressive model to
study business cycles, Eithreim and Terasvirta (1996) tested adequacy of smooth
transition autoregressive model, and Luukkonen et al. (1988) used smooth
transition autoregressive model to study nonlinearities. Bidarkota (1999-2000)
found a robust evidence of nonlinearities even after accounting for time varying
volatility, outliers and long memory in the data series. Likewise, Kiani and
Bidarkota (2004) analyzed data from the group of seven (G 7) industrialized
countries i.e. Canada, France, Italy, Germany, Japan, UK, and USA real GDP
growth rates to test possible existence of nonlinearities in these countries and
concluded that strong evidence of asymmetries does exists in Canada, Italy,
Germany, Japan, and USA, but did not find any evidence of asymmetries in
France and UK series. This necessitated undertaking another study to detect
possible existence of business cycle asymmetries in all the series and possible
existence of neglected nonlinearities in France and UK real GDP growth rates.
The present work employs neural network test for neglected nonlinearities
(nnw1) due to Lee et al. (1993), and neural network test for possible existence of
neglected nonlinearities (nnw2) that was proposed by Terasvirta et al. (1993) for
testing possible existence of nonlinearities in Canada, France, Japan, UK, and
USA real GDP growth rates. In addition, time series linearity tests e.g. Keenan
(1985), Tsay (1986), and Ramsey (1969) and some of its improved versions
(RESET1, and RESET2) are also employed to test possible existence of
asymmetric business cycle fluctuations in all the series. Finally, results obtained
from both the neural network tests are compared with that of the entire set of
4 Khurshid M. Kiani
linearity test that are constructed from time series linearity tests for each of the
series.
The remaining work is split into the following sections. Section 2
incorporates the structure of neural network models and neural network tests.
Section 3 consists of a number of time series models and nonlinearity tests based
on forecasts from these models. In section 4 hypotheses tests, and nonlinearity
tests results are discussed, and finally section 5 incorporates brief conclusion.
2. ARTIFICIAL NEURAL NETWORKS

An artificial neural network (ANN) is an artificial intelligence technology that
imitates human brain’s learning, and its executive procedure. While any
individual neuron or node would be worthless, a combination of a number of
neuron would make an efficient network. ANN does possess an efficient
information processing capacity that enables them to learn from examples and
generalize these learnings to solve complex problems never seen before (Reilly
and Cooper, 1990).
The traditional statistical modeling approach requires distribution of the
underlying data generating processes, ANN are data driven, which can easily be
implemented to solve obscure problems. That is the reason why ANN modeling
approach is useful for forecasters, and researchers especially in problems where
data is available but the data generating process is unknown. Therefore, ANN are
treated as highly flexible functional form of nonlinear models that are independent
of the distribution of the underlying data generating processes (White 1989).
While a number of ANN have been proposed so far, one of the most
influential neural network models is the multilayer perceptrons (MLP), which
consists of several layers of nodes where information from sources external to
ANN is included, whereas the output layer is the highest node, where the solution
of the problem is realized. Both input and output layers are separated by a number
of layers called hidden nodes (Zhang et. al. (1998).
ANN can be used in time series modeling because these are highly nonlinear
functional form of models, and can be structured to fit any nonlinear time series
model as well as these can approximate any continuous function with desired
level of precision (Hornik et al.1990). The functional relationship of a typical
ANN model can be shown in following Equations:
yt = f ( yt −1, yt − 2 , yt −3 , ......, yt − k ) (1.1)

A typical single layer feed forward neural network as of Lee, et al. (1993) can
be written as follows.
k
f ( x, ξ ) = θ 0 + ∑θ {ψ (γ w )},
i =1
j
'
t k∈N (1.2)
~ ' )' ,
where, wt = (1, w and ~ = ( y , y , y , ......, y )' . Equation 1.2
w
t t t −1 t −2 t −3 t −k
shows flexible functional forms (Lee et al. (1993), and White (1989) where ψ is
a transfer function. The transfer function can be either sigmoid (logistic) or
hyperbolic (tangent) cumulative distribution function. Here sigmoid function is
used as transfer function, which is proposed by Lee, et al. (1993). The inputs
( y t −1 , y t − 2 , y t −3 , ......, y t − k ) are fed to hidden / signal-processing units that
augment/diminish the signal by some factor γ ij where, γ i = γ i 0 , γ i1 , γ i 2 ,......, γ ik .
The processed signals ψ ( ~x 'γ i ), i = 1,......., k are fed to the transfer

−x
function (ψ ( x) = (1 + e )) , which passes them to output.
A single layer feed forward neural networks model is estimated which is
sufficient to approximate any complex nonlinear function with desired level of
accuracy (Zhang et. al. (1998)). Most researchers use one single hidden layer for
forecasting purposes. However, one hidden layer may require a large number of
nodes, which is not desirable because it would worsen the network generalization
and increase the training time unnecessarily. Therefore, neural networks with
fewer hidden nodes are preferable for forecasting as they usually have better
generalization, and less over fitting problems. Likewise, a few nodes would be
preferred for neural network linearity tests because the test will have less power to
reject linearity when more nodes would be used.
Neural networks literature contains a number of rules of thumb to select the
number of nodes in a neural networks model. For example Lippmann (1987)
suggests 2n + 1 nodes, Wong (1991) proposes 2n , Tang and Fishwich (1993)
suggest n , Kastens et al. (1995) suggests 2 , and Kang (1991) recommends n / 2
nodes and Kiani (2005), and Kiani and Kastens (2008) recommended as low as
two nodes to be added in a neural network, where n is the number of the input
nodes. However, it is hard to say if any or all of these rules of thumb would be
adequate for real life problems.
Following Zhang et al. (1998), using trial and error approach, five nodes are
selected in the neural network architecture employed which appears to be
6 Khurshid M. Kiani
adequate when using neural network linearity test for test statistic that is proposed
by Lee et al. (1993). Contrary to the evidence from the empirical research
discussed in the preceding paragraphs, more biased nodes are used to approximate
neural networks for constructing neural network test 1 to calculate principal
components of ψ t matrix based on the selection of maximum q * < q , where q is
the number of maximum nodes in a neural networks model. However, for
approximating nnw2 model the number of nodes is restricted to as low as two.
These tests are established in the following paragraphs.
Before performing neural network tests, one lag for Canada and USA, two
lags for France and UK, and four lags for Japan are chosen, using Schwarz
Bayesian Criterion (SBC) for the lag order selection. These lags are employed in
restricted as well as unrestricted models for constructing neural network tests.
yt = π ' wt + ut (1.3)
where
u t ~ N (0, σ 2 ), ~ ' )' ,

wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = ( π 0 , π 1 ........, π p )' .
2.1. Neural Network Test 1 (nnw1)
This test is based on in-sample forecasts obtained from nnw1 model that was
proposed by Lee et al. (1993). This test is designed to test possible existence of
business cycle nonlinearities in time series data. The test compares a linear model
forecasts with approximations from a neural networks model. Both neural
network model and its linear counterpart can be described by the following two
equations:
yt = π ' wt + ut (1.4)
where,
~ ' )' ,
u t ~ N (0, σ 2 ), wt = (1, w and ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = (π 0 , π 1 ,........,π p )'

k
yt = π ' wt + ∑θ
j =1
0 j {ψ (γ
'
wt )} + vt (1.5)
where, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π 0 is intercept. Equation 1.5 is a neural

network model that nests a linear model of the form of Equation 1.4 . A slightly
modified version of Equation 1.5 is employed to recover model predictions. The
test consists of the following four-step procedure.
First of all, the contemporaneous variable ( yt ) is regressed on an intercept
and its lags ( y t −1 ,.......... yt −k ) using the regression equation of the form of
Equation 1.4 to recover residuals ( û t ), and predictions. In the second part of this
test ANN of the form of Equation 1.5 are approximated, where contemporaneous
variable ( yt ) is employed as dependent variable, and its lags ( y t −1 ,.......... y t − k ) are
used as independent variables in the model. From the ANN approximations, ψ t
matrix is recovered for principal component analyses1. This is done because the
test statistics prepared by Lee et al. (1993) is hard to implement, therefore the
present work employs an alternate test statistics proposed by Lee et al. However,
the elements of the ψ t matrix might be collinear with themselves or with the data
matrix X t . Therefore, principal component analysis of theψ t matrix is
undertaken to eliminate collinearity between its elements selecting
maximum q < q to obtain a matrix of principal components i.e. ψ t* matrix.
*
Finally, residuals ( û t ) are regressed on the matrix of principal components (ψ t* ),

and the data matrix ( X t ) to recover R 2 that is used to construct the test statistics
of the form of Equation 1.6 .
TS = nR 2 (1.6)
The results obtained from neural network linearity test 1 (nnw1) are shown in
Table 2 . In this Table column 2 rows 1 through 5 show test statistics for
1
See Dunteman (1989) for principal component explanation
8 Khurshid M. Kiani
Canada, France, Japan, UK and USA respectively. p-values for each of the test
statistic are juxtaposed in the next column of the Table.
2.2. Neural Network Test 2 (nnw2)
This test is based on forecasts obtained from neural network test for possible
existence of nonlinearities (nnw2) that was proposed by Tersvirta et al. (1993).
The test compares a linear model forecasts with approximations from a neural
networks model. Both the neural networks and its linear counterpart can be
described by the following two equations:
yt = π ' wt + ut (1.7)
where,
u t ~ Nid (0, σ 2 ), ~ ' )' ,

wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = (π 0 , π 1 ,........,π p )'

k
uˆt = π ' wt + ∑θ
j =1
0 j {ψ (γ j wt )} + vt
'
(1.8)
where, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π 0 is intercept. Equation 1.7 shows a

linear model whereas the Equation 1.8 shows a neural networks model that nests
the linear model represented by Equation 1.7 . In the first step yt is regressed on
an intercept and, y t −1 ,.......... y t − k as per regression Equation 1.7 . From this
regression, residual ( û t ) are recovered that are subsequently employed
with yt −1 ,.......... y t −k to approximate a neural network model as per Equation 1.8 .
From this model, residuals ( v̂t ) are recovered for constructing the test statistics in
conjunction with residual ( û t ) recovered from the earlier part as per Equation 1.9 .
TS = {( SSE1 − SSE 2) / m} /{SSE 2 /(n − p − m − 1)} (1.9)
where, m denotes the number of restrictions in the unrestricted model, n is the

number of observations, and p is the number of lags in the regression equations.
The test statistics (TS) is distributed F [m, (n − p − m − 1)] under normality

hypothesis.
Table 2 presents results from the test that is constructed form neural network
2 approximations for Canada, France, Japan, UK, and USA. In this Table row 6
to 10 in column 2 show test statistics respectively for Canada, France, Japan,
UK, and USA real GDP growth rates. p-values for each test are juxtaposed in
parentheses in the next column.
2.3. Estimation Issues
Neural network models are employed to test nonlinearities in all the series,
which is a nonparametric approach. While estimating neural network models
convergence appeared to be difficult because of the additive linear term in the
neural network models. To avoid the likelihood of obtaining local optima, genetic
algorithm (GA) was employed with a couple of random starts to obtain the best
parameter vector for neural network approximations. GA is considered to be a
reliable estimation algorithm but it appeared to be very slow, therefore, a
nonlinear optimization routine was combined with the GA, which worked
satisfactorily although it was still slower than any other optimization routine.
Following De Jong (1975), GA was used in biology, engineering and
operation research (Goldberg 1989). The first ever, economic application of GA
was implemented by Axelord (1987), and thereafter by Marimon, McGartten, and
Sargent (1990), and Dorsey and Walter (1995).
2.4. Data Sources
Quarterly data on real GDP growth rates series for Canada, France, Japan,
UK and USA is obtained from the International Financial Statistic (IFS)’s CD-
ROM for the month of September 2001. Canada, Japan, UK and USA data range
from 1957:1 to 2000:4, whereas the France data range from 1970:1 to 2000:4.
The analysis started with the data on real GDP growth rates for G7 countries,
i.e. Canada, France, Japan, UK and USA, but there was a jump of 35.43 percent in
Germany real GDP growth rates in 1991:1. This likely would have been because
of its reunification. Italy showed an inexplicable spike of 87.5 percent in 1970:1.
Therefore, series for Italy and Germany were excluded from the analysis. Table
1.1 describes the data, and Figure 1 shows plots of all the data series employed.
10 Khurshid M. Kiani
Canada F ra n c e
0.04 0.02
0.01
0.02
0
0
-0 . 0 1
-0 . 0 2 -0 . 0 2
1950 1960 1970 1980 1990 2 0 00 2010 1970 1 9 75 1980 1985 1990 1995 2000 2005
Ja p a n UK
0.15 0.06
0.04
0.1
0.02
0.05
0
0
-0 . 0 2
-0 . 0 5 -0 . 0 4
1950 1960 1970 1980 1990 2 0 00 2010 1950 1 96 0 1970 1980 19 9 0 2000 2010
US A
0.04
0.02
-0 . 0 2
-0 . 0 4 G D P G ro w t h R a t e s O ve r Tim e
1950 1960 1970 1980 1990 2 0 00 2010
Figure 1. Plot of Real GDP Growth Rates over Time.
2.5. Preliminary Data Analysis
Table 1.2 provides some useful statistics on the raw data and Table 1.2
summarizes results of routine statistical tests for some interesting data hypotheses.
The average annualized quarterly GDP growth rates range from 2.44 to 8.84
percent, with the average growth rates for all countries being significantly
positive. UK has the lowest, and Japan has the highest growth rate. The quarterly
(annualized) standard deviations range from 2.40 percent for France to 7.80
percent for Japan. Skewness measure ranges from –0.54 for France to 0.64 for
Japan, being statistically significant for all countries except Canada. Excess
kurtosis measures range from 0.14 for Canada to 2.85 for the UK, with significant
fat tails found in UK, and USA, and marginally in Japan. The Jarque-Bera test
rejects normality for all countries except Canada. These preliminary results are
subject to some qualification because the tests are based on the assumption that
the growth rates are independently and identically distributed (iid) normal. As will
be evident later, this assumption is not appropriate. The Augmented Dicky Fuller
(ADF) test indicates unit roots in levels (with constant and time trend) for all
countries but not in growth rates (with constant only). The only exception is
Japan, for which the test fails to reject unit roots in growth rates with a constant
term only but does reject with constant and time trend. A Goldfeld-Quandt test
fails to reject homoskedasticity in all countries, and the Lagrange Multiplier (LM)
test detects autoregressive conditional heteroskedasticity (ARCH) only in Japan.
Table 1.1. Data Description
Canada France Japan UK USA

Data Series Quarterly Quarterly Quarterly Quarterly Quarterly
Real GDP Real GDP Nominal GDP Real GDP Real GDP
Sample Period 1957:1- 1970:1- 1957:1- 1957:1- 1957:1-
2000:4 2000:4 2000:4 2000:4 2000:4
Sample Length 176 124 176 176 176
Notes:
1. The quarterly seasonally adjusted real GDP growth rates for Canada, France, Japan, UK,
and USA were obtained from the September 2001 edition of the International
Financial Statistics (IFS) CD-ROM.
2. The nominal GDP series are employed for Japan because seasonally adjusted data were
only available for the nominal series and not for the real series on the IFS CD-ROM.
Table 1.2. Summary Statistics

Mean 0.91 0.63 2.21 0.61 0.83
(p-value for mean = 0) (0.00) (0.00) (0.00) (0.00) (0.00)
Standard Deviation 0.97 0.60 1.95 1.05 0.95
Skewness 0.17 -0.54 0.64 0.36 -0.44
(p-value for skewness = 0) (0.49) (0.01) (0.00) (0.03) (0.01)
Excess Kurtosis 0.14 0.48 0.57 2.85 1.26
(p-value for excess kurtosis = 0) (0.32) (0.14) (0.06) (0.00) (0.00)
Jarque-Bera Test 0.21 7.16 14.15 62.8 17.16
(p-value for normality) (0.89) (0.03) (0.00) (0.00) (0.00)
ADF Test for Unit Roots
Levels (constant + trend) -2.24 -2.25 -1.83 -1.39 0.14
First Differences (constant only) -4.59 -3.99 -1.80 -4.64 -4.83
Orders of Significant 1-3 1,2 1-10, 12, 13, 3 1,2
Autocorrelations (at 5% level) 16
Orders of Significant 1,11 1,2 1-5,8,9 3,8,16,31 1,2
Partial Autocorrelations (at 5%
level)
Goldfeld-Quandt Test 24.81 20.1 36.21 14.72 13.67
(p-value for homoskedasticity) (1.00) (1.00) (0.99) (1.00) (1.00)
Table 1.2. (Continued)

LM Test for ARCH 0.01 0.00 24.51 0.01 0.03
(p-value for no ARCH effects) (0.91) (0.95) (0.00) (0.91) (0.85)
Notes:
1. All statistics reported in this Table are for quarterly growth rates defined as 100Δ log yt ,
except the ADF test in levels for which the test statistics are for log levels log y t .
2. p-values for statistics pertaining to skewness, excess kurtosis, Jarque-Bera, Goldfeld-
Quandt, and the LM test for ARCH are reported in parentheses beneath each test
statistic.
3. Optimal lag order for the ADF tests is chosen using the minimum SBC criterion.
Asymptotic critical values at the 10% significance level for the model with constant,
and with constant and time trend, are -2.57 and -3.13, respectively.
4. Optimal lag order for the LM test is chosen using the minimum SBC criterion.
Table 1.3. Summaries of Statistical Inferences

Nonzero Mean Y Y Y Y Y
Skewness N N Y Y N
Fat Tails N N N Y Y
Non-Normality N Y Y Y Y
Unit Roots
Levels (constant + trend) Y Y Y Y Y
First Differences (constant N N Y N N
only)
Homoskedasticity Y Y Y Y Y
ARCH Effects N N Y N N
Notes:
1. ‘ Y ’ means that a given hypothesis is true.
2. ‘ X ’ means that a given hypothesis is false.
Table 2. Neural Network Tests Results
Tests Test Statistics p-values

Neural Network Test 1 124.83 (0.00)
31.36 (0.00)
85.58 (0.00)
Table 2. (Continued)
Tests Test Statistics p-values

35.79 (0.00)
26.31 (0.00)
Neural Network Test 2 342.69 (0.00)
178.35 (0.00)
73.04 (0.00)
268.47 (0.00)
503.72 (0.00)
Notes:
1. Neural network test compares neural network approximations to forecasts from linear
model that is nested in neural network model.
2. Both neural network and linear models are shown in the following Equations:
yt = π ' wt + ut (2a)
where,
~ ' )' ,
u t ~ N (0, σ 2 ), wt = (1, w and ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = (π 0 , π 1 ,........,π p )'

k
yt = π ' wt + ∑θ
j =1
0 j {ψ (γ
'
wt )} + vt (2b)
~ ' )' ,
where, wt = (1, wt
~ = ( y , y , y , ......, y )' , and
wt t −1 t −2 t −3 t −k ψ is a transfer
function.
3. Test statistics for neural network test 1 are presented in column 2 . In this column the
first row denotes test statistic constructed from neural network test 1 for Canada,
second France, third Japan, fourth UK, and fifth USA. The test statistics are
constructed using Equation shown below:
TS = nR 2 (2c)
2
where, n is the number of observation, and R is obtained from regressing regress
û t on X t' and ψ t* ( a matrix of principal components).
4. ‘‘TS” denotes calculated values of ‘Test Statistics” that is distributed χ ( p ) under

2 *
the assumption of normality.

5. p-values for test statistics for Canada, France, Japan, UK, and USA are juxtaposed with
each of the test statistic in parenthesis.
6. The neural network and linear models for neural network test 2 are shown in the
following Equations:
yt = π ' wt + ut (2d)
where,
u t ~ Nid (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = (π 0 , π 1 ,........,π p )'

k
j =1
0 j {ψ (γ j ' wt )} + vt (2e)
where, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π0 is intercept.

7. Column 2 rows 5 to 10 show test statistics for neural network test 2 for Canada,
France, Japan, UK, and USA respectively. The test statistic is constructed using
following Equation:
TS = {( SSE1 − SSE 2) / m} /{SSE 2 /(n − p − m − 1)} (2f)
where in Equation 2 f , m denotes the number of restrictions in the unrestricted
model, n is the number of observations, and p is the number of lags in the regression
equations selected by SBC criterion. The test statistics (TS) is distributed
F [m, (n − p − m − 1)] under normality hypothesis.
8. p-values for test statistics for Canada, France, Japan, UK and USA are juxtaposed with
the test statistics in parentheses.
3. TIME SERIES LINEARITY TESTS
3.1. Keenan Test
Keenan test is a linearity test that was proposed by Keenan (1985). This test is
designed to test if squared forecasts have any additional forecasting power. The
test consists of the following four steps. In the first step of this test yt is regressed
on an intercept and lags ( ys −1 , ys −2 ,..................., ys − M ) of yt to recover residuals
( u t ) for s = M + 1,.........., n . Thereafter, squared predictions ( f t 2 ) are regressed
on an intercept and lags ( y s −1 , y s − 2 ,..................., y s − M ) of yt . Thereafter,
residuals from linear regression ( û ) are regressed on residuals from nonlinear
regression ( v̂ ) for obtaining predictions η̂ and estimates of η 0 ( η̂ 0 ) using
Equation 1.10 .
n
ηˆ = ηˆ0 ( ∑v )
i = M +1
2 1\ 2
t (1.10)
Finally, the test statistics ( TS ) is calculated using SSE1 and η̂ as per

Equation 1.11 .
TS = (ηˆ 2 (2 − 2M − 2)) /( SSE1 − ηˆ 2 ) (1.11)
where, TS is distributed F [1, (n-2M-2)] under the assumptions of normality.
Test results based on Keenan model predictions are presented in Table 3 . In

this Table rows1 to 5 in column 2 show test statistics for Keenan Test for
Canada, France, Japan, UK, and USA real GDP growth rates. p-values for each
test statistic are juxtaposed in parenthesis in the subsequent column.
Table 3. Other Nonlinearity Test Results
Test Test Statistics p-values

Keenan Test 6.78 (0.99)
6.86 (0.98)
6.79 (0.99)
6.79 (0.99)
6.78 (0.99)
Tsay Test 163.60 (0.00)
116.98 (0.00)
163.10 (0.00)
167.47 (0.00)
169.37 (0.00)
RESET 6.00e+29 (0.00)
5.00e+30 (0.00)
1.42e+00 (0.32)
6.00e+24 (0.00)
2.00e+30 (0.00)
RESET1 7.35 (0.01)
0.53 (0.59)
0.00 (1.00)
0.17 (0.95)
8.89 (0.99)
RESET2 19.26 (0.00)
29.41 (0.00)
Test Test Statistics p-values
17.50 (0.00)
174.91 (0.00)
26.61 (0.00)
Notes:
1. Model 1 is the most general nonlinear model that nests linear model to construct
linearity test.
2. All test statistics for each of the series employed for all the tests are shown in column 2
of the Table.
3. In this Table, rows 1 to 5 show test statistics for Keenan Test, 6 to 10 for Tsay Test,
11 to 15 for RESET, 16 to 20 for RESET1, and finally rows 21 to 25 for
RESET2 respectively for Canada, France, Japan, UK, and USA.
4. Test Statistic for Keenan Test is distributed F [1, n-2p-2] under the assumption of
normality which is calculated using Equation 3a .
TS = (ηˆ 2 (2 − 2M − 2)) /( SSE1 − ηˆ 2 ) (3a)
where, ηˆ = ηˆ 0 ( SSE )1 / 2 η̂ = η 0^ (SSE2)1/2. η̂ 0 is the regression coefficient, which is

obtained from regressing û (errors from the linear model) on v̂ (errors from the
nonlinear models).
5. Test statistic for Tsay test is distributed as χ2 with p ( p + 1) / 2 degrees of freedom. The
test statistic for Tsay test is calculated using Equation 3b .
TS = T ( SSE0 − SSE (3b)
where, SSE1 are the residual sum of squares from unrestricted model, SSE 0 from the
restricted model, and T is the number of the observations.
6. Test statistic for RESET is approximately distributed F [k-1, n-k] when Ho is true. This
test statistic is calculated using Equation 3c .
TS = ((SSE0 − SSE1 ) /(k − 1)) /( SSE (3c)
where SSE1 are the residual sum of squares from unrestricted model, and n is the
number of the observations in the model, and k is the number of the parameter
estimated in the model.
7. The test statistic for RESET1 is calculated using Equation 3d .
TS = ((SSE2 − SSE1 ) /( p* )) /( SSE2 /(n − k )) (3d)
where, SSE1 is constructed from the residual sum of squares from unrestricted model
and n is the number of the observations in the model, k is the number of the
parameter estimated in the model, and p * is chosen in such a way that p* < (k − 1) .
The test statistic (TS) is approximately distributed F [p*, n-k] when Ho is true.
8. The test statistic for RESET2 is approximately distributed χ ( p * ) when Ho is true

which can be calculated from the following Equation.
TS = nR 2 (3e)
where, R 2 is obtained from the second part of the test and, n is the total number of
observation in the data series being tested.
9. p-values for each test statistics for Canada, France, Japan, UK and USA are juxtaposed
in the next column in parentheses.
3.2. Tsay Test
Tsay (1986) proposed this test for possible existence of nonlinearities in time
series data, and is also being employed in the present work for testing possible
existence of business cycle asymmetries in all the series. This test is an
improvement over the Keenan test because additional cross terms p( p + 1) / 2 are
added in the nonlinear model. This test can be performed in three steps. In the first
step the contemporaneous variable y t is regressed on an intercept and lags
( y t −1 ,.........., y t − k ) of y t for k = 1,..........., p to recover residuals ( ût ) for
t = 1,......., T . In the second step, residuals ( ût ) obtained from the first step are
regressed on the cross products ( y t − j y t − k ) for k ≥ j , j , k = 1,..........., p to recover
residuals ( v̂t ) for t = 1,......., T . Finally, in the third step, the test statistics (TS) is
computed using Equation 1.12 . The test statistic is distributed χ2 with p( p + 1) / 2
degrees of freedom.
TS = T ( SSE0 − SSE1 ) / SSE0 (1.12)
Test statistics for linearity tests based on Tsay model for all the series are
presented in Table 3 . In this Table rows 6 to 10 in column 2 show test statistics
for Tsay Test for Canada, France, Japan, UK, and USA real GDP growth rates. p-
values for each test statistic are juxtaposed in parentheses in the subsequent
column.
3.3. Ramsey Test RESET
Another test for testing possible existence of business cycle asymmetries due
to Ramsey (1969) is also employed for testing possible existence of business
cycle asymmetries in all the series. This test employs polynomial of the forecast
( Yhat1 = f t ), for an alternate formulation of nonlinear models that nests linear
model. This test can be completed in three steps. In the first step of this test, the
contemporaneous variable y t is regressed on an intercept, and lags
( y t −1 ,.........., y t − k ) of y t for k = 1,..........., p for recovering residuals ( û t ) for
t = 1,......., T . In the second step, y t is regressed on forecasts ( f t ), and
( )
polynomials of forecasts f t , f t1 , f t 2 , f t3 ,.... using Equation 1.13 for any k ≥ 2
to recover residuals ( v̂t ), for t = 1,......., T .
yt = f t + β1 f t 2 + β 2 f t3 + ........ + β k f t k + (1.13)
In the last step of this test, the test statistics (TS) is computed using
Equation 1.14 .
TS = ((SSE0 − SSE1 ) /(k − 1)) /( SSE0 /(n − k )) (1.14)
where, in Equation 1.14 , k is the number of polynomials of forecasts, and n is

the total number of observation in the data series being tested. The test statistic is
approximately distributed F [k-1, n-k] assuming that the null hypothesis of
linearity is true.
Table 3 shows test statistics for linearity tests for all the series that are based
on RESET model predictions. Rows11 to 15 in column 2 show test statistics for
RESET Test for Canada, France, Japan, UK, and USA real GDP growth rates. p-
values for each of the test statistic are juxtaposed in parentheses in the subsequent
column.
3.4. RESET1
This linearity test incorporates principal component analysis of the

polynomial of forecasts ( f t1 ,.............., f t k ) for the largest p * < (k − 1) , and
regressing y t on these principal components and linear term. In the first part of
this test the contemporaneous variable y t is regressed on an intercept, and its lags
[ y t −1 ,.........., y t − k ] for k = 1,..........., p to recover residuals ( ût ) for t = 1,......., T .
In the second step of this test y t is regressed on f t , and principal components of
the polynomials of f t ( f t1 ,.............., f t k ) as per regression Equation 1.15 to
recover residuals ( v̂t ).
yt = f t + β1 f t 2 + β 2 f t3 + ........ + β k (1.15)
Finally, the test statistics are computed using Equation 1.16 .
TS = ((SSE2 − SSE1 ) /( p* )) /( SSE2 /(n − k (1.16)
where, in Equation 1.16 , k is the number of polynomials of forecast, n is the

total number of observation in the data series being tested, and p * is chosen in
such a way that p * < (k − 1) . The test statistic (TS) is approximately distributed
*
F [p , n-k] when Ho is true.
Test statistics for linearity tests based on RESET1 predictions for all the
series are presented in Table 3 , where, rows 16 to 20 in column 2 show test
statistics for RESET1 Test respectively for Canada, France, Japan, UK, and USA
real GDP growth rates. p-values for each test are juxtaposed in parentheses in the
next column.
3.5. RESET2
RESET2 is a different version of the RESET linearity test that also

encompasses principal components of the polynomials of the forecast
( f t1 ,.............., f t k ) from the regression in the first part of the test. This test
consists of the following three-step procedure. In the first step of this test the
contemporaneous variable y t is regressed on an intercept and lags
[y t −1 ,.........., y t −k ] of y t for k = 1,..........., p in order to recover residuals ( ût )
for t = 1,......., T . In the second step, residuals ( ût ) are regressed on an intercept,
and lagged dependent variables [ y t −1 ,.........., y t − k ] of y t for j = 1,..........., p , and

principal components of polynomials of f t using Equation 1.17 for the largest
p * < (k − 1) to recover residuals ( v̂t ) for t = 1,......., T and R 2 , where both k and
p * are arbitrarily selected.
yt = f t + β1 f t 2 + β 2 f t3 +,........,+ β k f t k + vt (1.17)
where, in Equation 1.17 f t = Xˆ t'θ , and finally the test statistic is computed using
Equation 1.18.
TS = nR 2 (1.18)
where R 2 is obtained from the second step of the test, n is the total number of
observations in the data series being tested. The test statistic is approximately
distributed χ 2 ( p * ) assuming that the null hypothesis is true. Test statistic for
linearity test for each of the series based on RESET2 predictions is shown in
Table 3 , where rows 21 to 25 in column 2 show test statistics for RESET2 tests
respectively for Canada, France, Japan, UK, and USA real GDP growth rates. p-
values for each test are juxtaposed in parentheses in the next column.
4. EMPIRICAL RESULTS
4.1. Hypotheses Tests
Two types of neural networks tests are employed to find possible existence of
business cycle asymmetries in real GDP growth rates for Canada, France, Japan,
UK, and USA. Results from neural network linearity tests are compared with the
results for linearity tests that are constructed from other time series models
employed. Comparing linearity tests results that are constructed from forecast
from time series models with that of neural network linearity tests, it transpires
that the neural network approximations appeared to be better for constructing
linearity tests for all the series.
The null hypothesis both for neural network and other linearity tests2 is
linearity, against the alternative hypotheses of nonlinearity. Nonlinearities would
prevail if the null is not true, alternately nonlinearities would not prevail when the
null would be true.
4.2. Empirical Results on Hypotheses Tests
Empirical results from various linearity tests based on neural network tests
are shown in Table 2 . Likewise, empirical results based on linearity test
constructed from nonlinear time series models i.e. Keenan Test, Tsay Test, and
Ramsey Test (RESET), RESET1, and RESET2 are presented in Table 3 . All these
tests are based on test statistics proposed by different researchers, which are
constructed in most cases from the residuals from a linear model, and nonlinear
model that nests the linear model.
Neural networks test for neglected nonlinearities (nnw1) accepts the
alternative hypothesis against the null of linearity for Canada, France, UK and
USA at 5 percent level of significance. Likewise, a neural network test for
possible existence of nonlinearities (nnw2) also accepts the alternative hypothesis
against the null of linearity for Canada, France, UK and USA at 5 percent level of
significance. The inferences do not change when significance level is changed
from 5 to 10 percent.
The Keenan test does not reject the null of linearity for Canada, France,
Japan, UK and USA. Alternately, the Tsay test rejects linearity hypothesis in
Canada, France, Japan, UK and USA real GDP growth rates. Similarly, RESET
rejects the null hypothesis of linearity against alternative of nonlinearities for
Canada, France, UK and USA. However, the test fails to rejects null hypothesis
for Japan only. On the other hands the RESET1 fails to rejects the null hypothesis
of linearity against alternative for France, Japan, UK and USA. Test rejects
linearity hypothesis for Canada at 5 percent level of significance. However, the
RESET2 rejects the null hypothesis against alternative for Canada, France, Japan,
UK and USA. All the inferences do not change when significance level is changed
from 5 to 10 percent.
2
Keenan Test, Tsay Test, and Ramsey Test (RESET), RESET1, and RESET2
5. CONCLUSION
Linearity tests constructed from various time series models i.e. Keenan test
due to Keenan (1985), Tsay test due to Tsay (1986), Ramsay test due to Ramsay
(1969), and some of its improved versions i.e. RESET1 and RESET2 are
employed to test possible existence of business cycle asymmetries in Canada,
France, Japan, UK, and USA real GDP growth rate series. Additionally, neural
network test for neglected nonlinearities (nnw1) due to Lee et al (1993), and
neural network test for possible existence of nonlinearities (nnw2) proposed by
Terasvirta et al. (1993) that are constructed from in-sample approximations from
neural networks are employed to test possible existence of business cycle
asymmetries in Canada, France, Japan, UK, and USA series.
The results show that linearity hypothesis is rejected for linearity tests
constructed from neural network test for neglected nonlinearities (nnw1), and
neural network test for possible existence of nonlinearities (nnw2) for Canada,
France, Japan, UK, and USA real GDP growth rates. Likewise, the null
hypothesis is easily rejected for Tsay and RESET. RESET1 rejects the null
hypothesis in five whereas RESET accepts the null in four out of five series
studies. These results show that both the neural network linearity tests (nnw1 and
nnw2) outperform the linearity hypotheses in all the series when compared to the
linearity test constructed from all the time series models employed. The results on
linearity tests are in line with previous studies including Lee, et al. (1993) and
Terasvirta, et al. (1993).
The results on linearity in the conditional mean for USA are in line with
Bidarkota (2000) and Kiani and Bidarkota (2004). This shows that evidence
against linearity for USA is robust to changes in sample as well as change in
testing approach. Similarly evidence against linearity is also robust for Canada,
and Japan based on testing approach.
Neural Networks models outperformed nonlinear time series models for
testing possible existence of nonlinearities, which is in line with the previous
studies due to Lee, et al. (1993), and Terasvirta, et al. (1993). Compared to Kiani
and Bidarkota (2004), the results on nonlinearity show robust evidence of
neglected nonlinearities in France and UK. Thus, the results show statistically
significant evidence of business cycle asymmetries in Canada, France, Japan, UK,
and USA. However, compared to Kiani and Bidarkota (2004), there exists an
evidence of neglected nonlinearities in France, and UK real GDP growth rates.
6. REFERENCES
Axelrod, R. (1987). The evolution of the strategies in the iterated prisoner’s
dilemma: Algorithm and Simulated Annealing. Morgan Kaufmann, Los Altos:
CA.
Balke, N. & Fomby, T. (1994). Large shocks, small shocks and economic
fluctuations: outliers in macroeconomic time series. Journal of Applied
Econometrics,9, 181-200.
Beaudry, P., & Koop, G. (1993). Do recessions permanently change output?
Journal of Monetary Economics, 31, 149-163.
Bidarkota, P. (1999). Sectoral investigation of asymmetries in the conditional
mean dynamics of the real US GDP. Studies in Nonlinear Dynamics and
Econometrics, 3, 191-200.
Bidarkota, P. (2000). Asymmetries in the conditional mean dynamics of real
GNP: robust evidence. The Review of Economics and Statistics, 82, 153-157.
Blanchard, O., & Watson, M. (1986). Are Business Cycles All Alike? In Gordon,
R.J., (Eds.), The American Business Cycle: Continuity and Change,
University of Chicago Press, Chicago.
Brunner, A. (1992). Conditional asymmetries in real GNP: a semi-nonparametric
approach. Journal of Business & Economics Statistics, 10, 65-72.
Brunner, A. (1997). On the dynamic properties of asymmetric models of real
GNP. The Review of Economics and Statistics, 79, 321-326.
DeLong, J., & Summers, L. (1986). Are business cycles symmetrical? in Gordon
R (ed.), The American Business Cycle: Continuity and Change. Chicago, IL:
University of Chicago Press.
De Jong, K. (1975). An analysis of the behavior of a class of Genitive Adaptive
System. Unpublished Ph.D. dissertation, University of Michigan, Department
of Computer Science.
Dorsey, R., & Walter, J. (1995). Algorithm for Estimation Problems with Multiple
Optima, Nondifferentiability, and other Irregular Features. Journal of
Business Economics and Statistics, 13, 53-66.
Diebold, F., & Rudebusch, G. (1990). A nonparametric investigation of duration
dependence in the American business cycle. Journal of Political Economy,
98, 596-616.
Dunteman, G. (1989). Principal components analysis. Sage: Newbury Park.
Eitrheim, O., & Terasvirta, T. (1996). Testing adequacy of smooth transition
autoregressive models. Journal of Econometrics, 74, 59 -75.
Falk, B. (1986). Further evidence on the asymmetric behavior of economic time

series over the business cycle. Journal of Political Economy, 94, 1096-1109.
French, M., & Sichel, D. (1993). Cyclical patterns in the variance of economic
activity. Journal of Business and Economic Statistics, 11, 113-119.
Granger, C. (1995). Modeling nonlinear relationships between extended-memory
variables. Econometrica, 63, 265-279.
Goldberg, D. (1989). Genetic Algorithm in Search, Optimization and Machine
Learnings. Addison Wisely: Readings, MA.
Hornik, K., Stinchcombe, M., and White, H. (1990). Universal approximation of
an unknown mapping and its derivatives using multilayer feed forward neural
networks. Neural networks, 3, 551-560.
Kang, S. (1991). An introduction of the Use of Feed Forward Neural Networks
for Forecasting. Ph.D. Thesis: Kent State University.
Kastens T., Featherstone, A., and Biere, W. (1995). A neural network premier for
agricultural economists. Agricultural Finance Review, 55, 54-73.
Keenan, D. (1985). A Tucky nonlinear type test for time series nonlinearities.
Biometrika, 72, 39- 44.
Kiani, K.M., & Bidarkota, P. V. (2004). On business cycle asymmetries in G7
countries. Oxford Bulletin of Economics and Statistics, 66, 333-353.
Kiani, K. M. (2005). Detecting business cycle asymmetries using artificial neural
networks and time series models. Computational Economics, 2, 65-89.
Kiani, K. M., & Kastens, T.L. (2008). Testing forecast accuracy of foreign
exchange rates: predictions from feed forward and various recurrent neural
network architectures,” Computational Economics, (forthcoming).
Kydland, F., Prescott, E. (1982). Time to build and aggregate fluctuations.
Econometrica, 50, 1345-70.
Lee, T., White, H., & Granger, C. (1993). Testing for neglected nonlinearity.
Journal of Econometrics, 56, 269-290.
Lippmann, R. (1987). An introduction to computing with neural nets. IEEE ASSP
magazine, April, 4-22.
Long, J., & Plosser, C. (1983). Real business cycles. Journal of Political
Economy, 91, 39-69.
Lucas, R. (1976). Econometric policy evaluation: a critique. Carnegie-Rochester
Conference Series in Public Policy, 1, 19-46.
Luukkonen, R., Saikkonen, P., & Terasvirta, T. (1988). Testing linearity against
smooth transition autoregressive models. Biometrika, 75, 491- 499.
Marimon, R., McGratten, E., & Sargent, T. (1990). Money as medium of
exchange in an economy with artificially intelligent agents. Journal of
Economics Dynamics and Control, 14, 329- 373.
Neftci, S. (1984). Are Economic time series asymmetric over the business cycle?
Journal of Political Economy, 92, 307-328.
Potter S. (1995). A non-linear approach to U.S. GNP. Journal of Applied
Ramsey, J. (1969). Tests for specification errors in classical linear least-squares
regression analysis. Journal of the Royal Statistical Society, 31, 350-371.
Ramsey, J., & Rothman, P. (1996). Time irreversibility and business cycle
asymmetry. Journal of Money Credit and Banking, 28, 1-21.
Reilly, D., & Cooper, L. (1990). An overview of neural networks: early models to
real world systems. In Zornetzer, S. F., Davis, J. L., & Lau, C. (Eds.), An
Introduction to Neural and Electronic Networks, Academic Press, New York.
Scheinkman, J., and LeBaron, B. (1989). Non-linear dynamics and GNP data. In
Barnett, W. et al. (Eds.), Economic Complexity: Chaos, Sunspots, Bubbles,
and Non-linearity, Cambridge University Press, Cambridge.
Schumpeter, J. (1939). Business cycles: a theoretical, historical, and statistical
analysis of the capitalist process. McGraw-Hill:New York.
Sichel, D. (1989). Are business cycles asymmetric? a correction. Journal of
Political Economy, 97, 1255-1260.
Sims, K. (1980). Macroeconomic reality. Econometric, 48, 1-49.
Tang, Z., & Fishwick, A. (1993). Feed forward neural networks as models for
time series forecasting. ORSA Journal of Computing, 5, 374-385.
Terasvirta, T., Lin, C., & Granger, C. (1993). Power of neural network test.
Journal of Time Series Analysis, 14, 209-220.
Terasvirta, T., & Anderson, H. (1992). Characterizing nonlinearities in business
cycles using smooth transition autoregressive models. Journal of Applied
Econometrics, 7, S119-S136.
Tsay, R. (1986) Non-linearity tests for time series. Biometrika, 73, 461-466.
Tsay, R. (1988). Nonlinear time series analysis of blowfly population. Journal of
Time Series Analysis, 9, 247-263.
White, H. (1989). Learning in artificial neural networks: a statistical perspective.
Neural Computations, 1, 425-464.
Wong, F.W. (1991). Time series forecasting using back propagation using neural
networks. Neurocomputing, 2, 147-149.
Zhang, G., Patuwo, B., & Hu, M. (1998). Forecasting with artificial neural
networks: the state of arts. International Journal of Forecasting, 14, 35-62.
Chapter 2
NEURAL NETWORK FORECASTS

EVALUATIONS AND BUSINESS CYCLE
FLUCTUATIONS
ABSTRACT
The present work employs in-sample forecasts from linear models in
conjunction with neural network approximations to construct neural network
linearity tests for testing possible existence of business cycle asymmetries in
Canada, France, Japan, UK, and USA, real GDP growth rates. Additionally,
neural network linearity tests are constructed from jackknife out-of-sample
forecasts from linear models as well as neural networks approximations for
testing possible existence of business cycle asymmetries in Canada, France,
Japan, UK, and USA real GDP growth rates. Finally in-sample as well as
jackknife out-of-sample forecasts from linear models are compared with
those of neural network approximations using a pair-wise test of forecast
evaluation due to Ashley, Granger, and Schamalensee (1980) to determine
suitable forecasting model for a particular situation.
These results show that statistically significant evidence of business
cycle asymmetries does exist in all the series. Moreover, in-sample forecast
performance of neural networks is superior to linear models, however, when
using out-of-sample forecasts linear models would be a better choice.
Keywords: AGS test; asymmetries; business cycles; in-sample forecasts; neural

network tests; out-of-sample forecasts
JEL codes: C32, C45, E37
1. INTRODUCTION
Most studies on business cycle research are focused on univariate nonlinear
econometric modeling. Using univariate nonlinear time series models Neftci
(1984), Brunner (1992, 1997), Beaudry and Koop (1993), Potter (1995), and
Ramsey and Rothman (1996) concluded that business cycles are characterized by
asymmetric fluctuations. However, Falk (1986), Sichel (1989), Delong and
Summers (1989), and Diebold and Rudebusch (1990), either found a weak
evidence of asymmetries in business cycles fluctuations or failed to conclude that
business cycles are asymmetric.
Studies by Bidarkota (1999, 2000) provided strong evidence of asymmetries
in USA macroeconomic time series data. However, when Kiani and Bidarkota
(2004) conducted similar analyses on Canada, France, Germany, Italy, Japan, UK,
and USA, they found strong evidence of asymmetries in real GDP growth rates
for Canada, Germany, Italy, Japan, and USA, but they were not able to find
asymmetries in France and UK real GDP growth rates. Therefore, the basic
research question whether business cycle asymmetries in all the countries of the
world particularly in G7 countries are alike remained unanswered. Therefore,
using artificial neural networks (ANN), the present work focuses on two issues.
One, detecting business cycle asymmetries in Canada, Germany, Japan, and USA,
via in-sample testing as well as jackknife out-of-sample testing, and two, testing
in-sample as well as jackknife out-of –sample forecasts performance of linear
models versus artificial neural networks.
A number of nonlinear time series models were developed in the recent past,
this includes models due to Auerbach (1982), Beaudry and Koop (1993), Gordon
(1986), Kling (1987), Koch and Rasche (1988), Diebold and Rudebusch (1989),
Hamilton (1989), and Estrella and Mishkin (1998). However, because of ample
need for additional nonlinear time series models in this area, Kuan and White
(1994) including others specifically discussed neural networks and their
applications in economics.
Neural networks have been used successfully in science, engineering,
medical, business and economics. Although neural networks have demonstrated
some success in financial and economics applications, only a few studies, for
example Vishwakarma (1995), and Qi (2001) have focused on business cycles.
Therefore a study of business cycles might benefit from additional nonlinear
models, especially neural networks. In particular, the present work uses neural
network models to find possible existence of business cycle asymmetries in
Canada, France, Japan, UK, and USA real GDP growth rates, using in-sample as
Neural Networks Forecasts Evaluations … 29
well as jackknife out-of-sample forecasts from linear models and artificial neural
networks approximations. Moreover, the present work also seeks to test in-sample
as well as jackknife out-of-sample forecast performance of neural networks versus
linear models for all the series.
ANN are able to estimate any continuous function with a desired level of
precision (Hornik et al. (1989)). The ability of information processing makes
ANNs powerful computational devices that can learn from examples and
generalize these learning to solve problems that are never seen before (Reilly and
Cooper (1990)). Performance of ANNs in biology, engineering, and finance is
well known, and this should not be surprising given that ANNs are typically
highly flexible nonlinear models that can easily fit any data series. On the other
hand in terms of generalizing important predictive relationship, ANNs might
actually tend to over fit data, finding relationships where there really are none.
Thus, studies using ANNs would benefit from more out-of-sample testing. Then,
ANN superiority would be less of foregone conclusion.
Using a test of pair-wise forecast evaluation due to Ashley, Granger, and
Schamalensee (1980), henceforth referred to as the AGS test, the present work
employs in-sample forecasts from linear models, and neural networks. Thereafter,
in-sample and jackknife out-of-sample forecasts are employed to test possible
existence of business cycle asymmetries in Canada, France, Japan, UK, and USA
real GDP growth rates. Finally, forecast performance of neural networks is tested
via in-sample as well as jackknife out-of-sample forecast from linear models
versus approximations from neural networks.
The remaining paper is organized as follows. Section 2 incorporates
description on neural networks, estimation issues, neural network linearity tests,
jackknife re-sampling, AGS test, data sources and preliminary data analysis,
whereas section 3 presents empirical results, and forecast performance of neural
networks. Finally, section 4 presents conclusion.
2. NEURAL NETWORKS
An artificial neural network (ANN) is said to be an advanced artificial
technology that mimics human-brain's learning, and decision-making process.
ANN can learn through trial and experience, and use these learnings to solve
problems never seen before. The ability of information processing makes ANNs
powerful computational devices that can learn from examples, and generalize
learning to solve the problems (Reilly and Cooper (1990).
ANN are treated as nonlinear, nonparametric statistical methods that are

independent of the distributions of the underlying data generating processes
(White 1989). Therefore, the ANN modeling approach is useful for forecasters
and researchers, especially in problems where data are available but the data
generating processes as well the underlying laws for data generating processes are
unknown. The general form of a neural network model employed in the present
work is presented in equation 2.1 .
⎡ n ⎛ k ⎞⎤
f (x ) = sig ⎢α 0 +
⎢⎣
∑α j sig ⎜ ∑
⎜ β ij xi + β 0 j ⎟⎥ + ε
⎟⎥
(2.1)
j =1 ⎝ i =1 ⎠⎦
where, n is the number of the hidden nodes in the network, k is the number of
explanatory variables in the network, sig (x) = 1/(1+e-x), α j represents a vector of
parameters or weights linking the hidden to the output layers’ units, β ij (i = 1, …
, k; j = 1 , … , n) denotes a matrix of parameters linking the input to the hidden
layers’ units, and ε is the error term. The error term ε can be made arbitrarily
small if a sufficient number of explanatory variables is included, and if n is
chosen to be large enough. However, the present work considers up to only 2
explanatory variables (k = 1, 2), so that n is fixed to be 2. The parameters in
Equation 2.1 are chosen to minimize the sum of squared errors ( ∑ε 2
) , referred
to as SSE.
2.1. Neural Network Linearity Tests
Neural network linearity test tests the null hypothesis of linearity against the
alternative hypothesis of nonlinearity. The test is employed to test possible
existence of business cycle asymmetries in Canada, France, Japan, UK, and USA
real GDP growth rates. The test statistic for this test is constructed using forecasts
from a linear model of the form of the Equation 2.2 , and approximations from an
unrestricted nonlinear model i.e. a neural network model of the form shown in
Equation 2.3 .
yt = π ' wt + ut (2.2)
where,
u t ~ N (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = ( π 0 , π 1 ........, π p )'

k
j =1
0 j {ψ (γ j ' wt )} + vt (2.3)
where
u t ~ Nid (0, σ 2 ), ~ ' )' ,

wt = (1, w ~ = ( y ,..., y )' ,
w
t t t −1 t− p
and π = (π 0 , π 1 ,........, π p )' ,
ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 , and π 0 is intercept. Equation 2.3 shows a

nonlinear neural network model that nests the linear model (Equation 2.2 ). The
test consists of three-step procedure. In the first step real GDP growth rates ( y t )
are regressed on an intercept and lags ( y t −1 ,.......... y t − k ) of yt to recover residual
( û t ) and squared sum of errors ( SSE1 ). Likewise, ANN is approximated using
Equation 2.3 to recover residuals ( v̂t ), and squared sum of errors ( SSE 2 ). Finally,
the test statistic is computed using Equation 2.4 , which is distributed
F [m, (n − p − m − 1)] under normality hypothesis.
TS = {( SSE1 − SSE 2) / m} /{SSE 2 /(n − (2.4)
where, in Equation 2.4 , m denotes the number of restrictions in the unrestricted

model, n is the number of observations, and p is the number of lags.
Due to relatively large number of parameters and the nonlinearity inherent in

the ANN specification, the objective function is unlikely to be globally convex,
and thus can have many local minima. Thus, for the required SSE minimization
we started by using a genetic algorithm (GA), which often is considered among
the most reliable (does not get “hung up”) algorithms for estimating highly
nonlinear functional forms, but which typically is much slower than others. To
further increase the probability of obtaining a global minimum, 4 independent

runs of the GA were considered, each with a fixed number of iterations. The
parameter vector (of the 4) that had the smallest SSE was used as starting
conditions for Matlab’s fminsearch algorithm, which is a Nelder-Mead simplex
algorithm that worked well for closing in on the final optimum.
2.3. Jackknife Out-of-sample Forecasts
Researchers often use jackknife re-sampling techniques when the distribution

of the parameters under review is either unknown, when it cannot be characterized
by a mathematical function, or when the mathematical function is especially
difficult to estimate. Quenouille (1949) used this technique to reduce bias in
estimators, and thereafter Tuckey (1958) employed it for estimating variances.
The standard jackknife estimate is calculated deleting one observation and
estimating parameters using n − 1 observations of the data. The model and
estimated observations are than used to conditionally predict the dependent value
for the deleted observation. This process continues, detecting and predicting a
different observation for each model, until each of the observations in the data has
been predicted out-of-sample.
Sub-sample jackknife technique was initially proposed by Wu (1990), and
extended by Politis and Romano (1994). Politis et al. (1997) and Ziari el.al (1997)
also used these re-sampling techniques for their empirical work. Compared to
standard jackknife, in sub-sample jackknife more than one observation is dropped
to estimate out-of-sample forecasts of the remaining m = n − d observations,
where n is the total number of observations, and d = 2, 3,........., n − 1 . The model
and the estimated observations are than used to conditionally predict the
dependent value for the deleted observation, and this process continues until each
observation in the data is predicted or until all possible sub-samples are
considered, depending on the statistical assumptions and the application.
2.4. AGS Test
The test of pair-wise forecast evaluation by Ashley, Granger, and

Schmalensee (1980), henceforth called as AGS test provides a test for statistical
significance of the difference between root mean square errors of the two
competing forecasts. The procedure for the AGS test is also described in Bessler
and Brandt (1992), Bradshaw and Orden (1990), and Kastens, & Brester (1996).
The AGS test statistics are obtained by estimating the regression shown in
Equation 2.5 . However, if the sample mean of the forecast error from either
model is negative, that forecast error series must be multiplied by –1 before
estimating the regressions.
dt = β1 + β 2 ( st − smean) + et (2.5)
where,
d t = difference between forecast errors (lower RMSE forecasts are subtracted

from higher RMSE forecasts).
st = sum of forecast errors
smean = sample mean of s t
et = white noise
The test statistic for the AGS test is calculated from the residuals obtained
from estimating an unrestricted model represented by Equation 2.5 , and its
restricted version. The restricted model is obtained restricting β1 = β 2 = 0 in
Equation 2.5 . The test statistic is distributed F but the comparable value of this
test statistic is one fourth of the tabulated value of this test statistic because it does
not take into account the sign of the coefficient estimates. The test statistic is
calculated using Equation 2.6 given below.
TS = ((SSER − SSEUR ) /(k − 1)) /( SSEUR /(n (2.6)
where,
SSE R = Sum of squared residuals from restricted model

SSEUR = Sum of squared residuals from un-restricted model
n = Number of observations in the sample
k = Number of variables in the regression model
2.5. Data Sources
Quarterly data on real GDP growth rates series for Canada, France, Japan,
UK and USA is obtained from the International Financial Statistic (IFS)’s CD-
ROM for the month of September 2001. Canada, Japan, UK and USA data range
from 1957:1 to 2000:4 whereas the France data range from 1970:1 to 2000:4.
Table 1. Data Description

Data Series Quarterly Quarterly Quarterly Quarterly Quarterly
Real GDP Real GDP Nominal GDP Real GDP Real GDP
Sample Period 1957:1- 1970:1- 1957:1- 1957:1- 1957:1-
2000:4 2000:4 2000:4 2000:4 2000:4
Sample Length 176 124 176 176 176
Notes:
1. The quarterly seasonally adjusted real GDP growth rates for Canada, France, UK, and
USA were obtained from the September 2001 edition of the International Financial
Statistics (IFS) CD-ROM.
2. The nominal GDP series are employed for Japan because seasonally adjusted data were
only available for the nominal series and not for the real series on the IFS CD-ROM.
Table 2.1. In-sample Performance of Neural Network vs. Linear Models
LIN ANN2 ANN5 ANN10

STD RMSE RMSE RMSE RMSE
Canada 0.126 0.123 0.121 0.121 0.120
France 0.070 0.070 0.066 0.062 0.062
Japan 0.226 0.228 0.170 0.170 0.152
UK 0.139 0.134 0.134 0.127 0.123
USA 0.119 0.119 0.118 0.118 0.116
Notes:
1. The results presented in this Table are based on in-sample forecasts from linear models
and Jackknife out-of-sample approximations from artificial neural networks.
2. In this Table ‘STD’ represents standard deviation of the series, ‘Lin” represents linear
model, ANN2 represents, artificial neural network with two hidden nodes, ANN5,
artificial neural network with five hidden nodes, and ANN10 represents an artificial
neural network models with ten hidden nodes.
3. This Table presents, standard deviation, in-sample forecasts from linear models and in-
sample approximations from artificial neural networks (ANN), and root mean squared
errors (RMSE) for linear models ANN2, ANN5, and ANN10 for Canada, France,
Japan, UK, and USA series.
4. Column 2 shows standard deviation (STD), column 3 shows RMSE for linear models,
column4 RMSE for ANN2, column5 RMSE for ANN5, and column5 shows RMSE
for ANN10 model for each of the series.
The analysis started with the data on real GDP growth rates for G7 countries
but there was a jump of 35.43 percent in Germany GDP in 1991:1. This likely
would have been because of its reunification. Italy showed an inexplicable spike
of 87.5 percent in 1970:1. Therefore, series for Italy and Germany were excluded
from the analysis. Table 1 presents additional information on Canada, France,
Japan, UK, and USA real GDP growth rate series.
Table 2.2. Out-of-sample Performances of Neural Network vs. Linear Models
LIN ANN2 ANN5 ANN10

STD RMSE RMSE RMSE RMSE
Canada 0.126 0.125 0.257 0.287 0.335
France 0.070 0.070 0.168 0.179 0.195
Japan 0.226 0.223 0.586 0.688 0.718
UK 0.139 0.134 0.344 0.349 0.434
USA 0.118 0.118 0.279 0.305 0.377
Notes:
1. See notes on table 2.1.
2. Results presented in this Table are based on Jackknife out-of-sample forecasts from
linear models and jackknife out-of-sample approximations from artificial neural
network for Canada, France, Japan, UK, and USA real GDP growth rates.
In-sample forecasts from linear model and neural network approximations for
Canada, France, Japan, UK, and USA real GDP growth rates are presented in
Table 2.1 . This Table shows root mean squared errors for linear model estimates
(RMSE1) as well as for neural network model approximations (RMSE2) with two
(ANN2), five (ANN5), and ten (ANN10) hidden nodes. Similarly, Table 3.1
shows mean of the residuals from the linear models (M1), and neural network
models (M2), and root mean squared errors from linear (RMSE1), as well as
neural network models (RMSE2).
In addition to in-sample forecasts/approximations from linear models and

ANN, the present work employs jackknife re-sampling technique for estimating
out-of-sample forecasts from linear models for all the series. Root mean errors
calculated from linear modes for all the series are shown in Table 2.2. Finally, the
Table also show root mean square errors from jackknife out-of-sample
approximations from neural networks with two (ANN2), five (ANN5), and ten
(ANN10) hidden nodes. Likewise, using out-of-sample forecasts, mean of the
residuals from linear models (M1), and neural network models (M2) as well as
root mean square errors from linear models (RMSE1), and neural networks
(RMSE2) are presented in Table 3.2 for all the series.
Table 3.1. In-sample Forecast Results: Neural Network vs. Linear Models
M1 M2 RMSE1 RMSE2
Canada 0.000 0.000 0.123 0.121
France 0.000 0.000 0.070 0.066
Japan 0.000 0.000 0.224 0.170
UK 0.000 0.000 0.134 0.134
USA 0.000 0.000 0.119 0.118
Notes:
1. The results presented in this Table are based on in-sample forecasts from linear models
and Jackknife out-of-sample approximations from artificial neural networks.
2. Row 1 shows M1, M2, RMSE1, RMSE2 for Canada real GDP growth rates which are
respectively the mean of residuals series recovered from linear model, mean of
residuals from NNW model, RMSE from linear model and RMSE from NNW model
residuals.
3. Row 2 shows M1, M2, RMSE1, and RMSE2 for France real GDP growth rates, row 3
for Japan, row 4 for UK, and row 5 for USA residual series.
4. These statistics are the basis for the hypothesis if an artificial neural network is an
improvement over the linear models.
Table 3.2. Out-of-sample Forecast Results: Neural Network vs. Linear

Models
M1 M2 RMSE1 RMSE2
Canada -1.80e-4 -4.00e-4 0.125 0.257
France -1.80e-4 -3.00e-4 0.069 0.168
Japan -1.50e-4 -0.003 0.223 0.586
UK -5.00e-5 0.004 0.134 0.344
USA -7.00e-5 -0.003 0.118 0.279
Notes:
1. See notes on Table 3.1.
2. Results presented in this Table are based on Jackknife out-of-sample forecasts from
linear models and jackknife out-of-sample approximations from artificial neural
network for Canada, France, Japan, UK, and USA real GDP growth rates.
Table 4. Neural Network Nonlinearity Tests
In-sample Forecasts Jackknife Out-of-sample Forecasts

RMSE Test RMSE Test Statistics
Statistics
Canada 0.121 274.235 0.169 123.339
(0.000) (0.000)
France 0.069 270.082 0.096 121.633
(0.000) (0.000)
Japan 0.211 184.470 0.362 44.967
(0.000) (0.000)
UK 0.132 400.694 0.156 268.614
(0.000) (0.000)
USA 0.117 507.968 0.150 267.132
(0.000) (0.000)
Notes:
1. The results presented in this Table are based on in-sample as well as jackknife out-of-
sample forecasts from linear models and in-sample and jackknife out-of-sample
approximations from artificial neural networks.
2. Column 1 in this Table show names of the countries included in the study.
3. Test statistics presented in the Table are based on linear as well as neural network
models represented primarily by the following two Equations:
y t = π ' wt + u t (4.1a)
where,
u t ~ N (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = ( π 0 , π 1 ........, π p )'

k
yt = π ' wt + ∑θ
j =1
0 j {ψ (γ j ' wt )} + vt (4.1b)
4. The test statistics are F distributed with m numerator and (n − p − m − 1)

denominator degrees of freedom under normality hypothesis. The test statistic for this
distribution is calculated using following Equation:
TS = {( SSE1 − SSE 2) / m} /{SSE 2 /(n − p − m − 1)} (4.1c)
where, in Equation 4.1c , m denotes the number of restrictions in the unrestricted

model, n is the number of observations, and p is number of lags both in the
restricted (linear) and unrestricted (neural network) models.
5. Column 3 shows test statistics constructed from in-sample forecasts whereas column 6
shows test statistics constructed from jackknife out-of-sample forecasts. p-values for
each test statistic are shown beneath each test statistic in parenthesis.
6. In column 2 , we present root mean squared error (RMSE) for in-sample forecasts
whereas in column 5 , we show RMSE calculated from jackknife out-of-sample
forecasts.
3.1. Results on Linearity Tests
The present work tests the null hypothesis of linearity versus the alternative
hypothesis of nonlinearity. This hypothesis is tested on Canada, France, Japan,
UK, and USA real GDP growth rates based on nonlinearity test statistic
constructed from in-sample forecast approximated from neural network in
conjunction with forecasts from its linear counterparts. These hypotheses are
repeated using jackknife out-of sample forecasts approximated from neural
network models, and jackknife out-of-sample forecasts from linear models.
If null hypothesis of linearity is true, nonlinearities do not exist, however, if
the null is false, nonlinearities do exist in the data series being tested. Under
linearity hypothesis, the test statistic is distributed F [m, (n-p-m-1)], where, m is
the number of restrictions in the nonlinear model, n is the total number of
observations in the series being tested, and p is the numbers of lags used in each
regression.
Table 4 show results from in-sample as well as jackknife out-of-sample
forecast from linear model. In this Table column 2 rows 1 − 5 show in-sample
RMSE respectively for Canada, France, Japan, UK, and USA. The test statistic
constructed from in-sample forecasts are shown in column 3 rows 1 − 5
respectively for Canada, France, Japan, UK, and USA real GDP growth rates.
Likewise, RMSE as well as the test statistics obtained from the jackknife out-of-
sample forecast for all the series are shown respectively in column 4 and 5 .
The results based on in-sample forecasts from linear model as well as
jackknife out-of-sample approximations show that the hypothesis of linearity is
rejected in Canada, France, Japan, UK, and USA real GDP growth rates.
Likewise, neural network linearity test constructed from jackknife out-of-sample
forecast, overwhelmingly rejects linearity hypothesis in all the series. However,
test statistics for the tests based on jackknife out-of-sample forecasts though
significant have lower magnitude than that of the linearity tests that are
constructed from in-sample forecasts.
3.2. Forecast Performance of Neural Networks
The AGS test that is a test of pair-wise forecast evaluation between any two
competing models is employed for testing forecast performance of ANN versus
linear models for all the series. For doing so, in the present work, in-sample as
well as jackknife out-of-sample forecast from linear models as well as in-sample
and jackknife out-of-sample approximations from neural networks are employed.
If the coefficients β 1 and β 2 shown in Equation 2.5 are both positive, the
appropriate test is an F-test of the joint hypothesis that β 1 = β 2 = 0 . Because the
F-test does not consider the sign of the coefficient estimates, the actual
significance levels are only one fourth of that is reported in a standard F-
distribution Table. That means that the probability of obtaining F-Statistic is
greater than the critical value and having both estimates positive is equal to one
fourth of the significance level normally associated with the critical value.
If one of the estimated coefficients is significantly negative then one cannot
conclude that the neural network model is a significant improvement over the
linear model. However, if one of the estimated coefficients is negatively
insignificant then a 1 − tail t-test is used to judge the significance of the positive
coefficient. Contrary to that, if the null hypothesis is not accepted against the
alternative that at least one of the coefficients is significant then one might
conclude that the neural networks model is an improvement over the linear model.
The null hypothesis for in-sample forecast performance is that the neural
networks model is not a significant improvement over the linear model (RMSEs
of the competing forecasts are equal) versus the alternative hypothesis that the
neural networks model is an improvement over the linear model. If the null is
rejected, the neural networks model forecast performance is superior to that of the
linear model. However, if the null is accepted, then the two model forecasts
cannot be distinguished, meaning that the evidence does not suggest that the
neural network model is better than the linear model.
The results from in-sample AGS test statistics that are presented in Table 5.1
for all the series reveal that neural network models perform better than linear
models when in-sample forecasts are employed. However, the results based on
jackknife out-of-sample forecasts shown in Table 5.2 demonstrate that linear
model performance is superior to neural network models. Therefore, linear models

appear to offer better forecasts when out-of-sample forecast are to be worked
with. While neural network forecasts are superior in an in-sample framework, it
might not be the only reason for being chosen for out-of-sample forecasts.
Table 5.1. AGS TEST: In-sample Performance of Linear vs. ANN Model
Parameters Canada France Japan UK USA

β1 2.00e-6 1.70e-5 3.88e-4 7.60e-5 4.00e-6
(0.999) (0.929) (0.621) (0.691) (0.960)
β2 0.807 0.032 0.154 0.015 0.003
(0.323) (0.048) (0.000) (0.114) (0.395)
Test Statistic 0.492 1.946 13.224 1.329 0.365
(0.419) (0.466) (0.419) (0.419) (0.419)
Notes:
1. The results presented in this Table are based on in-sample forecasts from linear models,
and in-sample approximations from artificial neural networks.
2. In AGS test is the test of the joint significance of the parameter estimates of β 1 and
β 2 i.e. β 1 = β 2 = 0 is tested to conclude that the neural network models are
improvement over the linear models. The AGS test is contemplated using the model
shown in Equation 5.1 a :
dt = β1 + β 2 b + et (5.1a)
3. β1 is the coefficient on the intercept term, and β 2 is the slope coefficient of the AGS
regression model.
4. p-values for parameter estimates of the coefficients of intercept ( β 1 ) , and the slope
coefficients for each of the series are shown in parentheses below each parameter
estimate for all the series.
5. Test statistics for AGS tests are presented in last row of the Table, and the relevant p-
values for the joint significance of the parameter estimates are presented below each
statistic in parentheses.
Table 5.2. AGS TEST: Out-of-sample Performance of Linear vs. ANN

Models

β1 2.00e-4 -4.00e-4 0.003 0.004 0.003
(0.852) (0.694) (0.182) (0.024) (0.001)
Table 5.2. (Continued)

β2 0.565 0.616 0.604 1.033 0.521
(0.000) (0.000) (0.000) (0.000) (0.000)
Test Stat 32.862 30.458 49.151 49.572 46.111
(0.000) (0.000) (0.000) (0.000) (0.000)
Notes:
1. See note on Table 5.1
2. The results presented on this Table are based on jackknife out-of-sample forecasts from
linear models, and jackknife out-of-sample approximation from artificial neural
network.
4. CONCLUSION
The present work employs artificial neural networks (ANN) as well as linear
models to compare in-sample, and jackknife out-of-sample approximations from
neural networks and forecasts from linear models. Neural network linearity test is
constructed from in-sample forecasts/approximations from linear models and
ANN that nests the linear model for testing possible existence of nonlinearities in
Canada, France, Japan, UK, and USA real GDP growth rates. Likewise, neural
network linearity tests are constructed based on jackknife out-of-sample forecasts
from linear models and neural networks for testing possible existence of business
cycle asymmetries in Canada, France, Japan, UK, and USA real GDP growth
rates.
For testing pair-wise forecast evaluation between two competing models the
present work employs a test of forecast evaluation due to Ashley, Granger and
Schmalensee (1980) using in-sample as well as jackknife out-of-sample forecasts
from the two competing models i.e. linear model and ANN for Canada, France,
Japan, UK, and USA real GDP growth rates.
The results show that in-sample forecast performance of ANN is superior to
that of linear models. These results are in line with the previous studies due to
Lee, et al. (1993), and Terasvirta, et al. (1993). However, out-of-sample
performance of linear models is better than neural networks models for real GDP
growth rates for all series. This means that a superior in-sample forecast
performance of neural networks models should not be the only reason for
employing them for out-of-sample forecasting.
The results based on in-sample forecasts provide a robust evidence of

asymmetries in Canada, France, Japan, UK, and USA. In addition, the results
based on jackknife out-of-sample forecast strengthen the evidence for existence of
business cycle asymmetries in Canada, France, Japan, UK, and USA real GDP
growth rates. Therefore, the policymakers might not able to evaluate the impact of
monetary policy or any other shock on real GDP series in these countries based on
forecast from linear models.
The results on nonlinearity for USA are in line with Bidarkota (2000) and
Andreano and Savio (2002). Similarly the results on nonlinearity for Canada,
Japan, and USA are in line with Kiani and Bidarkota (2004). This shows that
evidence against linearity for USA is robust to change in samples as well as
change in testing approach. Neural network models outperform the traditional
statistical tests for remaining nonlinearities, which is in line with the previous
studies (Terasvirta, et al. 1993).
5. REFERENCES
Andreano, M., and G. Savio, (2002). Further evidence on business cycle
asymmetries in G7 countries, Applied Economics, 34, 895-904.
Ashley, R, Granger, C.W.J. & Schmalensee, R. (1980). Advertising and aggregate
consumption: an analysis of causality, Econometrica, 48, 1149-1167.
Auerbach, A. J., 1982, The index of leading indicators: measurement without
theory, thirty-five year later, Review of Economics and Statistics, 64, 589-
595.
Beaudry, P. & Koop, G. (1993). Do recessions permanently change output?
Bessler, D.A.& Brandt, J.A. (1992). An analysis of forecasts of livestock prices,
Journal of Economic Behavior and Organization, 18, 249-63.
Bidarkota, P.V. (1999). Sectoral investigation of asymmetries in the conditional
mean dynamics of the real US GDP, Studies in Nonlinear Dynamics and
Bidarkota, P.V. (2000). Asymmetries in the conditional mean dynamics of real
GNP: robust evidence, The Review of Economics and Statistics, 82, 153-157.
Bradshaw, G.W. & Orden, D. (1990). Granger causality from the exchange rates
to agricultural prices and export sales, W. J. Agricultural Economics, July,
100-110.
Brunner, A.D. (1992). Conditional asymmetries in real GNP: a semi-

nonparametric approach, Journal of Business & Economics Statistics, 10, 65-
72.
GNP, The Review of Economics and Statistics, 79, 321-326.
DeLong, J.B. & Summers, L.H. (1986). Are business cycles symmetrical? In R.J.
Gordon (Eds.), The American Business Cycle: Continuity and Change.
Chicago, IL: University of Chicago Press, pp. 166-179.
Diebold, F.X. Rudebusch, G.D. (1990). A nonparametric investigation of duration
dependence in the American business cycle, Journal of Political Economy,
98, 596-616.
Estrella, A., and Mishkin, F. S., (1998). Predicting U.S. business cycle regimes,
Journal of Business and Economic Statistics, 17, 313-323.
series over the business cycle, Journal of Political Economy, 94, 1096-1109.
Gordon, R. J., (1986). The American Business Cycle : continuity and change,
NBER studies in business cycles, Vol. No. 25, University of Chicago Press,
Chicago.
Hamilton, J.D., (1989). A new approach to the economic analysis of nonstationary
time series and business cycle, Econometrica, 57, 357-384.
Hornik, K., Stinchcombe, M. & White, H. (1989). Multilayer feed forward neural
networks are universal approximations, Neuralnetworeks, 2, 359-356.
Kastens, T. L., & Brester, G. W. (1996). Model selection and forecasting ability
of theory-constrained food demand system, American Journal of Agricultural
Economics, 78, 301-312.
Kiani, K. M., & Bidarkota, P.V. (2004). On business cycle asymmetries in ‘G7”
countries,” Oxford Bulletin of Economics and Statistics, 66, 333-353.
Kling, J.L., (1987). Predicting the turning points of business and economic time
series, Journal of Business, 60, 201-238.
Koch, P.D. & Rasche, R. H., (1988). An examination of the commerce department
leading indicator approach, Journal of Business and Economic Statistics, 6(2),
167-187.
Kuan, C., and White, H., (1994). Artificial neural networks: an econometric
perspective. Econometric Review, 13, 1–91.
Lee, T.H., White, H. & Granger, C.W.J. (1993). Testing for neglected
nonlinearity, Journal of Econometrics, 56, 269-290.
Neftci, S.N. (1984). Are economic time series asymmetric over the business
cycle? Journal of Political Economy, 92, 307-328.
Politis, D.N., & Romano, J.P. (1994). Large sample confidence region based on
sub-samples under minimal assumptions, Annals of Statistics, 22, 2031-2052.
Politis, D.N., Romano, J.P.& Wolf, M. (1997). Subsampling for heteroskedastic
time series, Journal of Econometrics, 81, 281-317.
Potter, S.M. (1995). A non-linear approach to U.S. GNP, Journal of Applied
Qi, M., (2001). Predicting US recessions via leading indicators via neural network
models, International Journal of Forecasting, 17, 383-401.
Quenouille, M.H. (1949). A note on bias in estimation, Biometrika, 43, 353-60.
Ramsey, J.B. & Rothman, P. (1996). Time irreversibility and business cycle
asymmetry, Journal of Money Credit and Banking, 28, 21.
Reilly, D.L., & Cooper, L.N. (1990). An overview of neural networks: early
models to real world systems. In S.F. Zornetzer, J.L. Davis, and C. Lau
(Eds.), An Introduction to Neural and Electronic Networks, New York:
Academic Press, pp. 227-248.
Sichel, D.E. (1989). Are business cycles asymmetric? a correction, Journal of
Terasvirta, T., Lin, C.F. & Granger, C.W.J. (1993). Power of neural network test,
Tuckey, J.W. (1958). A bias and confidence in not-quite large samples, Annals of
Mathematical Statistics (abstracts), 29, 614-623.
Vishwakarma, K., (1995). A neural networks to forecast business cycle indicators,
Elsevier Science, Mathematics, and Computers in Simulations, 39, 287-291.
White, H. (1989). Learning in artificial neural networks: a statistical perspective,
Neural Computations 1, 425-464.
Wu, C.F.J. (1990). On the asymptotic properties of the jackknife histogram,
Annals of Statistics, 18,1438-1452.
Ziari, H.A., Leatham, D.J. & Ellinger, P.N. (1997). Development of statistical
discriminant mathematical programming model via resampling estimation
techniques, American Journal of Agricultural Economics, 79, 1352-1362.
Chapter 3
A ROBUST EVIDENCE OF BUSINESS CYCLE

ASYMMETRIES IN G7 COUNTRIES
ABSTRACT
Artificial neural networks are employed for possible existence of
business cycle asymmetries in Canada, France, Germany, Italy, Japan, UK,
and USA real GDP growth rates. The relationship between real GDP growth
rates and business cycle fluctuations in these countries is modeled using
artificial neural networks. Finally, neural network linearity tests are
constructed from linear model predictions as well as neural network
approximations for all the series.
The results from neural network linearity tests based on in-sample
forecasts show statistically significant evidence of business cycle
asymmetries in Canada, France, Germany, Italy, Japan, UK, and USA real
GDP growth rates. Similarly, neural network linearity test results based on
jackknife out-of-sample forecasts are not much different. However,
compared to Andrano and Savio (2002), there exists statistically significant
evidence of neglected nonlinearities in France, Germany, and UK, and
France, and UK when compared to Kiani and Bidarkota (2004). Therefore,
due to underlying nonlinear data generating processes all types of linear
models cannot be employed to forecast the impact of monetary policy or any
other shock on output in Canada, France, Germany, Italy, Japan, UK, and
USA series.
Keywords: asymmetries; business cycles; neural networks; nonlinearities; vector

autoregressions
JEL codes:C32, C45, E32
1. INTRODUCTION
The empirical research on business cycle asymmetries that employed USA
macroeconomic time series data showed that asymmetries do exist in business
cycle fluctuations. Beaudry and Koop (1993), Brunner (1992-1997), and
Bidarkota (1999-2000) including many others who investigated USA gross
national product (GNP) concluded that business cycles are asymmetric. Similarly,
Neftci (1984), Ramsey and Rothman (1996) including others investigated USA
unemployment rates, and concluded that business cycle asymmetries are prevalent
in these series. Likewise studies by Potter (1995), Anderson and Vahid (1998) and
Anderson and Ramsey (2002) also showed that business cycle asymmetries do
exist in time series data. Alternately, Falk (1995), Sichel (1989), Delong and
Summers (1986), and Diebold and Rudebusch (1990) were unable to show
statistically significant evidence of business cycle asymmetries in the series they
studied.
A number of studies including Auerbach (1982), Gordon (1986), Kling
(1987), Koch and Rasche (1988), Diebold and Rudebusch (1990), Hamilton
(1989), Klein (1990), Estrella and Mishkin (1998) including others focused on
business cycle research, however, only a few studies investigated possible
existence of asymmetries in business cycle fluctuations using international data.
For example, Andreano and Savio (2002) investigated business cycle asymmetries
in G7 countries using Markov Switching models but were not able to detect
business cycle asymmetries in France, Germany, and UK series. Similarly, Kiani
and Bidarkota (2004) investigated possible existence of business cycle
asymmetries in G7 countries but despite using nonlinear and switching time series
models with stable distributions, and long memory they were not able to detect
asymmetries in France, and UK real GDP growth rates. While this study made
contribution over its other counterparts including Andreano and Savio (2002) who
also studied business cycle asymmetries in G7 countries, the basic question
whether “business cycles in G7 countries are alike” remained unanswered. This
posed a challenge for the macroeconomic theorists to develop new theories of
business cycles if empirical research is unable to show an evidence of business
cycle asymmetries in France, and UK that are two prominent members of G7
countries, and the European Union.
A Robust Evidence of Business Cycle Asymmetries … 47
It is important to detect business cycle asymmetries in macroeconomic time

series because asymmetries imply that the effects of contractionary and
expansionary monetary policy, and other shocks on output are asymmetric.
Therefore, any nonlinearity would invalidate the measures of the persistence of
monetary policy, and other shocks on output that are based on linear models
including those observed from linear vector autoregressions. Therefore, it is felt
that the present study can contribute to fill this gap adequately. Therefore, to
further this work, and contrary to Andreano and Savio (2002) who employed
Markov switching models, and Kiani and Bidarkota (2004) who employed
alternative regime switching models, we prefer to use artificial neural networks
for modeling asymmetries in Canada, France, Germany, Italy, Japan, UK, USA
real GDP growth rates, that is approximated using genetic algorithm, because
neural networks are flexible form of nonlinear models which can fit data well
even when distribution of the data generating process as well underlying laws
pertaining the data generating process are unknown (White 1989b).
Artificial neural networks (ANN) have been applied successfully in many
disciplines including business and economics. For example, Kuan and White
(1994) and Swanson and White (1995, 1997a, 1997b), Hutchinson, Lo, and
Poggio (1994), Garcia and Gencay (2000), and Qi and Madala (1999), Gencay
(1999), Vishwakarma (1995), Qi (2001), and Kiani, Bidarkota, and Kastens
(2005) employed neural networks in economics and finance. However, ANN have
been under heavy criticism because of their tendency to overfit the data although
this tendency can be mitigated with careful construction of neural network
architecture (Kiani 2005). Therefore, artificial neural networks (ANN) are
employed to approximate in-sample as well as jackknife out-of-sample forecasts
from linear models as well neural networks to construct neural network linearity
tests to investigate possible existence of business cycle asymmetries in Canada,
France, Germany, Italy, Japan, UK, and USA real GDP growth rates.
The remaining work is split into the following sections. Section 2 gives a
brief description of theoretical justification, and description of the nonlinear
empirical model i.e. neural network models and underlying tests, and section 3
shows data sources, hypotheses tests, empirical results on hypotheses tests, and
forecast performance of neural network models. Section 4 incorporates brief
conclusions.
2. EMPIRICAL MODEL
2.1. Theoretical Justification of Asymmetric Models
With the advent of the Great Depression in the United States of America
(USA), and its impact on other nations in the world, economists started of
thinking on its causes, and there is a voluminous literature that discusses the
causes of the great depression (GD) of 1930s in the USA. Monetarists’ believed
that the Federal Reserve System (FED) of USA was responsible for it, and that
FED’s adequate action would have averted the recessions but the inapt decision of
the FED was solely responsible for pushing the USA economy into the Great
Depression. Keynesians opposed the monetarist’s viewpoint saying that the FED
was responsible for the GD. The debate between Keynesians and monetarists
started soon after GD of 1930s, which lasted for some time, and this was the time
when people started thinking clearly about the business cycles. In this context,
Schumpeter (1939) classified business cycles into four phases i.e. peak,
expansion, recession, depression or trough based on their peak-to-peak and
trough-to-trough duration. Likewise, Mitchell and Burns (1938), and Burns and
Mitchell (1946) developed techniques for analysis of the classical business cycles.
National Bureau of Economic Research (NBER) constituted business cycle
chronology on retrospective identification of business cycle turning points.
NBER’s business cycle dating committee used data on a number of
macroeconomic variables including output, income, employment and trade etc. to
identify the business cycles dates when they occur. This type of business cycle
tells about an absolute decline in output (Stock and Watson 1999). However,
historically, studies on forecasting recessions remained quite sparse until lately.
For business cycle research, post 1930s Great Depression, Keynesian
macroeconomics remained dominant until the 70s when Lucas (1976) pioneered
the rational expectation revolutions. In 80s economists such as Kydland and
Prescott (1982), and Long and Plosser (1983) put forth the idea of real business
cycles without monetary factors and monetary management. This was the time
when Bernanke (1983) in a seminal paper discussed the non-monetary effects that
were the cause of the propagation of the Great Depression. Bernanke was pioneer
to unveil nonmonetary affects that propagated the great depression.
Beaudry and Koop (1993) including others concluded that recessions are
short lived whereas expansions are of longer duration showing that business cycle
fluctuations are asymmetric. Thus, assuming that data-generating process is linear,
the impulse responses function generated from one unit positive shock will have
span and amplitude equal to the impulse response function generated by negative
shock of equal magnitude though in an opposite direction. On the other hand, an
impulse response function generated from one unit positive shock is not equal to
an impulse response generated by one unit negative shock of equal magnitude
when the underlying data generating process in nonlinear. Therefore, business
cycles fluctuations associated with linear models are symmetric, and those
associated with nonlinear model are asymmetric.
Policymakers would be interested to know the impact of monetary policy or
other shocks on output. However, one may not be able to forecast the impact of
such shocks based on linear models when the underlying data generating process
is nonlinear. Therefore, it is imperative to investigate possible existence of
nonlinearities in data series so that appropriate linear or nonlinear forecasting
models are employed to anticipate the impact of monetary policy or other shocks
on output. Moreover, it is of interest to know whether business cycles in different
countries are alike. If they are not, then this poses a serious challenge to
macroeconomic theorists to develop new theories of business cycles that can
explain fluctuations in economic activity without considering country specific
institutional factors.
That is the reason why the present work considers artificial neural networks
that are considered to be highly flexible functional form of nonlinear models that
are data driven and can fit to any time series data adequately without taking into
account the distribution of the data generating process. These models are
elaborated in the following sub-sections in detail.
2.2. Nonlinear Empirical Model
Artificial neural network (ANN), which pertains to artificial intelligence

technology, imitates human brain's learning and decision-making processes. The
ability of information processing makes ANN powerful computational devices,
which can learn from examples and generalizes these learning’s to solve problems
never seen before (Reilly and Cooper (1990). ANN are nonlinear, and
nonparametric statistical methods which are independent of the distributions of
underlying data generating processes (White 1989b).
This work is intended to find possible existence of nonlinearities in Canada,
France, Germany, Italy, Japan, UK, and USA real GDP using in-sample
approximations from neural networks, which is an extension to Kiani and
Bidarkota (2004). Further, these analyses are extended to jackknife out-of sample
framework to observe forecast performance of these models in out-of-sample
framework when compared to the in-sample forecast as well as to investigate the

performance of the models with respect to previous studies (Kiani el al. 2005).
Artificial neural networks (ANN) of the form shown in Equation 3.2 is
employed which nests a linear model of the form shown in Equation 3.1 to
construct neural network linearity tests for testing possible existence of
asymmetries in Canada, France, Germany, Italy, Japan, UK, and USA real GDP
growth rates. An ANN linearity test comprises of three parts. In the first part a
linear model of the form of Equation 3.1 is estimated to recover model forecasts
and residuals for Canada, France, Germany, Italy, Japan, UK, and USA real GDP
growth rates.
y t = π ' wt + u t
(3.1)
where
u t ~ N (0, σ 2 ), ~ ' )' ,

wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = ( π 0 , π 1 ........, π p )'
In the second part forecasts from a neural network model are approximated
using residual obtained from the first part as endogenous variables and lagged real
GDP growth rates as exogenous variables for all the series using Equation 3.2 to
recover residuals and forecasts.
k
ut = π ' wt + ∑θ
j =1
0 j {ψ (γ j ' wt )} + vt
(3.2)
where, ψ (γ ' wt ) = (1 + exp{−γ ' wt }) −1 and π 0 is intercept.

Finally, in the third part using residuals from linear model as well as neural
networks, a test statistic is computed using Equation 3.3 , which is distributed
approximately1 F [m, (n − p − m − 1)] under normality hypothesis. Terasvirta et
al. (1993) and Kiani et al. (2005) also implemented this type of test.
TS = {(SSE1 − SSE 2) / m} /{SSE 2 / (n − p − m − 1)} (3.3)
1
This test statistic is approximate because of the nuisance parameter that appears under the
alternative hypothesis (Davies, 1977; Andrews, 2001).
where, in Equation 3.3 , m denotes the number of restrictions, n is number of

observations, and p is the number of lags in each of the models
estimated/approximated.
Neural networks linearity tests are constructed using in-sample forecasts from
linear models and neural networks approximations. In addition, neural networks
linearity tests using jackknife2 out-of-sample forecasts are also constructed using
linear model forecasts, and neural networks approximations. Jackknife out-of-
sample forecasts are employed to study the behavior of these models in out-of-
sample testing, and to compare these results with those of previous studies (White
1989a; Kiani at al 2005).
To avoid the likelihood of obtaining local optima, genetic algorithm is

employed with a couple of random starts to obtain best parameter vector for
neural network approximations which is considered to be a reliable estimation
algorithm but it appeared to be very slow. Therefore, a nonlinear optimization
routine is combined with genetic algorithm3, which worked satisfactorily although
it was still slower than any other optimization routine.
3.1. Data Sources
Quarterly real GDP data for Canada, France, Germany, Italy, Japan, the
2
Jackknife re-sampling technique is used when the distribution of the parameters under review is
either unknown, when it cannot be characterized by a mathematical function, or when the
mathematical function is especially difficult to estimate. Standard jackknife out-of-sample
forecasts were used by Quenouille (1949), and thereafter, Tuckey (1958). However, sub-sample
jackknife technique was initially proposed by Wu (1990) and thereafter by Politis and Romeo
(1994), Politis et al. (1997), and Ziari et al. (1997). Compared to the standard jackknife, sub-
sample jackknife drops more than one observation to estimate out-of-sample forecast of the
remaining m = n − d observations, where, n is the total number of observations
and d = 2,3,......... ....., n − 1 .
3
Following De Jong (1975) genetic algorithm was used in biology, engineering and operation
research (Goldberg 1989). The first ever, economic application of genetic algorithm was
implemented by Axelord (1987), and thereafter by Marimon, McGartten, and Sargent (1990),
and Dorsey and Walter (1995).
United Kingdom (UK), and the United States of America (USA) were obtained
from the November 2006 version of the International Financial Statistics’ CD-
ROM. The dataset spans from 1957:1 to 2006:2 for all countries except for France
for which the data spans from 1965:1 to 2006:2, and for Germany, and for Italy,
the data series span from 1960:1 to 2006:2. Table 1 presents additional
information on the data series employed, and Figure 1 plots real GDP growth
rates for Canada, France, Germany, Italy, Japan, UK, and USA
Canada France Germany Italy Japan UK USA

Data Quarterly Quarterly Quarterly Quarterly Quarterly Quarterly Quarterly
Series Real GDP Real Real Real GDP Nominal Real GDP Real
GDP GDP GDP GDP
Sample 1957:1- 1970:1- 1960:1- 1970:2- 1957:1- 1957:1- 1957:1-
Period 2006:2 2006:2 2006:2 2006:2 2006:2 2006:2 2006:2
Sample 197 165 185 185 197 197 197
Length
Notes
1. Quarterly seasonally adjusted GDP data is used for all countries from the November
2006 edition of the International Financial Statistics (IFS) CD-ROM.
2. Nominal GDP is employed for Japan because seasonally adjusted data was only
available for the nominal series and not for the real series on the IFS CD-ROM.
The chief hypothesis of this chapter is linearity versus the alternative

hypothesis of nonlinearity for Canada, France, Germany, Italy, Japan, UK, and
USA real GDP growth rates. The hypothesis is based on the test statistics
constructed from in-sample approximations from neural networks, and forecasts
obtained from linear models. These hypotheses are repeated for Canada, France,
Germany, Italy, Japan, UK, and USA based on jackknife out-of sample forecasts
from linear model as well as neural network approximations. When the null
hypothesis of linearity is true, nonlinearities do not prevail in the series being
studied, however, when the null is false, nonlinearities do exist in the data series
being tested.
Figure 1. Annualized Real GDP Growth Rates.
3.3. Results on Hypothesis Test
Table 2 shows neural network linearity test results for Canada, France,
Germany, Italy, Japan, UK and USA real GDP growth rates. In this Table for
example, the neural network test statistic for Canada that is constructed from in-
sample forecasts from linear models and neural networks is shown in column 2
row 1 of the Table. The relevant critical values from F distribution are shown
beneath each of the test statistic in brackets, and p-value are presented in
parenthesis. Test statistics for France, Germany, Italy, Japan, UK, and USA are
shown in column 2 , row 2 through 7 respectively, and relevant critical values are
shown in brackets, and p-values in parentheses beneath each test statistics.
Table 2. Neural Network Tests with Linear Models
In-Sample Forecasts Jackknife Out-of-Sample Forecasts

Test Statistics RMSE Test Statistics RMSE
Canada 280.8325 0.1404 41.4805 0.2918
[2.0586] [2.0586]
(0.0001) (0.000)
France 2428.0671 0.1568 520.4057 0.3247
[2.0687] [2.0687]
(0.0000) (0.000)
Germany 380.9606 0.2169 191.1076 0.2905
[2.0613] [2.0613]
(0.0001) (0.000)
Italy 672.838 0.2687 218.7765 0.4444
[2.0613] [2.0613]
(0.0001) (0.000)
Japan 209.5664 0.1793 36.6850 0.3405
[2.0586] [2.0586]
(0.0001) (0.000)
UK 202.0907 0.2201 59.7420 0. 3512
[2.0586] [2.0586]
(0.000) (0.000)
USA 346.0786 0.1247 52.9753 0. 2641
[2.0586] [2.0586]
(0.000) (0.000)
Notes:
1. Column 1 in this Table show name of the countries included in the study.
2. Test statistics presented in the Table are based on linear as well as neural network
models that are shown in the following two Equations:
yt = π ' wt + ut
2.1a
where,
u t ~ N (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = ( π 0 , π 1 ........, π p )'

k
ut = π ' wt + ∑θ
j =1
0 j {ψ (γ j ' wt )} + vt
2.1b
3. The test statistic is distributed F [m, (n − p − m − 1)] under normality hypothesis are
calculated using following Equation:
TS = {(SSE1 − SSE 2) / m} /{SSE 2 / (n − p − m − 1)} 2.1c

where, in Equation 2.1c , m denotes the number of restrictions in the unrestricted
model, n is the number of observations, and p is number of lags.
4. Column 2 shows test statistics constructed from in-sample forecasts whereas column
4 shows test statistics constructed from jackknife out-of-sample forecasts.
5. In this Table, for each test statistic constructed from in-sample as well as jackknife out-
of-sample forecasts, the relevant critical values from F distribution are shown
beneath each test statistic presented in this Table in addition to the p-values that are
shown beneath each test statistic in parenthesis.
6. In column 3 , root mean squared errors (RMSE) for in-sample forecasts are presented
whereas in column 5 , RMSE calculated from jackknife out-of-sample forecasts are
shown.
Neural network test statistic for Canada is statistically significant at 5 percent

level of significance based on relevant p-values shown in the Table. Inferences do
not change when significance level is switched from 5 percent to 10 percent
level. Neural network linearity test statistics for the remaining countries i.e.
France, Germany, Italy, Japan, UK, and USA are also significant at 5 as well as
10 percent level of significance. These results reveal that business cycle
asymmetries do prevail in Canada, France, Germany, Italy, Japan, UK, and USA
real GDP growth rates.
For the purpose of detecting possible existence of nonlinearities in all the
series, these results would suffice, however, the analysis is extended to jackknife
out-of-sample framework to detect possible existence of nonlinearities in all the
series in out-of-sample framework, and to study out-of-sample performance of the
models compared to their in-sample performance and with the results from the
previous studies in this area. In this context, for example, column 4 rows 1
through 7 in Table 2 presents neural network tests constructed from jackknife
out-of-sample forecasts for Canada, France, Germany, Italy, Japan, UK, and USA
series. The relevant critical values from F distribution are shown beneath each
test statistic in brackets, and p-value in parenthesis.
Neural network linearity test for possible existence of business cycle
nonlinearities, using in-sample forecasts from neural networks with univariate
linear models, show statistically significant evidence of asymmetries in all the
series. Similarly, neural network nonlinearity tests constructed from jackknife out-
of-sample forecasts from neural networks also show statistically significant
evidence of business cycle asymmetries in Canada, France, Germany, Italy, Japan,
UK, and USA real GDP growth rates.
3.4. Forecast Performance of Neural Network Models
Table 2 shows in-sample root mean square error (RMSE) computed from in-
sample as well as jackknife out-of-sample forecasts from neural network models
for all the series. In this Table, RMSE for in-sample forecasts are shown in
column 3 rows 1 to 7 and RMSE for jackknife out-of-sample forecasts are
shown in column 5 rows 1 to 7 respectively for Canada, France, Germany, Italy,
Japan, UK, and USA. Comparing RMSE from in-sample to jackknife out-of-
sample forecasts, for example for Canada, it transpires that in-sample forecast
performance of neural network models for Canada is superior to jackknife out-of-
sample forecasts from neural networks because RMSE from in-sample forecasts is
lower than what is computed from jackknife-out-of-sample forecasts. These
results hold for the remaining series, and are in line with previous studies
including Terasvirta, et al. (1993), and Kiani et al. (2005).
The results on nonlinearity for USA are in line with Bidarkota (2000) and
Andreano and Savio (2002). Similarly the results on nonlinearity for Canada,
Italy, Japan, and USA are in line with Andreano and Savio (2002). Moreover, the
results for Canada, Italy, Germany, Japan, and USA are in line with Kiani and
Bidarkota (2004). This shows that evidence against linearity for Canada, Italy,
Japan, and USA is robust to change in-samples as well as change in testing
approach.
Neural network models outperform the traditional statistical tests for
remaining nonlinearities, which is in line with the previous studies including
Terasvirta, et al. (1993). Similarly, in-sample forecast performance of neural
network models is superior to linear model which is in line with Kiani et al.
(2005).
4. CONCLUSION
In the present work artificial neural networks (ANN) are employed with
univariate linear models for possible existence on business cycle asymmetries in
Canada, France, Germany, Italy, Japan, UK, and USA real GDP growth rates
based on in-sample and jackknife out-of-sample forecasts from linear models and
those approximated from neural networks. Neural network linearity tests are
constructed from in-sample as well as jackknife out-of-sample forecasts for all the
series.
The study results based on in-sample forecasts provide a robust evidence of

asymmetries in business cycle fluctuations in Canada, France, Germany, Italy,
Japan, UK, and USA real GDP growth rates. In addition, the results based on
jackknife out-of-sample forecasts strengthen the evidence for existence of
business cycle asymmetries in Canada, France, Germany, Italy, Japan, UK, and
USA real GDP growth rates. Therefore, the impact of monetary policy or any
other shock on output in these countries cannot be determined using linear models
because the underlying data generating processes are nonlinear. These findings
also reveal that in-sample forecasts from neural networks are superior to the
jackknife out-of-sample forecasts from neural networks that is in line with
previous studies (Kiani at al. 2005).
The results show statistically significant evidence of business cycle
asymmetries in all the countries studied. Therefore, the impact of monetary policy
or any other shock on output in Canada, France, Germany, Italy, Japan, UK, and
USA cannot be determined using linear model and those derived from vector
autoregressions.
5. REFERENCES
Anderson, H., & Ramsey, J. (2002). U.S. and Canadian industrial production
indices as coupled oscillators. Journal of Economic Dynamics and Control,
26, 33-67.
Anderson, H., & Vahid, F. (1998). Testing multiple equation systems for common
nonlinear components. Journal of Econometrics, 84, 1-36.
Andreano, M., & Savio, G. (2002). Further evidence on business cycle
asymmetries in G7 countries. Applied Economics, 34, 895-904.
Andrews, D. (2001). Testing when parameter is under the boundary of the
maintained hypothesis. Econometrica, 69, 683–734.
Auerbach, A. (1982). The index of leading indicators: measurement without
theory, thirty-five year later. The Review of Economics and Statistics, 64,
589-595.
dilemma. In Davis, L. D. (Eds.), Algorithm and Simulated Annealing, Morgan
Kaufmann, Los Altos, CA: pp 32-41.
Bernanke, B. (1983). Nonmonetary Effects of the Financial Crisis in Propagation

of the Great Depression. American Economic Review, 73(3), 257-76.
GNP: robust evidence. The Review of Economics and Statistics, 82, 153-157.
Burns, A., & Michell, W. (1946). Measuring business cycles. NBER, New York.
Davies, R. (1977). Hypothesis testing when a nuisance parameter is present only
under the alternative. Biometrika, 64, 247–254.
System unpublished Ph.D. dissertation. University of Michigan, Department
DeLong, B., & Summers, L. (1986). Are business cycles symmetrical? In Gordon,
R.J. (Eds.), The American Business Cycle: Continuity and Change, Chicago,
IL, Chicago University Press.
Diebold, F., & Rudebusch, G. (1990). A nonparametric investigation of duration
dependence in the American business cycle. Journal of Political Economy,
98, 596-616.
Estrella, A., & Mishkin, F. (1998). Predicting U.S. Business Cycle Regimes.
series over the business cycle. Journal of Political Economy, 94, 1096-1109.
Garcia, R., &Gencay, R. (2000). Pricing and hedging derivative securities with
neural networks and a homogeneity hint. Journal of Econometrics, 94, 93-
115.
Gencay, R. (1999). Linear, nonlinear and essential foreign exchange prediction
with simple technical trading rules. Journal of International Economics, 47,
91-107.
Learning’s. Addison Wisely, Readings, MA.
Gordon, R. (1986). The American Business Cycle: continuity and change. NBER
studies in business cycles, Vol. No. 25, University of Chicago Press, Chicago,
IL.
Hamilton, J. (1989). A new approach to the economic analysis of nonstationary
time series and business cycle. Econometrica, 57, 357-384.
Hutchinson, J., Lo, A., & Poggio, T. (1994). A nonparametric approach to pricing
and hedging derivative securities via learning networks. Journal of Finance,
49, 851-889.
Kiani, K. (2005). Testing nonlinearities in using time series models and artificial
neural network. Computational Economics, 26, 65-89.
Kiani, K., Bidarkota, P., & Kastens, T. (2005). Forecast performance of neural
networks and business cycle asymmetries. Applied Financial Economics
Letters, 1, 205-210.
Kiani, K., & Bidarkota, P. (2004). On business cycle asymmetries in G7
Kling, J. (1987). Predicting the turning points of business and economic time
series. Journal of Business, 60, 201-238.
Klein, P. (1990). Analyzing Modern Business Cycles: Essays in Honor of
Geoffery H. Moore. M.E. Sharpe, M.E. (Eds.). NY: Armonk.
Koch, P., & Rasch, R. (1988). An examination of the commerce department
leading indicator approach. Journal of Business and Economic Statistics,
6,167-187.
Kuan, C., & White, H. (1994). Artificial neural networks: an economic
perspective. Econometric Review, 13, 1-91.
Kydland, F., Prescott, E. (1982). Time to build and aggregate fluctuations.
Econometrica, 50, 1345-70.
Long, J., & Plosser, C. (1983). Real business cycles. Journal of Political
Economy, 91, 39-69.
Lucas, R. (1976). Econometric policy evaluation: a critique. Carnegie-Rochester
Conference Series in Public Policy, 1, 19-46.
Marimon, R., McGratten, E., & Sargent, T. (1990). Money as Medium of
Economics Dynamics and Control, 14, 329-373.
Mitchell, W., & Burns, A. (1938). Statistical indicator of cyclical revivals. NBER,
New York.
Neftci, S. (1984). Are economic time series asymmetric over the business cycle?
Politis, D., & Romeo, J. (1994). Large sample confidence region based on sub-
sample under minimal assumption. Annals of Statistics, 22, 2031-2052.
Politis, D., Romeo, J., & Wolf, M. (1997). Sub-sampling for heteroskedastic time
series. Journal of econometrics, 81, 81-317.
Potter, S. (1995). A non-linear approach to U.S. GNP. Journal of Applied
Qi, M. (2001). Predicting US recessions via leading indicators via neural network
models. International Journal of Forecasting, 17, 383-401.
Qi, M., & Madala, G. (1999). Economic factors and the stock market: a new
perspective. Journal of Forecasting, 18, 151-166.
Quenouille, M. (1949). A note on bias in estimation. Biometrika, 43, 353-60.
real world systems. In Zornetzer, J., Davis, L., & Lau, C. (Eds.), An
Introduction to Neural and Electronic Networks, New York: NY.
Schumpeter, J. (1939). Business cycles: a theoretical, historical, and statistical
analysis of the capitalist process. McGraw-Hill:New York.
Sichel, D. (1989). Are business cycles asymmetric? a correction. Journal of
Swanson, N., & White, H. (1995). A model selection approach to assessing the
information in the term structure using linear model and artificial neural
networks. Journal of Business Economics and Statistics, 13, 265-275.
Swanson, N., & White, H. (1997a). A model selection approach for real time
macroeconomic forecasting using linear models and artificial neural
networks. The Review of Economics and Statistics, 79, 540-550.
Swanson, N., & White, H. (1997b). Forecasting economic time series using
flexible vs. fixed and linear vs. nonlinear economic models. International
journal of Forecasting, 13, 439-461.
Stock, J., & Watson, M. (1999). Business cycle fluctuations in US
macroeconomic time series. In Taylor, J.B., & Woodford, M. (Eds.),
Handbook of macroeconomics, Volume 1, Elsevier Science: B.V.
Terasvirta, T., Lin, C., & Granger, C. (1993). Power of neural networks test.
Tuckey, J. (1958). A bias and confidence in not-quite large samples. Annals of
Vishwakarma, K. (1995). A neural networks to forecast business cycle indicators.
Mathematics and Computers in Simulations, 39, 287-291.
White, H. (1989a). Some asymptotic results for learning in single hidden-layer
feed forward network models. American Statistical Association, 84, 1003-
1013.
White, H. (1989b). Learning in artificial neural networks: A statistical

perspective. Neural Computations, 1, 425-464.
Wu, C. (1990). On the asymptotic properties of the jackknife histogram. Annals of
Statistics, 18, 1438-1452.
Ziari,H., Leatham, D., & Ellinger, P. (1997). Developing of statistical
discriminant mathematical programming model via re-sampling estimation
techniques. American Journal of Economics, 79, 1352-1362.
Chapter 4
ASYMMETRIC BUSINESS CYCLE

FLUCTUATIONS AND CONTAGION EFFECTS IN
G7 COUNTRIES
ABSTRACT
Real gross domestic product (GDP) growth rates are employed for
studying possible existence of business cycle asymmetries in the group of
seven (G7) industrialized countries, i.e. Canada, France, Germany, Italy,
Japan, UK and USA. Asymmetries in these countries are modeled using
artificial neural networks (ANN) via in-sample as well as jackknife out-of-
sample forecasts. In addition, asymmetric business cycles and contagion
effects are explored in Canada, France, Germany, Italy, Japan, UK and USA
real GDP growth rates using bivariate framework.
The results show statistically significant evidence of asymmetries in
business cycle fluctuations in Canada, France, Germany, Italy, Japan, UK,
and USA real GDP growth rate series. The results on business cycle
asymmetries for Canada Italy, Japan, UK, and USA are in line with
Andreano and Savio (2002), and Kiani and Bidarkota (2004), in exception of
France and UK that show statistically significant evidence of neglected
nonlinearities. Therefore, the impact of monetary policy or any other shock
on the output in these countries cannot be determined using linear models
including those derived from linear vector autoregressions. In addition, using
bivariate framework, similar results are obtained on business cycle
asymmetries. In addition existence of statistically significant evidence of
contagion effects does exists in all the series.
Keywords: asymmetries; business cycles; neural networks; nonlinearities; vector

autoregressions
JEL codes: C32, C45, E32
1. INTRODUCTION
A wide body of empirical research that focused on detecting business cycle
asymmetries in economic fluctuations employed USA macroeconomic time series
and concluded that asymmetries in business cycle fluctuations do exist in
macroeconomic time series. In this context studies by Beaudry and Koop (1993),
Brunner (1992-1997), and Bidarkota (1999-2000) including others investigated
USA gross national product (GDP), and showed existenceof business cycles
asymmetries in the series. Likewise, Neftci (1984), Ramsey and Rothman (1996)
including others studied USA unemployment rates, and concluded for
predominance of asymmetric business cycle fluctuations in the series. Similarly,
Potter (1995), Anderson and Vahid (1998), and Anderson and Ramsey (2002) also
showed existence of business cycle asymmetries in macroeconomic time series.
Contrary to that Falk (1986), Sichel (1989), Delong and summers (1986), and
Diebold and Rudebusch (1990) either failed to reveal asymmetries in business
cycle fluctuations or were unable to find significant evidence of business cycle
asymmetries in the series they studied.
Nonlinearities imply that the effects of contractionary and expansionary
monetary policy and other shocks on output are asymmetric. Therefore, any
nonlinearity would invalidate the measures of the persistence of monetary policy
or any other shock on output that is based on linear models including those
derived from linear vector autoregressions when the underlying data generating
process is nonlinear. Policymakers would be interested to know the impact of
monetary policy or any other shock on output; however, one might not be able to
foretell the impact of such shocks on output that would be based on linear models
when the underlying data generating process is nonlinear. Therefore, it is
imperative to detect possible existence of nonlinearities in data series so that
appropriate forecasting models (linear or nonlinear) are employed to anticipate the
impact of monetary policy or other shocks on output. Moreover, it would be of
interest for macro theorists if business cycle fluctuations in all the countries of the
world are alike. If they were not, the macro-theorists would need to come up with
Asymmetric Business Cycle Fluctuations and Contagion Effects… 65
new theories of business cycles without taking into account the underlying
country specific institutional factors.
While studies by Auerbach (1982), Gordon (1986), Kling (1987), Koch and
Rasche (1988), Diebold and Rudebusch (1990), Hamilton (1989), Klein (1990),
Estrella and Mishkin (1998) focused on business cycle research, only a few
studies investigated possible existence of asymmetries in business cycle
fluctuations using international data. For example, Andreano and Savio (2002)
investigated business cycle asymmetries in G7 countries using Markov Switching
models but they were not able to detect asymmetries in France, Germany, and UK
series. Similarly, Kiani and Bidarkota (2004) studied possible existence of
business cycle asymmetries in real GDP series in G7 countries but despite using
nonlinear, and switching time series models with stable distributions, and long
memory, Kiani and Bidarkota were not able to find possible existence of business
cycle asymmetries in France and U.K series. This necessitated undertaking
another study to address the basic question whether business cycles in all the
countries of the world starting with the G7 countries are alike is still unanswered.
The present study fills this gap adequately modeling real GDP growth rates in G7
countries using artificial neural networks (ANN) that are considered to be highly
flexible functional form of nonlinear models that can fit to any data series without
taking into consideration the distribution of the underlying data generating
process.
Neural Networks have been applied successfully in many disciplines
including business and economics. For example, Kuan and White (1994), and
Swanson and White (1995, 1997a, 1997b) employed ANN in economics.
Hutchinson, Lo, and Poggio (1994), Garcia and Gencay (2000), and Qi and
Madala (1999), and Gencay (1999) employed ANN in finance. However, only
Vishwakarma (1995), Qi (2001), Kiani (2005), and Kiani, Bidarkota, and Kastens
(2005) focused on business cycles using neural networks. Therefore, in-sample as
well as jackknife out-of-sample forecasts are approximated from neural networks,
and linear models to construct neural network linearity tests that were originally
proposed by Terasvirta, Lin and Granger (1993) for possible existence of business
cycle asymmetries in Canada, France, Germany, Italy, Japan, UK, and USA real
GDP growth rates. Moreover, the present analysis is extended to bivariate
framework to reveal further evidence of asymmetric fluctuations, linkages, and
spillover effects within these countries.
The remaining study is split into the following sections. Section 2 shows a
brief description of empirical models, and underlying tests, and section 3 shows
data sources, preliminary data analysis, hypotheses tests, empirical results on
hypotheses tests, and forecast performance of neural network models. Section 4

incorporates brief conclusions.
2. EMPIRICAL MODEL: ARTIFICIAL NEURAL NETWORK

Artificial neural networks (ANN) pertain to artificial intelligence technology
that mimics human brain's learning and decision-making process. The ability of
information processing makes ANN powerful computational devices that can
learn from examples and generalize these learning to solve problems never seen
before (Reilly and Cooper 1990). ANN are nonlinear, nonparametric statistical
methods which are independent of the distributions of underlying data generating
processes (White 1989b). The present research employs ANN to investigate
possible existence of business cycle nonlinearities in Canada, France, Germany,
Italy, Japan, UK, and USA real GDP growth rates using in-sample forecasts
approximated from ANN that is an extension to Kiani and Bidarkota (2004).
Likewise, the analysis pertaining to testing business cycle asymmetries is
extended to neural network linearity tests using bivariate framework that are
constructed from in-sample as well as jackknife out-of sample approximations
from neural networks to investigate behavior of these models out-of-sample, and
to observe spillover and contagion effects within G7 countries in bivariate
framework.
Quenouille (1949) used jackknife re-sampling to reduce the bias in
estimators, and thereafter Tuckey (1958) employed jackknife re-sampling for
estimating variances. However, Wu (1990) introduced sub-sample jackknife
technique, which was also used by Politis and Romeo (1994). Later, Politis et al.
(1997), and Ziari et al. (1997) also used this re-sampling technique. In the sub-
sample jackknife, more than one observation is dropped to estimate out-of-sample
forecast of the remaining m = n − d observations, where, n is the total number of
observations, and d = 2,3,......... ....., n − 1 when compared to the standard
jackknife.
2.1. Neural Network Linearity Test
The model for constructing neural network linearity test due to Terasvirta et
al. (1993) comprise of a linear model shown in Equations 4.1 as well as a neural
network model as of Equation 4.2 . Although this model is constructed to work
with lagged exogenous variables from more than one data series, it can be
restricted to contemporaneous independent and lagged dependent variables from

one series only. A general form neural network linearity test can be shown in the
following two Equations.
yt = π ' wt + ut
4.1.
where,
u t ~ N (0, σ 2 ), ~ ' )' ,

wt = (1, w ~ = ( y ,..., y , x ,..., x )'
w
t t t −1 t− p t −1 t− p
and π = ( π 0 , π 1 ........, π p )'

k
ut = π ' wt + ∑θ
j =1
0 j {ψ (γ j ' wt )}
4.2.
where, in Equation 4.2 ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π 0 is intercept. Equation

4.2 shows a nonlinear neural network model that nests linear model shown in
Equation 4.1 . Under normality conditions, the test statistics for this test is
distributed approximately F [ m, ( n − p − m − 1)] , which is computed using
Equation 4.3 . The distribution of this test statistic is approximate because of the
nuisance parameter appearing under the alternative hypothesis (Davies, R. 1977;
and Andrews, W. 2001).
TS = {(SSE1 − SSE 2) / m} /{SSE 2 / (n − p − m − 1)} 4.3.
where, in Equation 4.3 , m denotes the number of restrictions, n the number of

observations, and p the number of lags in the linear, and ANN models.
Quarterly real GDP series for Canada, France, Germany, Italy, Japan, UK,
and USA were obtained from November 2006 version of the International
Financial Statistic’s CD-ROM. The dataset spans from 1957:1 to 2006:2 for all
countries except for France, for which the data spans from 1965Q1-2006Q2, and
Germany, and Italy for which the data ranges from 1960:1 to 2006Q2. Figure 1
plots real GDP growth rates for Canada, France, Germany, Italy, Japan, UK, and
USA series, and Table 1 shows additional information for all the series.
Canada France Germany Italy Japan UK USA

Data Quarterly Quarterly Quarterly Quarterly Quarterly Quarterly Quarterly
Series Real Real Real Real Nominal Real Real GDP
GDP GDP GDP GDP GDP GDP
Sample 1957:1- 1970:1- 1960:1- 1970:2- 1957:1- 1957:1- 1957:1-
Period 2006:2 2006:2 2006:2 2006:2 2006:2 2006:2 2006:2
Sample 197 165 185 185 197 197 197
Length
Notes:
1. The quarterly seasonally adjusted GDP data is obtained all countries from the November
2006 edition of the International Financial Statistics (IFS) CD-ROM.
2. Nominal GDP is used for Japan because seasonally adjusted data was available only for
the nominal series and not for the real series in the IFS CD-ROM.
Figure 1. Real GDP Growth.
In in-sample approximations are obtained from ANN in conjunction with

linear models to construct neural network linearity tests using test statistic of the
form shown in the Equation 4.3 for each of the series i.e. Canada, France,
Germany, Italy, Japan, UK, and USA. Likewise, jackknife out-of-sample forecasts
from ANN are also approximated to construct neural network linearity test for
each of the series studied.
While neural network linearity test are constructed from in-sample as well as
jackknife out-of-sample approximations from ANN in conjunction with univariate
linear model forecasts, the present work also considers constructing neural
network linearity tests using in-sample as well as jackknife out-of-sample
approximations from neural networks in conjunction with forecast from bivartiae
linear models. The rational for this extension is to find possible existence of
additional evidence of business cycle asymmetries in addition to considering
linkages or spillover, and contagion effects between these countries. These types
of linkages were also discussed in Anderson and Ramsay (2002) in a bivariate
framework for Canada, and USA time series. A general bivariate vector
autoregression (VAR) model can be represented by the following Equations:
p p
yt = α 1 + ∑β
i =1
1i y t − i + ∑β
i =1
2 i xt − i +
4.4.
p p
xt = δ 1 + ∑
j =1
γ 1 j yt − j + ∑γ
j =1
2 j xt − j +e
4.5.
where, y t and xt are contemporaneous whereas y t − p and x t − p are lagged real

GDP growth rates for example for a bivariate vector autoregression model CAFR
that comprises of two series i.e. Canada and France for all p ≥ 1 . The CAFR
bivariate model consist of two equations, i.e. CAFR1 which is of the form of
Equation 4.4 , and CAFR2 as of Equation 4.5 . These equations are employed to
construct two separate neural network linearity tests.
In the first part of neural network linearity test, a VAR model CAFR for
Canada and France is estimated to recover residuals ( μ̂ i for i = 1, 2 ), and residual
sum of squares (RSS i for i = 1,2) for each of the equations of this VAR are
recovered. This process is repeated for each of the twenty one bivariate models
i.e. CAFR, CAGR, CAIT, CAJP, CAUK, CAUS, FRGR, FRIT, FRJP, FRUK,
FRUS, GRIT1, GRJP, GRUK, GRUS, ITJP, ITUK, ITUS, JPUK, JPUS, and
UKUS that are estimated using Canada, France, Germany, Italy, Japan, UK, and
USA real GDP growth rates for recovering residuals, and (RSS i ) that are finally
employed for constructing neural network linearity tests.
In the second part of the bivariate neural network linearity test, in-sample
forecasts from ANN are approximated using Equation 4.2 wherein residuals ( û i
for i = 1, 2 ) from each of the VAR equation are employed as endogenous
variables, and lagged real GDP growth rates ( y t −1 ,.......... y t − k xt −1 ,.....xt − k
(for k ≥ 0) ) as exogenous variables. For example, for the bivariate VAR model
CAFR, two separate neural network linearity test are constructed using its
equations CAFR1 and CAFR2 . From this part of the test, the residuals ( v̂t
for i = 1, 2 ), and sum of squared residuals SSE i (for i = 1, 2 ) are recovered from
each of the neural network model approximated. Finally, test statistic is calculated
using Equation 4.6 .
TS = {( RSSi − SSE) / m} /{SSEi /(n − p − m −

4.6.
where in Equation 4.6 , m denotes the number of restrictions, n is the number of

observations, and p the number of lags in the linear, and ANN the models. In
Equation 4.6 RSS i ( for i = 1,2) are residual sum of squares from the first
equation whereas SSEi show squared residual sum from the second part of the
neural network linearity test.
The process employed for constructing neural network linearity tests using
bivariate CAFR model is repeated for all the equations of the remaining VAR
models estimated for each of the series. In addition to using in-sample forecast
from linear models in conjunction with ANNs, jackknife out-of-sample forecasts
from linear models and ANN approximations are also employed for constructing
neural network linearity tests for Canada, France, Germany, Italy, Japan, UK, and
USA real GDP growth rates.
To avoid the likelihood of obtaining local optima, genetic algorithm (GA) is
employed with a couple of random starts to obtain best parameter vector for
neural network approximations which is considered to be the most reliable
estimation algorithm for estimating any type of nonlinear functional form
including ANN but it appeared to be very slow. Therefore, GA is employed in
conjunction with fminsearch routine which is a canned optimization routine from
MATLAB which is simplex algorithm that worked satisfactorily although it was
still slower than any other optimization algorithm that could have been used for
estimating nonlinear models. After its applications in biology, and engineering,

GA was employed in operation research by Goldberg (1989). However, its
economic application was implemented by Axelord (1987), Marimon, McGratten,
and Sargent (1990), and Dorsey and Mayer (1995).
3.1. Preliminary Data Analysis
Preliminary data analysis (detailed results not shown for brevity) reveals that
skewness is statistically significant, and excess kurtosis show significant fat tails
in Canada, France, Germany, Italy, UK, and USA series including Japan that
show marginal fat tails. The Jarque-Bera test rejects normality for all countries
except Canada. Augmented Dicky Fuller (ADF) test indicates unit roots in levels
(with constant and time trend) for all countries but not in growth rates (with
constant only). The only exception is Japan for which the test fails to reject unit
roots in growth rates with a constant term only but does reject with constant and
time trend. Goldfeld-Quandt test fails to reject homoskedasticity in all countries,
and Lagrange Multiplier (LM) test detects autoregressive conditional
heteroskedasticity (ARCH) in Japan only.
3.2. Estimation Results
Table 2 shows test statistics for neural network linearity tests that are
constructed from in-sample, and jackknife out-of-sample forecasts from univariate
linear models as well as neural networks. In this Table column 2 rows 1 − 7
show test statistics for neural network linearity tests respectively for Canada,
France, Germany, Italy, Japan, UK, and USA series. The relevant p-values for
each of the test statistic are juxtaposed in the subsequent columns in the same
row. Column 5 , rows 1 − 7 show test statistics for neural network linearity tests
that are constructed using jackknife out-of-sample forecasts from linear models as
well as neural networks. Again the relevant p-values are juxtaposed in the
subsequent column in the same row. This Table also show root mean squared
errors (RMSEs) for in-sample as well as jack out-of-sample forecasts that are
shown in columns 4 and 7 in rows 1 − 7 respectively for Canada, France,
Germany, Italy, Japan, UK, and USA real GDP growth rates.
Table 2. Neural Network Tests with Univariate Linear Models
In-Sample Forecasts Jackknife Out-of-Sample Forecast

Model Test Statistic p-Values RMSE Test Statistic p-value RMSE
Canada 280.8325 0.0000 0.1404 41.4805 0.0000 0.2918
France 2428.0671 0.0000 0.1568 520.4057 0.0000 0.3247
Germany 380.9606 0.0000 0.2169 191.1076 0.0003 0.2905
Italy 672.8380 0.0000 0.2687 218.7765 0.0226 0.4444
Japan 209.5664 0.0000 0.1793 36.6850 0.0000 0.3405
UK 202.0907 0.0000 0.2201 59.7420 0.0000 0. 3512
USA 346.0786 0.0000 0.1247 52.9753 0.0080 0. 2641
Notes:
1. Column 1 in this Table shows the names of the countries included in the study.
2. Test statistics presented in this Table are based on in-sample as well as jackknife out-of-
sample forecasts from univariate linear and neural network. These models are shown
in the following two Equations:
3. yt = π ' wt + ut
(2.1a)
where,
u t ~ N (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
and π = ( π 0 , π 1 ........, π p )'

k
ut = π ' wt + ∑θ
j =1
0 j {ψ (γ j ' wt )} + (2.1b)
4. Test statistic for this test is distributed F [ m, ( n − p − m − 1)] under normality

hypothesis. The test statistic can be shown using following Equation:
TS = {(SSE1 − SSE 2) / m} /{SSE 2 / (n − p − m − 1)} (2.1c)
where in Equation 2.1 c m denotes the number of restrictions in the unrestricted
model, n is the number of observations, and p is the number of the lags.
5. Column 2 shows test statistics constructed from in-sample forecasts whereas column
5 shows test statistics constructed from jackknife out-of-sample forecasts. The p-
values for each test statistic are juxtaposed in the subsequent column in the same row.
6. In column 4 , root mean squared errors (RMSE) for in-sample forecasts are presented
whereas in column 7 , RMSE calculated from jackknife out-of-sample forecasts are
shown.
Table 3 shows test statistics for neural network linearity tests that are based
on in-sample as well as jackknife out-of-sample forecasts from bivariate linear
models, and neural networks. In this Table column 2 and 5 show test statistics
computed from all the bivariate linear models in conjunction with the relevant
neural networks for Canada, France, Germany, Italy, Japan, UK, and USA real
GDP growth rates. Relevant p-values for each of the test statistic are juxtaposed in
the subsequent column in the same row. The Table also show root mean squared
errors (RMSE) for in-sample and jackknife out-of-sample forecasts that are shown
in columns 4 and 7 respectively for all the models estimated for Canada, France,
Germany, Italy, Japan, UK, and USA real GDP growth rates.
Table 3. Neural Network Tests with Bivariate Linear Models

Model Test Statistic p-Values RMSE Test Statistic p-Value RMSE
CAFR1 270.1986 0.0000 1.0292 16.4824 0.0000 0.3735
CAFR2 2332.0594 0.0000 0.4591 33.2566 0.0000 1.2228
CAGR1 265.5669 0.0000 1.1298 5.6772 0.0003 0.4257
CAGR2 307.3003 0.0000 0.8458 2.9238 0.0226 0.8345
CAIT1 260.8549 0.0000 0.1218 8.4628 0.0000 0.4107
CAIT2 736.0613 0.0000 0.2289 15.2931 0.0000 1.2066
CAJP1 273.4044 0.0000 1.1353 3.5625 0.0080 0.4504
CAJP2 205.0926 0.0000 1.1700 3.3339 0.0117 0.5027
CAUK1 280.1213 0.0000 1.1268 21.8905 0.0000 0.3857
CAUK2 242.9319 0.0000 1.1096 6.2060 0.0001 0.5817
CAUS1 286.7682 0.0000 0.1182 15.4751 0.0000 0.4037
CAUS2 316.0419 0.0000 1.0988 8.2575 0.0001 0.4279
FRGR1 2017.3321 0.0000 0.4850 39.5745 0.0000 1.1219
FRGR2 290.9946 0.0000 0.7832 16.2750 0.9964 0.8074
FRIT1 983.0594 0.0000 0.6914 35.3386 0.0000 1.0014
FRIT2 530.1332 0.0000 0.1829 20.6363 0.0000 1.2042
FRJP1 2160.8583 0.0000 0.4770 48.6139 0.0000 1.1151
FRJP2 199.4415 0.0000 1.0736 0.0185 0.9993 0.4730
FRUK1 2139.9850 0.0000 0.4794 56.9604 0.0000 1.0577
FRUK2 255.1859 0.0000 0.9347 2.0882 0.0843 0.5425
FRUS1 2221.8645 0.0000 0.4704 54.8573 0.0000 1.0666
FRUS2 339.6425 0.0000 0.1197 6.3342 0.0843 0.3880
GRIT1 301.3993 0.0000 0.8540 6.9836 0.0000 0.7881
GRIT2 735.5024 0.0000 0.6535 16.3525 0.0000 1.1293
GRJP1 301.7980 0.0000 0.8540 4.1894 0.5239 0.7702

Model Test Statistic p-Values RMSE Test Statistic p-Value RMSE
GRJP2 193.2506 0.0000 1.1915 2.9310 0.0221 0.4795
GRUK1 313.5724 0.0000 0.8372 7.3175 0.0000 0.7536
GRUK2 241.2715 0.0000 1.0500 18.1193 0.0000 0.4684
GRUS1 377.9493 0.0000 0.7678 15.6002 0.0000 0.7434
GRUS2 307.2726 0.0000 1.0590 19.0447 0.0000 0.1139
ITJP1 788.0186 0.0000 0.6316 29.3709 0.0000 1.0908
ITJP2 199.9694 0.0000 1.1673 1.8838 0.1153 0.4927
ITUK1 661.9456 0.0000 0.6871 17.9637 0.0000 0.5019
ITUK2 250.8894 0.0000 3.7961 17.9637 0.0000 0.5019
ITUS1 667.8976 0.0000 0.6848 19.7459 0.0000 1.1672
ITUS2 332.7416 0.0000 1.0189 12.6557 0.0000 0.3922
JPUK1 169.5096 0.0000 1.3045 3.4933 0.0090 0.5052
JPUK2 184.0335 0.0000 1.2437 5.4035 0.0004 0.5915
JPUS1 185.8725 0.0000 1.2511 13.8287 0.0000 0.5219
JPUS2 306.3379 0.0000 1.0978 7.2141 0.0000 0.4275
UKUS1 180.6804 0.0000 1.2559 10.3327 0.0000 0.5658
UKUS2 294.7611 0.0000 1.1199 4.8548 0.0010 0.4380
Notes:
1. Column 1 in this Table shows the names of the neural network models employed for
approximating in-sample as well as jackknife out-of-sample forecasts with bivariate
linear models.
2. Test statistics presented in this Table are constructed using in-sample and jackknife out-
of-sample forecasts from bivariate linear models and those approximated from the
relevant neural networks using the following two Equations:
yt = π ' wt + ut
(3.1a)
where,
u t ~ N (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y , x ,..., x )'
w
t t t −1 t− p t −1 t− p
and π = ( π 0 , π 1 ........, π p )'

k
ut = π ' wt + ∑θ
j =1
0 j {ψ (γ j ' wt )} + vt
(3.1b)
Test statistics shown in columns 2 and 5 are constructed respectively using in-
sample and jackknife out-of-sample forecasts that are distributed
F [m, (n − p − m − 1)] under normality hypothesis. The test statistic is calculated

using Equation 3.1 c .
TS = {(SSE1 − SSE 2) / m} /{SSE 2 / (n − p − m − 1)}
(3.1c)
where, in Equation 3.1 c , m denotes the number of restrictions, n the number of
observations, and p the number of lags in the linear and ANN models. In Equation
3.1 c RSS i (for i = 1, 2 ) are residual sum of squares from the first equation whereas
SSE i show squared residual sum from the second part of the neural network linearity
test.
3. Test statistics for bivariate linear models using in-sample forecasts are presented in
column 3 and those with jackknife out-of-sample forecasts are shown in column 5 .
The relevant p-values for each test statistics are juxtaposed with each test statistics in
parenthesis.
4. Columns 4 and 6 show root mean squared errors (RMSE) respectively for in-sample
and jackknife out or sample forecast for each of the neural network model
approximated.
The chief hypothesis of this study is linearity versus the alternative hypothesis
of nonlinearity for testing possible existence of asymmetries in Canada, France,
Germany, Italy, Japan, UK, and USA real GDP growth. The linearity hypothesis
is based on the test statistic that is constructed from in-sample forecast
approximated from neural network models with its linear as well bivariate linear
counterparts for Canada, France, Germany, Italy, Japan, UK, and USA real GDP
growth rates. The linearity hypothesis is also tested using jackknife out-of-sample
forecasts from linear models as well as ANN approximations for all the series.
When the null hypothesis of linearity is true, nonlinearity does not prevail in
the series being tested, alternately when the null is false, nonlinearities do prevail
in the data series being tested. The test statistic for each of the series is calculated
using Equation 2.3 . This test statistic is distributed F [m, (n-p-m-1)] under
linearity hypothesis when considering neural network linearity test using
univariate linear models. Where, in Equation 2.3 , m is the number of restrictions
in the nonlinear model, n is the total number of observations in the series being
tested, and p is the numbers of lags used in each regression. Likewise, neural
network linearity tests are also constructed in conjunction with bivariate linear
models using in-sample as well as jackknife out-of-sample forecasts from each of

the model estimated using Equation 2.4 .
3.4.Results on Hypothesis Test
Results on hypotheses tests for neural network linearity tests that are based on
in-sample as well jackknife out-of-sample forecasts from univariate linear models,
and neural networks are presented in Table 2 . Similarly, the results on hypotheses
tests for neural network linearity tests that are constructed from bivariate linear
models in conjunction with neural network are shown in shown in Table 3 .
Neural network test statistics presented in Table 2 , column 2 , row 1 for
Canada is statistically significant at 5 percent level of significance based on p-
values that are shown in column 3 row 1 of this Table. Inferences do not change
when significance level is changed from 5 percent to 10 percent level. Neural
network test statistics for the remaining countries i.e. France, Germany, Italy,
Japan, UK, and USA are also significant at 5 as well as 10 percent level of
significance. These results reveal that a statistically significant evidence of
business cycle asymmetries does exists in Canada, France, Germany, Italy, Japan,
UK, and USA real GDP growth rates.
For the purpose of detecting possible existence of nonlinearities in all the
series, these results would suffice, however, this analysis is extended to jackknife
out-of-sample framework for possible existence of nonlinearities in all the series
after accounting for linkages between these countries in the form of contagion,
and spillovers that might be effecting business cycle asymmetries in these
countries in bivariate framework. Test statistics for all the series are statistically
significant at all levels based on the relevant p-values. Thus, the results on neural
network linearity tests show existence of a robust evidence of asymmetries in
business cycle fluctuations in Canada, France, Germany, Italy, Japan, UK, and
USA real GDP growth rates.
While the objective of the present study is to detect possible existence of
business cycle asymmetries in Canada, France, Germany, Italy, Japan, UK, and
USA series, contagion effects in G7 countries in bivariate framework is also
focused. Inclusion of such tests in this analysis is intended to test additional
evidence of nonlinearities in all the series (if any) in addition to exploring possible
existence of contagion effects and to observe how one country’s business cycle
would affect the magnitude of business cycle fluctuation of the other country. For
example, for the bivariate model CAUS, the neural network linearity test statistic
shown in Table 2 column 2 row 12 indicates that USA business cycle is not
affected much when Canadian GDP series is included in the bivariate model,
however, test statistic in the previous row (row11) in same column shows that
inclusion of USA series in bivariate model for Canada reduces the magnitude of
test statistic showing that USA economy has more pronounced impact on the
Canadian business cycles. This can be attributable to the fact that despite the size
of their economies, USA, and Canada being neighboring countries depend heavily
on each other’s products. For example, about 84 percent of Canada’s export
goods are sent to USA and 57 percent of Canada’s import come from USA.
Similarly, USA imports about 13 percent of its total import from Canada, and 23
percent of USA’s total export go to Canada. Using test statistic for the remaining
bivariate models shown in the Table, one can analyze the impact of France,
Germany, Italy, Japan and UK series on Canadian business cycles, and impact of
Canadian economy on these countries’ business cycles. Similarly, impact of
business cycle fluctuations among other countries included in the study can also
be analyzed in a similar manner.
Neural network nonlinearity test for detecting possible existence of business
cycle nonlinearities using in-sample forecasts from neural networks with
univariate linear, and bivariate linear models show statistically significant
evidence of asymmetries in all the series. Similarly, neural network nonlinearity
tests constructed from jackknife out-of-sample forecasts from neural networks,
and its linear counterparts also show statistically significant evidence of business
cycle asymmetries in Canada, France, Germany, Italy, Japan, UK, and USA real
GDP growth rates. However, in four out of forty cases that are studied using
jackknife out-of-sample forecasts, neural network linearity tests fail to reject
linearity hypothesis at all level whereas in two out of forty two cases linearity
could not be rejected at conventional (5 percent) level of significance.
3.5. Forecast Performance of Neural Networks
Figure 2 shows plots of in-sample forecasts from linear model against ANN
approximations for Canada, France, Germany, Italy, Japan, UK, and USA real
GDP growth rates. Comparing linear versus ANN forecasts this Figure shows that
neural network models explain data series better than those of linear models for
Canada, France, Germany, Italy, Japan, UK, and USA real GDP growth rates.
Figure 2. Forecasts from Linear Models and Neural Networks.
Table 2 shows in-sample root mean square error (RMSE) computed from in-
sample as well as jackknife out-of-sample forecasts from neural network models
for all the series. In this Table, RMSE for in-sample forecasts are shown in
column 4 rows 1 to 7 and RMSE for jackknife out-of-sample forecasts are
shown in column 7 rows 1 to 7 respectively for Canada, France, Germany, Italy,
Japan, UK, and USA. Comparing RMSE calculated from in-sample to jackknife
out-of-sample approximations from ANN for Canada for example, it transpires
that jackknife out-of-sample forecast performance of neural network models is not
superior to their in-sample forecasts. Similar results (not discussed for brevity) are
obtained for the remaining series studied.
4. CONCLUSION
The present work considers the usefulness of artificial neural networks to find
possible existence of business cycle asymmetries in Canada, France, Germany,
Italy, Japan, UK, and USA real GDP growth rates. Possible existence of business
cycle asymmetries in all the series are studied using in-sample, and jackknife out-
of-sample forecasts that are estimated from univariate and bivariate linear models
in conjunction with their relevant ANN models.
The results based on in-sample approximations from artificial neural
networks, and other models show robust evidence of asymmetries in business
cycle fluctuations in Canada, France, Germany, Italy, Japan, UK, and USA series.
Likewise, the results based on jackknife out-of-sample forecasts strengthens the
evidence for possible existence of business cycle asymmetries in all the series.
These results reveal that the impact of monetary policy or any other shock on
output in these countries cannot be determined that are based on forecasts
obtained from linear models, and linear vector autoregressions.
The results reveal that in-sample forecasts from neural networks are superior
to their jackknife out-sample forecasts. In addition, results from bivariate analysis
show spillover and contagion effects among G7 countries in bivariate framework.
While these linkages can be viewed as trade and other linkages, and trade linkages
between USA, and Canada business cycles can easily be explained because these
neighboring highly industrialized countries depend heavily on each other’s
products. Moreover, North American Free Trade Agreement (NAFTA) that
includes Mexico also facilitates trade between these countries. Similar type of
linkages can also be noticed between other countries in bivariate framework.
However, it might be expeditious to determine the linkages other than trade
linkages between these countries that might be effecting business cycle
fluctuations when using bivariate framework.
Neural network models outperform the traditional statistical tests for
remaining nonlinearities, which is in line with the previous studies (Terasvirta, et
al. 1993). However, in-sample forecast performance of neural network models is
superior to univariate linear models (Kiani et al. 2005). Additionally, the effects
of one country’s business cycle fluctuations on the other country in bivariate
framework can be viewed as business cycle linkages or contagion effects due to
trade and other linkages between them. For example the size of an economy and
magnitude of trade between any two countries appear to be two important factors
to determine high or a low magnitude of business cycle linkages or contagion
effects.
It might be of interest if future research employs additional models, and tests
using data from G7 countries to explore the nature of trade and other linkages and
their share in these countries in addition to detecting the impact of these linkages
on economic fluctuations within these economies.
5. REFERENCES
Anderson, H., & F. Vahid, (1998). Testing Multiple Equation Systems for
Common Nonlinear Components. Journal of Econometrics, 84, 1-36.
Anderson, H., & Ramsey, J. (2002). U.S. and Canadian Industrial Production
Indices as Coupled Oscillators. Journal of Economic Dynamics and Control,
26, 33-67.
Andreano, M., & Savio, G. (2002). Further Evidence on Business Cycle
Asymmetries in G7 Countries. Applied Economics, 34, 895-904.
Andrews, D. (2001). Testing When Parameter is Under the Boundary of the
Maintained Hypothesis. Econometrica, 69, 683-734.
Axelrod, R. (1987). The Evolution of the Strategies in the Iterated Prisoner’s
Dilemma. In Algorithm and Simulated Annealing, Davis, L. D. (Eds.), Los
Altos, CA: Morgan Kaufmann. 32 - 41.
Auerbach, A. (1982). The Index of Leading Indicators: Measurement without
Theory, Thirty-Five Year Later. Review of Economics and Statistics, 64, 589-
595.
Beaudry, P., & Koop, G. (1993). Do Recessions Permanently Change Output?
Bidarkota, P. (1999). Sectoral Investigation of Asymmetries in the Conditional
Mean Dynamics of the real US GDP. Studies in Nonlinear Dynamics and
Bidarkota, P. (2000). Asymmetries in the Conditional mean Dynamics of Real
GNP: Robust Evidence. The Review of Economics and Statistics, 82, 153-157.
Brunner, A. (1992). Conditional Asymmetries in Real GNP: A Semi-
Nonparametric Approach. Journal of Business & Economics Statistics, 10,
65-72.
Brunner, A. (1997). On the Dynamic Properties of Asymmetric Models of real
Davies, R. (1977). Hypothesis Testing When a Nuisance Parameter is Present
under the Alternative. Econometrica, 69, 683-733.
DeLong, B., & Summers, L. (1986). Are Business Cycles Symmetrical? In The
American Business Cycle: Continuity and Change, Gordon, R.J. (Eds.),
Chicago IL: Chicago University Press. 166-179.
Diebold, F., & Rudebusch, G. (1990). A Nonparametric Investigation of Duration
Dependence in the American Business Cycle. Journal of Political Economy,
98, 596-616.
Estrella, A., & Mishkin, F. (1998). Predicting U.S. Business Cycle Regimes.
Falk, B. (1986). Further Evidence on the Asymmetric Behavior of Economic
Time Series over the Business Cycle. Journal of Political Economy, 94, 1096-
1109.
Garcia, R. and R. Gencay, (2000). Pricing and Hedging Derivative Securities with
Neural Networks and a Homogeneity Hint. Journal of Econometrics, 94, 93-
115.
Gencay, R. (1999). Linear, Nonlinear and Essential Foreign Exchange Prediction
with Simple Technical Trading Rules. Journal of International Economics,
47, 91-107.
Learnings. Readings, MA: Addison Wisely.
Gordon, R. (1986). The American Business Cycle: Continuity and Change. NBER
Studies in Business Cycles, Vol. No. 25, Chicago Illinois: University of
Chicago Press.
Hamilton, J. (1989). A New Approach to the Economic Analysis of Nonstationary
Time Series and Business Cycle. Econometrica, 57, 357-384.
Hutchinson, J., Lo, A., and Poggio, T. (1994). A Nonparametric Approach to
Pricing and Hedging Derivative Securities via Learning Networks. Journal of
Finance, 49, 851-889.
Kiani, K. (2005). Detecting business cycle asymmetries using artificial neural
network and time series models, Computational Economics, 26, 65-85.
Kiani, K., Bidarkota, P., & Kastens, T. (2005). Forecast Performance of Neural
Networks and Business Cycle Asymmetries. Applied Financial Economics
Letters, 1, 205-210.
Kiani, K., & Bidarkota, P. (2004). On Business Cycle Asymmetries in G7
Countries. Oxford Bulletin of Economics and Statistics, 66, 333-351.
Klein, P. (1990). Analyzing Modern Business Cycles: Essays in Honor of
Geoffery H. Moore. M.E. Sharpe, M.E. (Eds.). NY: Armonk.
Kling, J. (1987). Predicting the Turning Points of Business and Economic Time
Series. Journal of Business, 60, 201-238.
Koch, P., & Rasch, R. (1988). An Examination of the Commerce Department
Leading Indicator Approach. Journal of Business and Economic Statistics, 6,
167-187.
Kuan, C., & White, H. (1994). Artificial Neural Networks: An Economic

Perspective. Econometric Review, 13, 1-91.
Marimon, R., McGratten, E., & Sargant, T. (1990). Money as Medium of
Exchange in an Economy with Artificially Intelligent Agents. Journal of
Economics Dynamics and Control, 14, 329-373.
Neftci, S. (1984). Are Economic Time Series Asymmetric over the Business
Cycle? Journal of Political Economy, 92, 307-328.
Politis, D., & Romeo, J. (1994). Large Sample Confidence Region Based on Sub-
Sample under Minimal Assumption. Annals of Statistics, 22, 2031-2052.
Politis, D., Romeo, J. & Wolf, M. (1997). Sub-Sampling for Heteroskedastic
Time Series. Journal of econometrics, 81, 281-317.
Potter, S. (1995). A Non-Linear Approach to U.S. GNP. Journal of Applied
Qi, M. (2001). Predicting U.S. Recessions via Leading Indicators via Neural
Network Models. International Journal of Forecasting, 17, 383-401.
Qi, M., & Madala, G. (1999). Economic Factors and the Stock Market: A New
Perspective. Journal of Forecasting, 18, 151-166.
Quenouille, M. (1949). A note on Bias in Estimation. Biometrika, 43, 353-60.
Ramsey, J., & Rothman, P. (1996). Time Irreversibility and Business Cycle
Asymmetry. Journal of Money Credit and Banking, 28, 1-21.
Reilly, D., & Cooper, L. (1990). An Overview of Neural Networks: Early Models
to Real World Systems. In An Introduction to Neural and Electronic
Networks, Zornetzer, J., Davis, L., & C. Lau. C. ed., New York: NY.
Sichel, D. (1989). Are Business Cycles Asymmetric? A Correction. Journal of
Swanson, N., & White, H. (1995). A Model Selection Approach to Assessing the
Information in the Term Structure Using Linear Model and Artificial Neural
Networks. Journal of Business Economics and Statistics, 13, 265-275.
Swanson, N., & White, H. (1997a). A Model Selection Approach for Real Time
Macroeconomic Forecasting Using Linear Models and Artificial Neural
Networks. Review of Economics and Statistics, 79, 540-550.
Swanson, N., & White, H. (1997b). Forecasting Economic Time Series Using
Flexible vs. Fixed and Linear vs. Nonlinear Economic Models. International
Journal of Forecasting, 13, 439 - 461.
Terasvirta, T., Lin, C., & Granger, C. (1993). Power of Neural Networks Test.
Tuckey, J. (1958). A Bias and Confidence in Not-Quite Large Samples. Annals of
Vishwakarma, K. (1995). A Neural Networks to Forecast Business Cycle

Indicators. Mathematics and Computers in Simulations, 39, 287-291.
White, H. (1989a). Some Asymptotic Results for Learning in Single Hidden-
Layer Feed Forward Network Models. American Statistical Association, 84,
1003-1013.
White, H. (1989b). Learning in Artificial Neural Networks: A Statistical
Perspective. Neural Computations, 1, 425-464.
Wu, C. (1990). On the Asymptotic Properties of the Jackknife Histogram. Annals
of Statistics, 18, 1438-1452.
Ziari, H., Leatham, D., & Ellinger, P. (1997). Developing of Statistical
Discriminant Mathematical Programming Model via re-Sampling Estimation
Techniques. American Journal of Economics, 79, 1352-1362.
Chapter 5
BUSINESS CYCLE ASYMMETRIES IN ASIAN

ECONOMIES VIA NONLINEAR TIME SERIES
MODELS AND NEURAL NETWORKS
ABSTRACT
The present work investigates possible existence of asymmetries in
business cycle fluctuations in the selected Asian economies that include
Armenia, Indonesia, Japan, Kazakhstan, Malaysia, Philippines, Russia,
Singapore, Thailand, and Turkey real GDP growth rates. Asymmetries in
these countries are modeled using nonlinear augmented and switching time
series models. These models account for the presence of outlier, time varying
volatility, and long memory that may be present in the series. Finally neural
network linearity tests, and neural network tests for neglected nonlinearities
are employed for testing existence of nonlinearities, and neglected
nonlinearities (if any) in all the series.
Results based on various tests constructed from nonlinear time series
models show statistically significant evidence of business cycle asymmetries
in Armenia, Indonesia, Japan, Malaysia, Philippines, Singapore, Thailand,
and Turkey real GDP growth rates. Further, the neural network linearity tests
overwhelmingly reject linearity hypothesis in all the series. Moreover, the
results for Japan are in line with Kiani and Bidarkota (2004), and Kiani
(2005). Finally, the results show that linear models cannot be employed to
forecast the impact of monetary policy or any other shock on output in any of
the countries studied.
Key phrases: Real GDP growth rates; fat tails; stable distributions; neural
networks; out-of-sample forecasts; long memory; nonlinearities; business
cycles
JEL Codes: C22, C32, C45, C53
1. INTRODUCTION
Detecting possible existence of business cycle asymmetries in
macroeconomic time series is important for several reasons. For instance
nonlinearities imply that the effects of expansionary and contractionary monetary
policy and other shocks on output are asymmetric. Therefore, any nonlinearity
would invalidate the measures of the persistence of monetary policy or any other
shocks on output that is based on linear models. The behavior of economic
variables during different phases of business cycles is important for policymakers
who might be interested in anticipating the impact of monetary policy or any other
shocks on output for timely policy action that might help avoid downturns in the
economy.
Studies by Neftici (1984), Brunner (1992, 1997), Beaudry and Koop (1993),
Potter (1995), and Ramsey and Rothman (1996), Anderson and Vahid (1998),
Bidarkota (1999-2000), Anderson and Ramsey (2002), Kiani and Bidarkota
(2004) provided evidence of business cycle asymmetries in real GDP growth rates
in G7 countries. With the fair idea that business cycle fluctuations are asymmetric
in developed countries it would be of interest to investigate the behavior of
business cycle fluctuations in developing economies especially Asian economies.
This is because Asian countries comprise of a group of countries1 that encompass
highly developed economies like Japan, fastest growing economies including
Malaysia, Singapore, and Indonesia, and transition economies like Kazakhstan
and Russia, oil rich countries like Saudi Arabia, Kuwait, and Kazakhstan as well
as a number of developing and less developed countries. However, the data
limitations hampered the present empirical exercise to a small number of
1
Afghanistan, Armenia, Azerbaijan, Bahrain, Bangladesh, Bhutan, Brunei, Burma, Cambodia, China,
Georgia, Hong Kong, India, Indonesia, Iran, Iraq, Israel, Japan, Jordan, Kazakhstan, Korea,
North, Korea, South, Kuwait, Kyrgyzstan, Laos, Lebanon, Malaysia, Maldives, Mongolia,
Myanmar, Nepal, Oman, Pakistan, Philippines, Qatar, Russia, Saudi Arabia, Singapore, Sri
Lanka, Syria, Taiwan, Tajikistan, Thailand, Turkey, Turkmenistan, United Arab Emirates,
Uzbekistan, Vietnam, and Yemen.
Business Cycle Asymmetries in Asian Economies… 87
countries2 because seasonally adjusted quarterly data for most of these countries
was not available. However, despite data limitations the sample employed
comprises of fastest growing economies, transition economies, highly developed
economy and developing economies. Indeed, the present empirical exercise would
help researchers, and policymakers to anticipate the impact of monetary policy or
other shock(s) on output particularly in the countries studied using appropriate
empirical models, which would help employing appropriate stabilizing policies
when needed. For instance linear models would be adequate for forecasting output
if the underlying data generating process is symmetric else nonlinear models
would be considered adequate. Taking appropriate policy measures would help
avoiding anticipated downturns, and likewise timely policy action might help
avoiding anticipated future financial crises (if any).
There is a growing perception that evidence of nonlinearity reported by
several studies could be because of the presence of outliers in the data. In this
context Tsay (1988) showed that linearity could be rejected due to the presence of
outliers in the data, Blanchard and Watson (1986) concluded for the presence of
outliers in GNP data, and Balke and Fomby (1994) and Scheinkman and LeBaron
(1989) demonstrated a weak evidence of linearity in US real GNP data once
outlier were taken into account.
There is a strong consensus that non-normality as well as time varying
volatility is present in macroeconomic time series data. However, most studies on
business cycle asymmetries do not encompass these features in their models3.
Inefficient estimation would result when such features are not included in the
models that are employed to detect business cycle asymmetries in macroeconomic
time series.
The present chapter investigates whether asymmetries in business cycle
fluctuations are present in Armenia, Indonesia, Iran, Japan, Kazakhstan, Malaysia,
Philippines, Russia, Singapore, Thailand, and Turkey real GDP growth rates using
nonlinear time series models that encompass long memory, time varying
volatility, and stable distributions. Moreover, considering the difficulties faced by
Andreano and Savio (2002), and Kiani and Bidarkota (2004) in detecting business
cycle asymmetries in G7 countries, the present work employs artificial neural
networks that are considered to be highly flexible functional forms of nonlinear
models to study possible existence of business cycle asymmetries (if any) in real
GDP growth rates in all the series. Kuan and White (1994) and Swanson and
2
Armenia, Indonesia, Iran, Japan, Kazakhstan, Malaysia, Philippines, Russia, Singapore, Thailand
and Turkey
3
Exception is Bidarkota (2000), and Kiani and Bidarkota (2004).
White (1995, 1997 a, 1997b) employed artificial neural networks in economic

time series data. Likewise, Vishwakarma, (1995), Qi, (2001), Kiani (2005), and
Kiani et al. (2005) used artificial neural networks in business cycles.
The remaining chapter is organized as follows. Section 2 elaborates nonlinear
time series models, and artificial neural networks. Estimation results and
hypotheses of interest are discussed in section 3 . Section 4 presents conclusion.
2. EMPIRICAL MODELS
In the present work nonlinear time series models and artificial neural
networks are employed for possible existence of business cycle asymmetries in all
the series. The model parameterizations for all the series are selected using
Schwarz Bayesian Criterion (SBC) due to Schwarz (1978). Empirical models
employed in this study are elaborated in the following sub-sections.
2.1. Time Series Models
Two classes of time series models are employed to detect possible presence of
business cycle asymmetries in Armenia, Indonesia, Iran, Japan, Kazakhstan,
Malaysia, Philippines, Russia, Singapore, Thailand and Turkey real GDP growth
rates. The models employed are termed as CDR-Augmented, and SETAR-
Switching models. Within each class of these models, three different versions of
the models are entertained. Model 1 incorporates stable distributions, conditional
heteroskedasticity, and fractional differencing parameters. Restricting fractional
differencing in Model 1 yields Model 2. Imposing homoskedasticity on Model 2
gives Model 3. Model 1 is the most general model, which nests Models 2, and 3.
Each of these classes of models is elaborated in the following three sub-sections.
2.1.1. CDR Augmented Models

CDR-Augmented model was initially proposed by Beaudry and Koop (1993)
who estimated an autoregressive moving average (ARMA) model, augmented
with an ad hoc nonlinear term. The most general model (Model 1) in this class of
models encompasses stable distributions, conditional heteroskedasticity, and long
memory. Following Kiani and Bidarkota (2004), the model employed in the
present work in shown in the following Equations:
Φ ( L)(1 − L) d (Δy t − μ ) = [Ω( L) − 1]CDRt + ε t (5.1a)

ε t | I t −1 ~ z t ct , z t ~ i.i.d . S α (0,1)
ctα = b1 + b2 ctα−1 + b3 | ε t −1 |α (5.1b)
where, Δy t ≡ 100 * Δ(ln GDPt ) is the growth rate of GDP, μ is its unconditional
mean, and d is the differencing parameter. The polynomial Ω(.) is of orders r ,
and polynomial Φ (.) of the order p in the lag operator L , with Ω(0) = Φ (0) = 1 .
This ad-hoc nonlinear term is labeled as the current depth of a recession ( CDRt )
which measures the gap between the current level of output and the economy’s
historical maximum level that is defined as CDRt = max{ y t − j } j ≥ 0 − y t . The term
CDRt is designed to permit recessions to be less or more persistent than
expansions depending on the parameter estimates.
A random variable X will have a symmetric stable distribution S α (δ , c ) if
its log characteristic function can be expressed as ln E exp(iXt ) = iδt − | ct |α . Here
c ∈ [0, ∞] is the scale parameter, δ ∈ [−∞, ∞] is the location parameter, and
α ∈ [0,2] is the characteristic exponent governing tail behavior, and its small
values shows thicker tails. However, when α = 2 normal distribution results.
When α = 2 in Equation 1b , normal GARCH (1,1) process prevails.
However, when d = 0 there exists a unit root in yt , but with d = −1 , y t ends up
being integrated of order zero I (0). ARFIMA models with long memory are
defined in terms of the rate of decay of their autocovariances, so the extension of
these models to infinite variance stable shocks is not immediate.
A stationary casual and invertible solution to an ARFIMA model with
Gaussian errors requires | d |< 0.5 (Brockwell and Davis (1991). However,
according to Kokoszka and Taqqu (1995), MA (∞ ) representation to an
ARFIMA model with stable shocks needs α (d − 1) < −1 . Therefore, d needs to be
positive when α > 1 . In addition, α > 1 and | d |< (1 − 1 / α ) for the ARFIMA
model to be a solution to an AR (∞ ) process. That is the reason why α and d is
restricted in Equation 1 within these limits.
Equation 1a reduces to an autoregressive (AR) model with non-integer
differencing when Ω( L) = 1 . Because, it nests AR models, likelihood ratio (LR)
statistic can be used to test statistical significance of the non-linear term governing
the conditional mean dynamics. When autoregressive lag order p is 0 and r is 1,
ω1 = 0 yields a random walk with drift. However, a positive ω1 implies that
negative shocks are less persistent whereas a negative ω1 implies that positive
shocks are less persistent.
The existence of nonlinearities shows that innovations are symmetric but the
impulse transmission mechanism is nonlinear, or the innovations are asymmetric
but the impulse transmission mechanism is linear, or the innovations are
asymmetric and the impulse transmission mechanism is also nonlinear. However,
it is difficult to separate the asymmetric innovations from the nonlinear
propagation mechanism, if they do exist in a data series. Although asymmetric α -
stable distributions exist and are well defined, the present work is merely
investigating possible existence of asymmetries in the conditional mean regardless
of how they can be characterized.
2.1.2. SETAR-Switching Models

The SETAR-switching model or self-exciting threshold autoregressive
(SETAR) model governs switching between the regimes that are defined in terms
of the observed series yt . Potter (1995) used SETAR-Switching models with
errors to come from a normal family, and thereafter Bidarkota (2000), and Kiani
and Bidarkota (2004) estimated SETAR-Switching models with long memory,
conditional heteroskedasticity, and stable distribution. The most general model
estimated within this class of models is shown in the following Equations:
In Regime 1:
(1 − φ1 L − φ2 L2 )(1 − L) d (Δyt − μ1 ) = ε t (5.2a)
ε t | I t −1 ~ zt ct z ~ iid Sα (0,1)
ctα = b1 + b2 ctα−1 + b3 | ε t −1 |α (5.2b)
In Regime 2:
(1 − φ3 L − φ4 L2 )(1 − L) d (Δyt − μ 2 ) = ε t (5.2c)

ε t | I t −1 ~ zt γct z ~ iid Sα (0,1)
ctα = b1 + b2 ctα−1 + b3 | ε t −1 / γ |α 5.(2d)
The switch between the two regimes in the model presented in Equations
5.2 a − d is governed by the term Δy t − d > r , where y t is log real GDP, d is the
delay, and r is the threshold parameter. Therefore, when, Δy t − 2 > 0 regime 1
prevails, but when Δy t − 2 ≤ 0 , regime 2 is obtained.
2.3. Artificial Neural Networks
An artificial neural network (ANN) consists of a number of interconnected

elements called neurons. The ANN are able to learn from examples that can be
employed to solve unknown problems (Reilly and Cooper 1990). ANN are
nonlinear, nonparametric models that are independent of the distributions of the
underlying data generating processes (White 1989) which enables them to
approximate any continuous function with desired level of precision (Hornick et
al. 1989). Although ANN are criticized heavily for being “Black Boxes” because
it is hard to know their exact functional form as well as overfitting issues
associated with them. However, the present work is focused on final prediction
( Yhat ) rather than parameter estimates so the issue pertaining to the functional
form is of least importance. Nevertheless, following Kiani (2005), overfitting can
be mitigated using careful construction of neural network. A general form of
neural network employed is shown in Equation 5.3.
⎡ n ⎛ k ⎞⎤
f ( x) = sig ⎢α 0 + ∑ α j sig ⎜⎜∑β ij xi + β 0 j ⎟⎥ + ε
⎟⎥
(5.3)
⎣⎢ j =1 ⎝ i =1 ⎠⎦
where, n is the number of hidden nodes in the network, k is the number of

explanatory variables in the network, sig (x) = 1/(1+e-x) is a transfer function, that
can either be sigmoid (logistic) or hyperbolic (tangent) cumulative distribution
function, α j represents a vector of parameters or weights that link the hidden
node to the output layers’ units, β ij (i = 1 , … , k; j = 1 , … , n) denotes a matrix
of parameters linking the input to the hidden layers’ units, and ε is the error term.
A neural network test for detecting possible existence of nonlinearities was

originally proposed by Terasvista et al. (1993) which was also employed by Kiani
(2005) including others. The test comprises of a linear model, and artificial neural
networks. The linear model is shown in Equation 5.4 whereas artificial neural
network is presented in Equation 5.5 .
yt = π ' wt + u t (5.4)
where, ~ ' )' ,

wt = (1, wt
~ = ( y ,..., y )' ,
wt t −1 t− p π = (π 0 , π 1 ,........,π p )' and
u t ~ Nid (0, σ 2 ) .
⎡ n ⎛ k ⎞⎤
uˆt = π ' wt + sig ⎢α 0 + ∑ α j sig ⎜⎜∑β ij xi + β 0 j ⎟⎥ + ε
⎟⎥ (5.5)
⎣⎢ j =1 ⎝ i =1 ⎠⎦
where in equation 5.5 , π 0 is intercept, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 is a transfer

function, û t represents residuals from linear model, and the residuals from neural
network are represented by v̂t . The test statistic of the form shown in Equation
5.6 is constructed using residuals and residual sum of squares from Equation 5.4 ,
and 5.5 .
TS = [(SSE1 − SSE2 ) / m] /[SSE2 / (n − p − m − 1)] (5.6)
where, in Equation 5.6 , p is the number of lags in the series, m denotes the
number of restrictions in the unrestricted model, and n is the number of
observations. The test statistic is distributed approximately F4 under normality
hypothesis with ( n − p − m − 1 ), and m degrees of freedom.
Neural network linearity test are constructed using in-sample forecasts from
linear models as well as neural networks. Likewise, neural network linearity test
are also constructed using jackknife out-of-sample5 forecasts from linear models
4
See Davies, R. (1977) and Andrews, W. (2001) for discussions on the nuisance parameter that
appears under the alternative hypothesis.
5
Sub-sample jackknife that is proposed by Wu (1990) and extended by Politis and Romeo (1994) are
employed in the present work. Sub-sample jackknife was also used by Politis et al. (1997), and
Ziari et al. (1997) including others.
as well as neural networks. The test statistic for this test is constructed using in-
sample as well as jackknife out-of-sample forecast from linear models, and
artificial neural networks that nest the linear model shown in Equation 5.4 .
To capture asymmetric fluctuations in the series Beaudry and Koop (1993)

added an additive nonlinear term in an autoregressive moving average (ARMA)
framework. For this purpose, Beaudry and Koop (1993) assumed that the shocks
are identically, and independently distributed (iid) normal. Further, they
confirmed that the additive ad hoc nonlinear term does not impose any estimation
problems. Alternatively, Bidarkota (1999-2000), and Kiani and Bidarkota (2004)
assumed that innovations follow more general stable distributions so that
computational algorithm due to McCulloch (1997) could be employed to obtain
stable densities for maximum likelihood estimation of the employed models
which works well particularly when the errors are symmetric stable. However, the
full information maximum likelihood (ML) method proposed by Sowell (1992a)
can only be adopted if the errors are iid normal, but in the present work more
complicated non-normal conditionally heteroskedastic models are used, therefore,
conditional sum of squares (CSS) estimator is employed that is discussed in Box
and Jenkins (1976) that was originally proposed by Hosking (1984). The CSS
procedure is equivalent to the full information MLE asymptotically as was noted
by Baillie et al. (1996), and it works better for complex models.
Following Kiani and Bidarkota (2004) we do not consider any moving
average (MA) terms in the specification of the model. Maximum likelihood
estimation (MLE) of mixed ARMA models with stable errors poses a challenge;
therefore, like Kiani and Bidarkota (2004), Whittle estimator due to Mikosch et al.
(1995), and minimum dispersion estimators due to Brockwell and Davies (1991)
are employed.
The artificial neural networks (ANN) employed here encompass a nested
linear model to construct neural network linearity tests, which poses difficulty in
convergence. To surmount this difficulty, genetic algorithm (GA)6 is selected as
an estimation algorithm for these complex models. The GA is considered to be the
most reliable estimation algorithm for estimating any nonlinear functional form
but it is notoriously slow. Therefore, to speed up the estimation process, and to
6
Initially, De Jong (1975) applied genetic algorithm, to mathematical optimization. Later, Goldberg,
(1989) employed it in biology, engineering, and operation research. In economics, the genetic
algorithm was employed by Axelord (1987), Marimon, McGratten and Sargent (1990), and
Dorsey and Walter (1995).
enhance the probability for obtaining global minima, estimation process is started
with a couple of random starts using GA. Thereafter, the parameter vector
encompassing the smallest sum of squares was chosen to run GA for upto 50,000
iteration, whereupon the parameter vector obtained from GA was used as the
starting conditions for Matlab’s fminsearch algorithm, which is a Nelder-Mead
simplex algorithm that worked well for closing on the final optimum.
3.1. Data Sources
Quarterly real GDP7 growth rates for Armenia, Indonesia, Iran, Japan,
Kazakhstan, Malaysia, Philippines, Russia, Singapore, Thailand, and Turkey were
obtained from DataStream. Table 1 provides additional information about that
data series on Asian economies that are included for testing asymmetries in
business cycles fluctuations in these series.
3.2. Specification Search
A wide-ranging specification search was performed for each country for all
the models for each of the two classes of models i.e. CDR-Augmented and
SETAR-Switching models, where the specification was done over all the
parameterization with lag orders or the autoregressive, and CDR t terms of three
or less to attain parsimony as was done by Beaudry and Koop (1993). However,
for SETAR-Switching models the search was performed with the autoregressive
lag polynomials in the two regimes restricted to be of orders (3,3), (2,2), (1,1), or
(0,0) only.
3.3. Hypotheses Test
Two types of hypotheses tests are employed in the present work. The first
type of hypotheses pertains to linearity tests in conditional mean, which relates to
augmented nonlinear time series models, and switching time series models.
7
Data limitations hampered our empirical exercise to eleven Asian Economies only.
However, the second type of hypotheses tests relates to neural network linearity
test. Both these tests are elaborated in the following paragraphs.
The first type of hypothesis of interest is the test for linearity in the
conditional mean, which pertains to SETAR-Switching models. Under the null
hypothesis of the linearity, the unconditional means in the two regimes i.e. μ 1 and
μ 2 , and the corresponding autoregressive coefficients (if present in the models
employed) in the two regimes are equal, and the scale ratio γ is equal to one.
Failing to reject the null hypothesis would result in one regime only, however, a
rejection of the null would result in two distinct regimes describing the GDP
growth rates. On the other hand for CDR-augmented models the asymptotic
distribution of t-test for significance of the nonlinear CDRt term, which is shown
in the Equation, 1a is non-standard especially when the dependent variable is
non-stationary (Hess and Iwata, 1977; Kiani and Bidarkota 2004). Therefore, the
present work refrains from constructing tests for linearity in conditional mean
based on these models.
The second type of hypothesis pertains to neural network linearity tests that
are constructed from in-sample forecasts from linear models as well as neural
networks approximations. In addition to linearity tests constructed from in-sample
forecasts, neural network linearity tests are also constructed using out-of-sample
forecasts from linear models as well neural networks. In the second set of
hypothesis, if the null hypothesis of linearity is rejected in favor of alternative
hypothesis of nonlinearity, the series under investigation is said to have
nonlinearity. However, if the null hypothesis is not rejected then nonlinearities do
not prevail in the series being studied.
Data Series Sample Period Observations

Armenia Quarterly Real GDP 1994Q4-2004Q1 37
Indonesia Quarterly Real GDP 1990Q1-2004Q1 56
Iran Quarterly Real GDP 1988Q1-2000Q1 48
Japan Quarterly Real GDP 1980Q1-2004Q2 97
Kazakhstan Quarterly Real GDP 1994Q1-2004Q1 40
Malaysia Quarterly Real GDP 1991Q1-2004Q1 52
Philippines Quarterly Real GDP 1981Q1-2004Q2 93
Russia Quarterly Real GDP 1993Q3-2004Q1 42
Singapore Quarterly Real GDP 1984Q3-2004Q2 79
Data Series Sample Period Observations

Thailand Quarterly Real GDP 1993Q1-2004Q2 45
Turkey Quarterly Real GDP 1987Q1-2003Q4 67
Table 2. Specification Search/Log Likelihood: CDR Augmented Models
Model 3 Model 2 Model 1

Armenia (3, 3) (3, 2) (3, 3)
172.682 170.363 170.097
Indonesia (3, 3) (3, 2) (3, 3)
336.870 333.238 328.901
Iran (3, 3) (3, 3) (1, 3)
288.465 288.447 197.312
Japan (3, 3) (3,3) (3, 3)
618.194 615.045 610.483
Kazakhstan (3, 2) (3, 2) (3, 2)
231.659 231.611 231.292
Malaysia (3, 3) (3, 3) (3, 3)
312.689 312.246 312.175
Philippines (3, 0) (3, 1) (3, 1)
482.579 481.979 481.887
Russia (3, 3) (3, 2) (3, 2)
240.228 240.122 170.176
Singapore (3, 3) (3, 3) (2, 3)
488.767 472.286 470.992
Thailand (3, 1) (2, 3) (3, 3)
268.537 268.049 266.273
Turkey (3, 2) (3, 1) (3, 0)
291.728 290.820 290.408
Notes:
1. For each country in columns 2, 3 , and 4 , the first row represents model
parameterization, and the second row show log likelihood estimates for the model.
For example for Armenia, the first row for each model (models 1-3) shows model
parameterization and the second row shows log likelihood estimates for each of the
models. The results for the remaining countries are presented in the similar manner.
2. The symbol “--“denotes that our nonlinear time series models failed to converge.
Table 3. Specification Search/Log Likelihood: SETAR Switching Models

Armenia (3, 3) (3, 3) (3, 3)
212.835 164.058 165.354
Indonesia (2, 2) (1, 1) (1, 1
333.599 333.852 328.305
Iran (3, 3) (3, 3) (3, 3)
280.809 280.193 279.308
Japan (1, 1) (2, 2) (2, 2)
614.997 613.895 603.364
Kazakhstan (3, 3) (3, 3) (3, 3)
223.364 223.108 222.850
Malaysia (3, 3) (3, 3) (3, 3)
302.221 299.004 298.136
Philippines (3, 3) (3, 3) (3, 3)
472.862 470.069 470.023
Russia (3, 3) (3, 3) (3, 3)
233.799 234.601 234.024
Singapore (3, 3) (3, 3) (3, 3)
1096.431 1077.185 1073.283
Thailand (3, 3) (3, 3) (3, 3)
256.778 257.169 256.499
Turkey (2, 2) (2, 2) (2, 2)
349.057 340.519 328.105
Notes:
1. For each country in columns 2, 3 , and 4 the first row represents model
parameterization, and the second row shows log likelihood estimates for the model.
For example for Armenia, the first row for each model (models 1-3) shows model
parameterization and the second row shows log likelihood estimates for the model.
The results for the remaining countries are presented in the similar manner.
Table 4. Test for Linearity: LR Tests Based on SETAR-Switching Models

Armenia -- 0.33462 80.27632
(0.953) (0.000)
Indonesia 8.2574 10.67372 8.80742
(0.016) (0.001) (0.003)
Iran 4.3048 5.31706 6.9144
(0.230) (0.149) (0.074)
Japan 12.83378 -- 6.84474
(0.000) (0.032)
Kazakhstan 4.67282 5.18034 5.13354
(0.197) (0.159) (0.162)
Malaysia 6.62028 10.979 10.97686
(0.085) (0.011) (0.011)
Philippines 10.17062 15.59288 15.47848
(0.001) (0.001) (0.001)
Russia 2.09926 0.80362 0.24182
(0.552) (0.848) (0.970)
Singapore -- -- 1214.6804
(0.000)
Thailand 8.57034 2.50884 3.41662
(0.035) (0.473) (0.331)
Turkey 26.33814 -- --
(0.000)
Notes:
1. For each country in columns 2, 3 , and 4 the number in the first row represents “test
statistic” and the number in the second row in parenthesis shows p-values for the test
statistics. For example for Armenia, the number in the first row for each of the models
(models 1-3) shows test statistic for that model and the number in the second row in
parenthesis shows p-values for the test statistic for that model. The results for the
remaining countries are presented in the similar manner.
2. The symbol “--“denotes that our nonlinear time series models failed to converge.
Table 5. Test for Linearity: Neural Network Tests
In-Sample Jackknife Out-of-sample

Test p-Values Test p-Values
Statistic Statistic
Armenia 393.502 (0.000) 22.237 (0.000)
Indonesia 71.662 (0.000) 0.517 (0.000)
Iran 108.912 (0.000) 67.164 (0.000)
Japan 109.722 (0.000) 55.676 (0.000)
Kazakhstan 59.121 (0.000) 36.524 (0.000)
Malaysia 64.148 (0.000) 46.217 (0.000)
Philippines 268.639 (0.000) 138.723 (0.000)
Russia 34.304 (0.000) 19.021 (0.000)
Singapore 1049.931 (0.000) 705.312 (0.000)
Thailand 63.222 (0.000) 44.707 (0.000)
Turkey 186.245 (0.000) 24.992 (0.000)
Notes:
1. Table presents neural network test results approximated from in-sample as well as
jackknife out-of-sample forecasts from linear models and neural network models.
2. In columns 2 and 4 , starting rows 1 to 11 this Table presents neural network test
statistics for Armenia, Indonesia, Iran, Japan, Kazakhstan, Malaysia, Philippines,
Russia, Singapore, Thailand, and Turkey respectively. p-values for each test statistic
are juxtaposed in parenthesis in next column.
3.4. Results on Hypotheses Tests
The empirical results on hypotheses tests for linearity in conditional mean as

well as neural network linarity tests for each of the series are performed as
elaborated in the earlier sub-sections. All tests are based on the likelihood ratio
(LR) test statistics that are reported in Table 4 . p-values for each test are
juxtaposed in parentheses in the next column. Results from neural network
linearity tests that are based on in-sample as well as jackknife out-of-sample
forecasts from all the series are reported in Table 5. p-values for each test statistic
are juxtaposed in the next column in parenthesis.
3.4.1. Results on Time Series Models

The results based on linearity in conditional mean for Indonesia, Japan,
Malaysia, and Philippines show strong evidence of asymmetries in business cycle
fluctuations. However, results on Iran, Kazakhstan, and Russia are largely linear.
On the other hand Armenia, Thailand, and Singapore show weak evidence of
business cycle asymmetries in real GDP growth rates. All statistical inferences for
these tests are drawn at 5 percent level of significance. However, the results do
not change when we change the significance level from 5 percent to 10 percent
level.
3.4.2. Results on Neural Network Tests

The neural network test for possible existence of nonlinearities tests the null
hypothesis of linearity against the alternative hypothesis of nonlinearities for
Armenia, Indonesia, Iran, Japan, Kazakhstan, Malaysia, Philippines, Russia,
Singapore, Thailand, and Turkey. Nonlinearity prevails in the data series being
tested when the null hypothesis of linearity is rejected against the alternate
hypothesis of nonlinearity.
Neural networks linearity test results based on in-sample forecasts from linear
models as well as neural network approximations show existence of statistically
significant evidence nonlinearities in real GDP growth rates of Armenia,
Indonesia, Iran, Japan, Kazakhstan, Malaysia, Philippines, Russia, Singapore,
Thailand, and Turkey. Similarly, neural network linearity test results based on
jackknife out-of-sample forecasts approximated from artificial neural networks
also support these results. All statistical inferences are drawn at the five percent
significance level. However, the results do not change when significance level is
changed from 5 to 10 percent level.
3.4.3. Results on Selected Parameter Estimates

Further to Kiani and Bidarkota (2004) the primary objective of this study is to
find possible existence of business cycle asymmetries in Asian economies.
However, data limitations hampered out analysis to a limited number of countries
because most countries included in the initial sample do not have sufficient
number of observation in quarterly series. Therefore, many countries8 including
India, and China that would be good examples of fastest growing economies, were
dropped from this empirical exercise, however, the sample used encompassed data
from highly developed countries, fastest growing countries, transition economies,
8
Afghanistan, Azerbaijan, Bahrain, Bangladesh, Bhutan, Brunei, Burma, Cambodia, China, Georgia,
Hong Kong, India, Iraq, Israel, Jordan, Korea, North, Korea, South, Kuwait, Kyrgyzstan, Laos,
Lebanon, Maldives, Mongolia, Myanmar, Nepal, Oman, Pakistan, Qatar, Saudi Arabia, Sri
Lanka, Syria, Taiwan, Tajikistan, Turkmenistan, United Arab Emirates, Uzbekistan, Vietnam,
and Yemen.
and other developing countries, which helped finding possible existence of

business cycle asymmetries in the selected Asian economies9.
Results for selected parameters from the most general switching models
estimated for each country are shown in Table 6. In this Table, columns 2 to 3
show parameter estimates for the switching parameter γ , characteristic exponent
α from SETAR-Switching models and parameter estimate for nonlinearity term
ω1 from CDR-Augmented for all countries starting Armenia in alphabetical order.
Standard errors for these parameter estimates are placed below each parameter
estimate in parenthesis.
Table 6. Selected Parameter Estimates for Most General Models
Models CDR-Switching SETAR-

Switching
Parameters of γ α ω1
interest
Armenia 3.0193 1.560 -2.097e-5
(1.604) (0.277) (0.011)
Indonesia 1.669 1.999 -0.035
(0.311) (0.000) (0.100)
Iran 1.279 1.999 0.810
(0.283) (0.005) (0.000)
Japan 0.964 1.999 -0.295
(0.154) (0.000) (0.334)
Kazakhstan 1.244 1.999 0.294
(0.385) (0.000) (0.023)
Malaysia 1.718 1.718 0.843
(0.143) (0.093) (0.954)
Philippines 1.246 1.400 -0.004
(0.197) (0.000) (0.013)
Russia 0.7301 1.999 0.001
(0.150) (0.000) (0.012)
Singapore 1.074 1.718 1.090
(0.166) (0.092) (0.374)
Thailand 2.383 1.652 0.028
(1.074) (0.203) (0.208)
9
Armenia, Indonesia, Iran, Malaysia, Philippines, Japan, Kazakhstan, Russia, Singapore, Thailand,
and Turkey,
Models CDR-Switching SETAR-

Switching
Parameters of γ α ω1
interest
Turkey 1.531 1.882 0.082
(0.086) (0.099) (0.090)
Notes:
1. Table presents selected parameter estimates in rows 1 to 11 respectively for Armenia,
Thailand, and Turkey. Standard error for each of the parameter estimates are shown in
parenthesis beneath each parameter estimates.
2. In this Table column 2 to 4 show parameter estimates for switching parameter gamma,
and characteristic exponent alpha ( α ) for CDR-switching models and parameter
estimated for omega ( ω1 ) from SETAR-switching models for all the countries
starting Armenia are shown in alphabetical order.
3.5. Nature of Asymmetries
The results reveal that parameter estimates for characteristic exponent ( α ) is

close to normal behavior in Indonesia, Iran, Japan, Kazakhstan, and Russia.
However, the values for the characteristic exponent ( α ) for Armenia, Malaysia,
Philippines, Singapore, Thailand, and Turkey show fat tails. The values of the
switching parameter ( γ ) are different across different countries. For example,
high values of this parameter for Armenia, Indonesia, Iran, Kazakhstan, Malaysia,
Philippines, Singapore, Thailand, and Turkey shows that volatility in high regimes
in these countries is lower than in the low regimes. The results for Japan, and
Russia though in sharp contrast reveal that compared to developing countries
developed countries have different volatility patterns during different phases of
business cycles. Developing countries’ volatility has similar patterns irrespective
of their geographical locations. The values of ω1 for Armenia, Indonesia, Japan,
and Philippines are positive implying that negative shocks are less persistence in
these economies. On the other hands the value of ω1 for Iran, Kazakhstan,
Malaysia, Singapore, Thailand, and Turkey being negative reveals that positive
shocks are less persistent in these economies.
The results on nonlinearity tests based on time series models provide

statistically significant evidence of business cycle asymmetries in Indonesia, Iran,
Japan, Malaysia, Philippines, Singapore, and Turkey only. Neural network
nonlinearity tests based on in-sample as well jackknife out-of-sample forecasts
approximated from artificial neural networks, in conjunction with linear models
show that nonlinearities do prevail in all the series, including Iran, Kazakhstan,
and Russia. The results based on neural network linearity test show statistically
significant evidence of nonlinearities in Armenia, Indonesia, Iran, Japan,
Kazakhstan, Malaysia, Philippines, Russia, Singapore, Thailand, and Turkey real
GDP growth rates which is an improvement over the linearity tests that are
obtained from nonlinear time series and switching models. This shows that neural
networks linearity tests outperformed the linearity tests that are constructed from
forecasts from the nonlinear time series models employed for detecting possible
existence of nonlinearities in all the series. These results are in line with those
obtained from previous studies including Lee, et al. (1993) Terasvirta, et al.
(1993), and Kiani et al. (2005). Additionally, volatility patterns in developed
countries are in sharp contrast with those from the developing countries when
switching from lower regime of the business cycle to the high regime and vice
versa. The characteristic exponent α , that governs the tail behavior in these
countries, did not show any pattern like the other parameters did. Similarly, the
results also show that negative shocks are less persistent in Iran, Kazakhstan,
Malaysia, Singapore, Thailand, and Turkey, whereas positive shocks are less
persistent for Armenia, Indonesia, Japan, and Philippines
The results on nonlinearity in the conditional mean for Japan are in line with
Bidarkota (2000) and Kiani and Bidarkota (2004). This shows that the evidence
against linearity in conditional mean for the Japan is robust to changes in the
sample period. Koop and Potter (2001) investigated whether nonlinearities could
arise from structural instability. Blanchard and Simon (2001) show a possible
structural change in the early 1980s; however, this is not accounted for in the
present work. Similarly, Diebold and Inoue (2001) show that the series that
undergo occasional structural change may show spurious evidence of long
memory or spurious evidence of unit roots (Perron, 1989). These limitations also
apply to the present work.
4. CONCLUSION
Nonlinear time series models are employed to test possible existence of
business cycle asymmetries (if any) in Armenia, Indonesia, Iran, Japan,
Kazakhstan, Malaysia, Philippines, Russia, Singapore, Thailand and Turkey real
GDP growth rates. Time series models employed are fully parametric, and are
capable to account for time varying volatility, long memory in the process, and
outliers that might be present in the series. In addition, artificial neural networks
are also employed in order to find possible existence of nonlinearities in all the
series.
The findings from nonlinear augmented and switching time series models
reveal strong evidence of business cycle asymmetries in Indonesia, Japan,
Malaysia, Philippines, real GDP growth rates. The results show weak evidence of
business cycle asymmetries in Armenia, Singapore, Thailand, and Turkey.
However, the results do not show statistically significant evidence of business
cycle asymmetries in real GDP growth rates of Iran, Kazakhstan, and Russia.
The results from neural network linearity tests that are based on in-sample
approximations from neural networks show statistically significant evidence of
business cycle asymmetries in real GDP growth rates in Armenia, Indonesia, Iran,
Japan, Kazakhstan, Malaysia, Philippines, Russia, Singapore, Thailand, and
Turkey. Similarly, results from neural network linearity tests that are constructed
from jackknife out-of-sample approximations also show statistically significant
evidence of business cycle asymmetries in real GDP growth rates in Armenia,
Thailand, and Turkey. Thus, these results reveal that business cycle fluctuations in
all the Asian economies studied are asymmetric, thus alike.
The study results suggest that forecasts from linear models as well as those
derived from linear vector autoregression cannot be employed to forecast the
impact of monetary policy or any other shock on output in Armenia, Indonesia,
Iran, Japan, Kazakhstan, Malaysia, Philippines, Russia, Singapore, Thailand, and
Turkey. Therefore, the policymakers in these countries will have to employ
appropriate nonlinear time series forecasting models to anticipate the impact of
monetary policy or any other shock on output in these countries.
5. REFERENCES
Anderson, H., & Vahid, F. (1998). Testing multiple equation systems for common
nonlinear components. Journal of Econometrics, 84, 1-36.
Anderson, H., & Ramsey, J. (2002). US and Canadian industrial production
indices as coupled oscillator. Journal of Economic Dynamics and Control, 26,
33-67.
Andreano, M., & Savio, G. (2002). Further Evidence on Business Cycle
Asymmetries in G7 Countries. Applied Economics, 34, 895-904.
Andrews, D. (2001). Testing when parameter is under the boundary of the
maintained hypothesis. Econometrics, 69, 683-734.
dilemma: Algorithm and Simulated Annealing. Morgan Kaufmann, Los Altos:
CA.
Baillie, R. (1996). Long Memory Processes and Fractional Investigation in
Econometrics. Journal of Econometrics (Annals), 73, 5-59.
Baillie, R., Chung, C., & Tieslau, M. (1996). Analyzing Inflation by the
Fractionally Integrated ARFIMA-GARCH Model. Journal of Applied
Balke, N., & Fomby, T. (1994). Large shocks, small shocks and economic
fluctuations: outliers in macroeconomic time series. Journal of Applied
GNP: robust evidence. The Review of Economics and Statistics, 82,153-157.
Blanchard, O., & Simon, J. (2001). The Long and Large Decline in U.S. Output
Volatility. Brookings Papers on Economic Activity, 1, 135-174.
Blanchard, O., & Watson, M. (1986). Are Business Cycles All Alike? In Gordon,
R.J., (Eds.), The American Business Cycle: Continuity and Change,
University of Chicago Press, Chicago.
Box, G., & Jenkins, G. (1976). Time Series Analysis, Forecasting and Control.
Holden-Day, San Francisco, CA.
Brockwell, P., & Davis, R. (1991). Time Series: Theory and Methods. Springer-
Verlag: New York.

Davies, R. (1977). Hypothesis testing when a nuisance parameter is present under
the alternative. Econometrica, 69, 683-733.
System. Unpublished Ph.D. dissertation, University of Michigan, Department
Diebold, F., & Inoue, A. (2001). Long Memory and Regime Switching. Journal of
Learnings. Addison Wisely: Readings, MA.
Hess, G., & S. Iwata, S. (1997), Asymmetric Persistence in GDP? A Deeper Look
at Depth. Journal of Monetary Economics, 40, 535-554.
Hosking, J. (1984). Modeling Persistence in Hydrological Time Series Using
Fractional Differencing. Water Resources Research, 20, 1898-1908.
Kuan, C., & White, H. (1994). Artificial neural networks: an economic
perspective. Econometric Review, 13, 1-91.
Kiani, K. M., & Bidarkota, P.V. (2004). On business cycle asymmetries in G7
Kiani, K. (2005). Detecting business cycle asymmetries using artificial neural
network and time series models, Computational Economics, 26, 65-85.
Kiani, K., Bidarkota, P & Kastens, T. (2005). Forecast performance if neural
networks and business cycle asymmetries. Applied Financial Economics
Letters, 1, 205-210.
Kokoszka, P., & Taqqu, M. (1995). Fractional ARIMA with Stable Innovations.
Stochastic Processes and their Applications, 60, 19-47.
Koop, G., & Potter, S. (2001). Are Apparent Findings of Nonlinearity Due to
Structural Instability in Economic Time Series? Econometrics Journal, 4, 37-
55.
Lee, T., White, H., & Granger, C. (1993). Testing for neglected nonlinearity.
Journal of Econometrics, 56, 269-290.
Marimon, R., McGratten, E., & Sargent, T. (1990). Money as Medium of
Economics Dynamics and Control 14, 329-373.
McCulloch, J. (1997). Measuring tail thickness in order to estimate the stable

index α: a critique. Journal of Business and Economic Statistics, 15, 74-81.
Mikosch, T., Gadrich, T., Kluppelberg, C., & Adler, R. (1995). Parameter
Estimation for ARMA Models with Infinite Variance Innovations. Annals of
Statistics, 23, 305-326.
Neftci, S. (1984). Are Economic time series asymmetric over the business cycle?
Perron, P. (1989). The Great Crash, the Oil Price Shock and the Unit Root
Hypothesis. Econometrica, 57, 1361-1401.
Politis, D., & Romeo, J. (1994). Large sample confidence region based on sub-
sample under minimal assumption. Annals of Statistics, 22, 2031-2052.
Politis, D., Romeo, J., & Wolf, M. (1997). Sub-sampling for heteroskedastic time
series. Journal of econometrics, 81, 281-317.
Potter S. (1995). A non-linear approach to U.S. GNP, Journal of Applied
Qi, M. (2001). Predicting US recessions via leading indicators via neural network
models. International Journal of Forecasting, 17, 383-401.
real world systems. In Zornetzer, S. F., Davis, J. L. & Lau, C. (Eds.), An
Introduction to Neural and Electronic Networks, New Academic Press, York.
Scheinkman, J., and LeBaron, B. (1989). Non-linear dynamics and GNP data. In
Barnett, W. et al. (Eds.), Economic Complexity: Chaos, Sunspots, Bubbles,
and Non-linearity, Cambridge University Press, Cambridge.
Sowell, F. (1992a). Maximum Likelihood Estimation of Stationary Univariate
Fractionally Integrated Time Series Models. Journal of Econometrics,
53,165-188.
Sowell, F. (1992b). Modeling Long-Run Behavior With the Fractional ARIMA
Model. Journal of Monetary Economics, 29, 277-302.
Swanson, N., & White, H. (1995). A model selection approach to assessing the
information in the term structure using linear model and artificial neural
networks. Journal of Business Economics and Statistics, 13, 265-275.
Swanson, N., & White, H. (1997a). A model selection approach for real time
macroeconomic forecasting using linear models and artificial neural
networks. Review of Economics and Statistics, 79, 540-550.
Swanson, N., & White, H. (1997b). Forecasting economic time series using
flexible vs. fixed and linear vs. nonlinear economic models. International
journal of Forecasting 13, 439-461.
Terasvirta, T., Lin, C. & Granger, C. (1993). Power of neural network test.
Tsay, R. (1988). Nonlinear time series analysis of blowfly population. Journal of
Time Series Analysis, 9, 247-263.
Vishwakarma, K. (1995). Recognizing business cycle turning points by means of
a neural network. Computational Economics, 7, 175-185.
White, H. (1989). Some asymptotic results for learning in single hidden-layer feed
forward network models. American Statistical Association, 84, 1003-1013.
Wu, C. (1990). On the asymptotic properties of the jackknife histogram. Annals of
Statistics 18, 1438-1452.
Ziari, H., Leatham, D., & Ellinger, P. (1997). Developing of statistical
discriminant mathematical programming model via re-sampling estimation
techniques. American Journal of Economics, 79, 1352-1362.
INDEX
Asian countries, 86
assumptions, 15, 32, 44
A asymmetry, 2, 25, 44, 60, 107
asymptotic, 44, 60, 61, 95, 108
accounting, 3, 76
asymptotically, 93
accuracy, 5, 24
autoregressive model, 3, 23, 24, 25
ad hoc, 3, 88, 93
Azerbaijan, 86, 100
Afghanistan, 86, 100
agents, 24, 59, 106
aggregate supply, 2 B
agricultural, 24, 42
algorithm, 9, 31, 47, 51, 70, 93 Bahrain, 86, 100
alpha, 102 Bangladesh, 86, 100
alternative, 21, 30, 38, 39, 47, 50, 52, 58, 67, basic research, 28
75, 92, 95, 100, 106 Bayesian, 6, 88
alternative hypothesis, 21, 30, 38, 39, 50, 52, behavior, 23, 24, 43, 51, 58, 66, 86, 89, 102,
67, 75, 92, 95, 100 103, 106
amplitude, 2, 49 Bhutan, 86, 100
ANN, 1, 4, 7, 28, 29, 30, 31, 34, 36, 39, 40, bias, 32, 44, 60, 66
41, 47, 49, 50, 56, 63, 65, 66, 67, 68, 69, biology, 9, 29, 51, 71, 93
70, 75, 77, 78, 79, 91, 93 bivariate analysis, 79
Annealing, 23, 57, 80, 105 brain, 4, 29, 49, 66
application, 9, 32, 51, 71 Burma, 86, 100
Arabia, 86, 100 business, vii, viii, 1, 2, 3, 6, 17, 18, 20, 22, 23,
Armenia, 85, 86, 87, 88, 94, 95, 96, 97, 98, 24, 25, 27, 28, 29, 30, 41, 42, 43, 44, 45,
99, 100, 101, 102, 103, 104 46, 47, 48, 49, 55, 56, 57, 58, 59, 60, 63,
artificial, vii, viii, 1, 4, 24, 25, 28, 29, 34, 35, 64, 65, 66, 69, 76, 77, 78, 79, 81, 85, 86,
36, 37, 40, 41, 44, 45, 47, 49, 56, 59, 60, 87, 88, 94, 99, 100, 102, 103, 104, 106,
61, 63, 65, 66, 78, 79, 81, 87, 88, 91, 92, 107, 108
93, 100, 103, 104, 106, 107 business cycle, vii, viii, 1, 2, 3, 6, 17, 18, 20,
artificial intelligence, 4, 49, 66 22, 23, 24, 25, 27, 28, 29, 30, 41, 42, 43,
Asian, v, vii, viii, 85, 86, 94, 100, 104 44, 45, 46, 47, 48, 49, 55, 56, 57, 58, 59,
110 Index
60, 63, 64, 65, 66, 69, 76, 77, 78, 79, 81,
85, 86, 87, 88, 94, 99, 100, 102, 103, 104,
D
106, 107, 108
data analysis, 29, 65, 71
dating, 48
C decay, 89
decision-making process, 29, 49, 66
Cambodia, 86, 100 degrees of freedom, 16, 17, 37, 92
Canada, 1, 3, 6, 8, 9, 10, 11, 12, 13, 14, 15, demand, 2, 43
16, 17, 18, 19, 20, 21, 22, 27, 28, 29, 30, dependent variable, 7, 20, 67, 95
34, 35, 36, 37, 38, 40, 41, 42, 45, 47, 49, depression, 2, 48
50, 51, 52, 53, 54, 55, 56, 57, 63, 65, 66, derivatives, 24
67, 68, 69, 70, 71, 72, 73, 75, 76, 77, 78, 79 developed countries, 86, 100, 102, 103
capacity, 4 developing countries, 101, 102, 103
capitalist, 25, 60 deviation, 34, 35
causality, 42 dispersion, 93
CDR, 88, 94, 95, 96, 101, 102 disseminate, vii
Chicago, 23, 43, 58, 59, 80, 81, 105 distribution, 4, 5, 32, 37, 39, 47, 49, 51, 53,
China, 86, 100 55, 65, 67, 89, 90, 91, 95
classes, 88, 94 distribution function, 5, 91
classical, 25, 48 duration, 2, 23, 43, 48, 58
classified, 2, 48
codes, 27, 46, 64
commerce, 43, 59
E
Commerce Department, 81
econometrics, 60, 82, 107
components, 2, 6, 7, 13, 18, 19, 23, 57, 105
economic, 2, 9, 23, 24, 43, 49, 51, 58, 59, 60,
computers, 44, 60, 83
64, 71, 79, 86, 88, 105, 106, 107
computing, 24
economic activity, 24, 49
conditional mean, 22, 23, 42, 58, 90, 94, 95,
economics, vii, 28, 47, 65, 93
99, 103, 105
economies, viii, 77, 79, 85, 86, 94, 100, 102,
confidence, 44, 59, 60, 107
104
consensus, 87
economy, viii, 24, 48, 59, 77, 79, 86, 87, 89,
construction, 47, 91
106
consumption, 42
employment, 48
continuity, 43, 59
endogenous, 50, 70
convergence, 9, 93
engineering, 9, 28, 29, 51, 71, 93
convex, 31
estimating, 9, 31, 32, 33, 36, 66, 70, 93
critical value, 12, 39, 53, 55
estimation problems, 93
criticism, 47
estimation process, 93
CSS, 93
estimator, 93
cumulative distribution function, 5, 91
estimators, 32, 66, 93
cycles, viii, 2, 3, 23, 24, 25, 27, 28, 43, 44, 45,
European, 46
46, 48, 49, 58, 59, 60, 63, 64, 65, 77, 79,
European Union, 46
86, 88, 94, 102
evidence, vii, 1, 3, 6, 22, 23, 24, 27, 28, 39,
42, 43, 45, 46, 55, 56, 57, 58, 63, 64, 65,
Index 111
69, 76, 77, 79, 85, 86, 87, 99, 100, 103, genetic, 9, 31, 47, 51, 70, 93
104, 105 Georgia, 86, 100
evolution, 23, 57, 105 Germany, 3, 9, 28, 35, 45, 46, 47, 49, 50, 51,
exchange rate, 24, 42 52, 53, 54, 55, 56, 57, 63, 65, 66, 67, 68,
exchange rates, 24, 42 69, 70, 71, 72, 73, 75, 76, 77, 78, 79
exercise, 86, 94, 100 GNP, 3, 23, 25, 42, 43, 44, 46, 58, 60, 80, 82,
exogenous, 50, 66, 70 87, 105, 106, 107
expansions, 48, 89 graduate students, viii
Great Depression, 48, 58
gross domestic product, 63
F gross national product, 46, 64
growth, 1, 3, 9, 10, 11, 12, 15, 17, 18, 19, 20,
false, 12, 38, 52, 75
21, 22, 27, 28, 29, 30, 31, 34, 35, 36, 37,
family, 90
38, 41, 42, 45, 46, 47, 50, 52, 53, 55, 56,
fat, 10, 71, 86, 102
57, 63, 65, 66, 67, 69, 70, 71, 73, 75, 76,
Federal Reserve, 48
77, 78, 85, 86, 87, 88, 89, 94, 95, 100, 103,
finance, 29, 47, 65
104
financial crises, 87
growth rate, 1, 3, 9, 10, 11, 12, 15, 17, 18, 19,
fluctuations, vii, 2, 3, 23, 24, 28, 45, 46, 48,
20, 21, 22, 27, 28, 29, 30, 31, 34, 35, 36,
49, 57, 59, 60, 63, 64, 65, 76, 77, 79, 85,
37, 38, 41, 42, 45, 46, 47, 50, 52, 53, 55,
86, 87, 93, 94, 100, 104, 105
56, 57, 63, 65, 66, 67, 69, 70, 71, 73, 75,
food, 43
76, 77, 78, 85, 86, 87, 88, 89, 94, 95, 100,
forecasting, viii, 5, 14, 25, 27, 41, 43, 48, 49,
103, 104
60, 64, 87, 104, 107
foreign exchange, 24, 58
France, 1, 3, 6, 8, 9, 10, 11, 12, 13, 14, 15, 16, H
17, 18, 19, 20, 21, 22, 27, 28, 29, 30, 34,
35, 36, 37, 38, 40, 41, 42, 45, 46, 47, 49, hands, 21, 102
50, 51, 52, 53, 54, 55, 56, 57, 63, 65, 66, hedging, 58, 59
67, 68, 69, 70, 71, 72, 73, 75, 76, 77, 78, 79 heterogeneous, viii
freedom, 16, 17, 37, 92 heteroskedasticity, 3, 11, 71, 88, 90
histogram, 44, 61, 108
homogeneity, 58
G Hong Kong, 86, 100
human, 4, 29, 49, 66
G7 countries, vii, 9, 24, 28, 35, 42, 46, 57, 59,
human brain, 4, 49, 66
65, 66, 76, 79, 86, 87, 106
hyperbolic, 5, 91
Gaussian, 89
hypothesis, 2, 9, 12, 14, 18, 20, 21, 22, 30, 31,
GDP, 1, 2, 3, 9, 10, 11, 15, 17, 18, 19, 20, 21,
36, 37, 38, 39, 50, 52, 54, 57, 67, 72, 75,
22, 23, 27, 28, 29, 30, 31, 34, 35, 36, 37,
77, 85, 92, 95, 100, 105
38, 41, 42, 45, 46, 47, 49, 50, 51, 52, 53,
55, 56, 57, 58, 63, 64, 65, 66, 67, 68, 69,
70, 71, 73, 75, 76, 77, 78, 80, 85, 86, 87, I
88, 89, 91, 94, 95, 96, 100, 103, 104, 105,
106 identification, 48
generalization, 5 Illinois, 81
112 Index
imports, 77
inclusion, 77
K
income, 48
Kazakhstan, 85, 86, 87, 88, 94, 95, 96, 97, 98,
independent variable, 7
99, 100, 101, 102, 103, 104
India, 86, 100
Keynesian, 2, 48
indicators, 42, 44, 57, 60, 107
Keynesians, 48
indices, 57, 105
Korea, 86, 100
Indonesia, 85, 86, 87, 88, 94, 95, 96, 97, 98,
Kuwait, 86, 100
99, 100, 101, 102, 103, 104
Kyrgyzstan, 86, 100
industrial, 57, 105
industrial production, 57, 105
industrialized countries, 3, 63, 79 L
inferences, 21, 100
infinite, 89 Laos, 86, 100
information processing, 4, 29, 49, 66 laws, 30, 47
instability, 103 learning, 4, 29, 49, 59, 60, 66, 108
intelligence, 4, 49, 66 Lebanon, 86, 100
international, 46, 65 likelihood, 9, 51, 70, 89, 93, 96, 97, 99
Iran, 86, 87, 88, 94, 95, 96, 97, 98, 99, 100, limitations, 86, 94, 100, 103
101, 102, 103, 104 linear, vii, viii, 2, 6, 7, 8, 9, 13, 14, 16, 18, 21,
Iraq, 86, 100 25, 27, 28, 29, 30, 31, 34, 35, 36, 37, 38,
Israel, 86, 100 39, 40, 41, 42, 44, 45, 47, 48, 49, 50, 51,
Italy, 3, 9, 28, 35, 45, 47, 49, 50, 51, 52, 53, 52, 53, 54, 55, 56, 57, 60, 63, 64, 65, 66,
54, 55, 56, 57, 63, 65, 66, 67, 68, 69, 70, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77,
71, 72, 73, 75, 76, 77, 78, 79 79, 85, 86, 87, 90, 92, 93, 95, 99, 100, 103,
iteration, 94 104, 107
linear model, vii, viii, 2, 6, 7, 8, 13, 16, 18, 21,
27, 28, 29, 30, 31, 34, 35, 36, 37, 38, 39,
J
40, 41, 42, 45, 47, 49, 50, 51, 52, 53, 55,
56, 57, 60, 63, 64, 65, 66, 67, 68, 69, 70,
jackknife, 27, 28, 29, 32, 35, 36, 37, 38, 39,
71, 73, 74, 75, 76, 77, 79, 85, 86, 87, 92,
41, 42, 44, 45, 47, 49, 51, 52, 55, 56, 57,
93, 95, 99, 100, 103, 104, 107
61, 63, 65, 66, 69, 70, 71, 72, 73, 74, 75,
linear regression, 14
76, 77, 78, 79, 92, 99, 100, 103, 104, 108
literature, 5, 48
Japan, 1, 3, 6, 8, 9, 10, 11, 12, 13, 14, 15, 16,
livestock, 42
17, 18, 19, 20, 21, 22, 27, 28, 29, 30, 34,
location, 89
35, 36, 37, 38, 40, 41, 42, 45, 47, 49, 50,
51, 52, 53, 54, 55, 56, 57, 63, 65, 66, 67,
68, 69, 70, 71, 72, 73, 75, 76, 77, 78, 79, M
85, 86, 87, 88, 94, 95, 96, 97, 98, 99, 100,
101, 102, 103, 104 M1, 35, 36
Jordan, 86, 100 macroeconomic, vii, viii, 2, 3, 23, 28, 46, 47,
judge, 39 48, 49, 60, 64, 86, 87, 105, 107
justification, 47 macroeconomics, 2, 48, 60
Index 113
Malaysia, 85, 86, 87, 88, 94, 95, 96, 97, 98, neural network, vii, viii, 1, 2, 3, 4, 5, 6, 7, 8, 9,
99, 100, 101, 102, 103, 104 13, 14, 20, 21, 22, 24, 25, 27, 28, 29, 30,
management, 2, 48 31, 34, 35, 36, 37, 38, 39, 40, 41, 43, 44,
mapping, 24 45, 47, 49, 50, 51, 52, 53, 54, 55, 56, 57,
market, 60 58, 59, 60, 61, 63, 64, 65, 66, 67, 68, 69,
Markov, 46, 47, 65 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 81,
mathematical, 32, 44, 51, 61, 93, 108 85, 86, 87, 88, 91, 92, 93, 95, 99, 100, 103,
mathematical programming, 44, 61, 108 104, 106, 107, 108
matrix, 6, 7, 13, 30, 91 Neural Network Model, 56, 82
Maximum Likelihood, 107 neural networks, vii, viii, 1, 5, 6, 8, 20, 22, 24,
measurement, 42, 57 25, 27, 28, 29, 34, 36, 37, 39, 40, 41, 43,
measures, 2, 10, 47, 64, 86, 87, 89 44, 45, 47, 49, 50, 51, 52, 53, 55, 56, 57,
mechanical, iv 58, 59, 60, 61, 63, 64, 65, 66, 69, 71, 73,
medium of exchange, 24 74, 76, 77, 78, 79, 86, 87, 88, 92, 93, 95,
memory, 3, 24, 46, 65, 85, 86, 87, 88, 89, 90, 100, 103, 104, 106, 107
103, 104 neurons, 91
Mexico, 79 New York, 25, 44, 58, 59, 60, 82, 105
MLE, 93 nodes, 4, 5, 30, 34, 35, 36, 91
modeling, vii, 2, 4, 28, 30, 47, 65 noise, 33
models, vii, viii, 1, 2, 4, 6, 9, 13, 16, 18, 20, nonlinear, vii, viii, 1, 3, 4, 5, 9, 14, 16, 17, 18,
21, 22, 23, 24, 25, 27, 28, 29, 34, 35, 36, 21, 22, 24, 25, 28, 29, 30, 31, 38, 44, 45,
37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 46, 47, 49, 51, 57, 58, 60, 64, 65, 66, 67,
49, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 70, 75, 85, 87, 88, 89, 90, 91, 93, 94, 95,
63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 96, 98, 103, 104, 105, 107
74, 75, 76, 77, 78, 79, 81, 85, 86, 87, 88, nonlinearities, 1, 2, 3, 6, 8, 9, 17, 21, 22, 24,
89, 90, 91, 93, 94, 95, 96, 97, 98, 99, 100, 25, 38, 41, 42, 45, 49, 52, 55, 56, 59, 63,
101, 102, 103, 104, 106, 107, 108 64, 66, 75, 76, 77, 79, 85, 86, 90, 92, 95,
monetarists, 48 100, 103, 104
monetary policy, viii, 1, 2, 42, 45, 47, 49, 57, nonparametric, 9, 23, 30, 43, 49, 58, 59, 66,
63, 64, 79, 85, 86, 87, 104 91, 106
money, 2 normal, 10, 89, 90, 93, 102
Mongolia, 86, 100 normal distribution, 89
Myanmar, 86, 100 North America, 79
North American Free Trade Agreement
(NAFTA), 79
N null hypothesis, 18, 20, 21, 22, 30, 38, 39, 52,
75, 95, 100
national product, 46, 64
Nepal, 86, 100
network, vii, 1, 2, 3, 4, 5, 6, 7, 8, 9, 13, 14, 20, O
21, 22, 24, 25, 27, 28, 29, 30, 31, 34, 35,
36, 37, 38, 39, 40, 41, 42, 44, 45, 47, 49, observations, 8, 14, 16, 20, 31, 32, 33, 38, 51,
50, 51, 52, 53, 54, 55, 56, 59, 60, 65, 66, 55, 66, 67, 70, 72, 75, 92
67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, oil, 86
79, 81, 85, 91, 92, 93, 95, 99, 100, 103, Oman, 86, 100
104, 106, 107, 108 operator, 89
114 Index
optimization, 9, 51, 70, 93 recession, 2, 48, 89

organ, 57 recessions, 23, 42, 44, 48, 57, 60, 89, 105, 107
oscillator, 105 recovery, 2
outliers, 3, 23, 87, 104, 105 regression, 7, 8, 14, 16, 19, 25, 33, 38, 40, 75
regression analysis, 25
regression equation, 7, 8, 14
P regressions, viii, 33
rejection, 95
Pakistan, 86, 100
relationship, 4, 29, 45
paper, 29, 48
relationships, 24, 29
parameter, 9, 16, 32, 40, 50, 51, 57, 58, 67,
research, vii, 2, 6, 9, 28, 46, 48, 51, 64, 65, 66,
70, 89, 91, 92, 94, 101, 102, 105, 106
71, 79, 93
parameter estimates, 40, 89, 91, 101, 102
researchers, viii, 4, 5, 21, 30, 87
perception, 3, 87
residuals, 7, 8, 14, 17, 18, 19, 21, 31, 33, 35,
performance, vii, 27, 28, 29, 39, 40, 41, 47,
36, 50, 69, 70, 92
49, 55, 56, 59, 66, 78, 79, 106
reunification, 9, 35
permit, 89
Russia, 85, 86, 87, 88, 94, 95, 96, 97, 98, 99,
Philippines, 85, 86, 87, 88, 94, 95, 96, 97, 98,
100, 101, 102, 103, 104
99, 100, 101, 102, 103, 104
policymakers, viii, 1, 42, 86, 87, 104
polynomial, 18, 89 S
polynomials, 18, 19, 94
population, 25, 108 sales, 42
power, 5, 14 sample, vii, 6, 22, 27, 28, 29, 32, 33, 34, 35,
prediction, 58, 91 36, 37, 38, 39, 40, 41, 42, 44, 45, 47, 49,
preparation, iv 51, 52, 53, 55, 56, 57, 59, 63, 65, 66, 68,
prices, 42 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
principal component analysis, 7, 18 86, 87, 92, 95, 99, 100, 103, 104, 107
probability, 32, 39, 94 sample mean, 33
production, 57, 105 sampling, 29, 32, 36, 51, 60, 61, 66, 107, 108
programming, 44, 61, 108 Saudi Arabia, 86, 100
propagation, 25, 48, 90 science, 28
search, 94
securities, 58, 59
Q selecting, 7
series, vii, viii, 1, 2, 3, 4, 6, 9, 11, 16, 17, 18,
Qatar, 86, 100
19, 20, 21, 22, 23, 24, 25, 27, 28, 29, 33,
34, 35, 36, 38, 39, 40, 41, 42, 43, 44, 45,
R 46, 47, 49, 50, 52, 55, 56, 58, 59, 60, 63,
64, 65, 66, 67, 68, 69, 70, 71, 75, 76, 77,
random, 9, 51, 70, 89, 90, 94 78, 79, 81, 85, 86, 87, 88, 90, 92, 93, 94,
random walk, 90 95, 96, 98, 99, 100, 103, 104, 105, 106,
range, 9, 10, 34 107, 108
real time, 60, 107 shock, viii, 1, 2, 42, 45, 48, 57, 63, 64, 79, 85,
reality, 25 87, 104
Index 115
shocks, viii, 2, 23, 47, 49, 64, 86, 89, 90, 93, Thailand, 85, 86, 87, 88, 94, 96, 97, 98, 99,
102, 103, 105 100, 101, 102, 103, 104
sigmoid, 5, 91 theoretical, 25, 47, 60
sign, 33, 39 theory, 42, 43, 57
signals, 5 thinking, 48
significance level, 12, 21, 39, 55, 76, 100 threshold, 90, 91
similarity, vii time series, vii, viii, 1, 2, 3, 4, 6, 17, 20, 21,
Singapore, 85, 86, 87, 88, 94, 95, 96, 97, 98, 22, 23, 24, 25, 28, 43, 44, 46, 47, 49, 58,
99, 100, 101, 102, 103, 104 59, 60, 64, 65, 69, 81, 85, 86, 87, 88, 94,
skewness, 11, 12, 71 96, 98, 103, 104, 105, 106, 107, 108
speed, 93 trade, 48, 79
spillover effects, 65 trading, 58
spillovers, 76 training, 5
Sri Lanka, 86, 100 transfer, 5, 13, 91, 92
stabilization, viii transition, vii, 3, 23, 24, 25, 86, 100
stabilize, viii transition economies, 86, 100
standard deviation, 10, 34, 35 transmission, 90
standard error, 101, 102 trend, 10, 11, 12, 71
statistical analysis, 25, 60 trial, 5, 29
statistical inference, 100 trial and error, 5
statistics, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, Turkey, 85, 86, 87, 88, 94, 96, 97, 98, 99, 100,
18, 19, 20, 21, 33, 36, 37, 38, 39, 40, 52, 101, 102, 103, 104
53, 54, 55, 67, 71, 72, 73, 74, 75, 76, 98, 99 Turkmenistan, 86, 100
STD, 34, 35
Stochastic, 106
stock, 60 U
strategies, 23, 57, 105
Uunemployment, 46, 64
students, viii
unemployment rate, 46, 64
superiority, 29
United Arab Emirates, 86, 100
supply, 2
United Kingdom (UK), K, 1, 3, 6, 8, 9, 10, 11,
switching, vii, viii, 46, 47, 65, 85, 90, 94, 101,
12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22,
102, 103, 104
27, 28, 29, 30, 34, 35, 36, 37, 38, 40, 41,
Syria, 86, 100
42, 45, 46, 47, 49, 50, 52, 53, 54, 55, 56,
systems, 25, 44, 57, 60, 105, 107
57, 63, 65, 66, 67, 68, 69, 70, 71, 72, 73,
75, 76, 77, 78, 79
T United States, 48, 52
univariate, 28, 55, 56, 69, 71, 72, 75, 76, 77,
Taiwan, 86, 100 79
Tajikistan, 86, 100 Uzbekistan, 86, 100
technology, 4, 29, 49, 66
test statistic, 6, 7, 8, 9, 12, 13, 14, 15, 16, 17,
18, 19, 20, 21, 30, 31, 33, 37, 38, 39, 50,
52, 53, 54, 55, 67, 68, 70, 71, 72, 73, 75,
76, 90, 92, 93, 98, 99
116 Index
volatility, 3, 85, 87, 102, 103, 104

V
values, 8, 9, 12, 13, 14, 15, 16, 17, 18, 19, 20, W
38, 40, 53, 55, 71, 72, 73, 75, 76, 89, 98,
99, 102 warrants, vii
VAR models, 70
variable, 7, 17, 18, 19, 89, 95
variables, 2, 7, 20, 24, 30, 33, 48, 50, 66, 70, Y
86, 91
Yemen, 86, 100
variance, 24, 89
vector, viii, 2, 9, 30, 32, 45, 47, 51, 57, 63, 64,
69, 70, 79, 91, 94, 104
Vietnam, 86, 100

Business Cycle Fluctuations

Uploaded by

Document Information

Original Description:

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

Business Cycle Fluctuations

Uploaded by

Copyright:

Available Formats

BUSINESS CYCLE FLUCTUATIONS

AND ECONOMIC POLICY

Nova Science Publishers, Inc.

NOTICE TO THE READER

LIBRARY OF CONGRESS CATALOGING-IN-PUBLICATION DATA

ASYMMETRIC BUSINESS CYCLE

output based on forecasts obtained from linear models including those

Keywords: asymmetries; business cycles; neural network; nonlinearities;

2. ARTIFICIAL NEURAL NETWORKS

yt = f ( yt −1, yt − 2 , yt −3 , ......, yt − k ) (1.1)

The processed signals ψ ( ~x 'γ i ), i = 1,......., k are fed to the transfer

u t ~ N (0, σ 2 ), ~ ' )' ,

and π = ( π 0 , π 1 ........, π p )' .

2.1. Neural Network Test 1 (nnw1)

and π = (π 0 , π 1 ,........,π p )'

where, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π 0 is intercept. Equation 1.5 is a neural

Finally, residuals ( û t ) are regressed on the matrix of principal components (ψ t* ),

u t ~ Nid (0, σ 2 ), ~ ' )' ,

and π = (π 0 , π 1 ,........,π p )'

where, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π 0 is intercept. Equation 1.7 shows a

TS = {( SSE1 − SSE 2) / m} /{SSE 2 /(n − p − m − 1)} (1.9)

where, m denotes the number of restrictions in the unrestricted model, n is the

The test statistics (TS) is distributed F [m, (n − p − m − 1)] under normality

2.3. Estimation Issues

2.4. Data Sources

Figure 1. Plot of Real GDP Growth Rates over Time.

2.5. Preliminary Data Analysis

Table 1.1. Data Description

Canada France Japan UK USA

Table 1.2. Summary Statistics

Canada France Japan UK USA

Table 1.2. (Continued)

Canada France Japan UK USA

Table 1.3. Summaries of Statistical Inferences

Canada France Japan UK USA

Table 2. Neural Network Tests Results

Tests Test Statistics p-values

Tests Test Statistics p-values

and π = (π 0 , π 1 ,........,π p )'

4. ‘‘TS” denotes calculated values of ‘Test Statistics” that is distributed χ ( p ) under

the assumption of normality.

and π = (π 0 , π 1 ,........,π p )'

where, ψ (γ ' wt ) = (1 + exp{−γ ' wt )) −1 and π0 is intercept.

3. TIME SERIES LINEARITY TESTS

3.1. Keenan Test

Finally, the test statistics ( TS ) is calculated using SSE1 and η̂ as per

TS = (ηˆ 2 (2 − 2M − 2)) /( SSE1 − ηˆ 2 ) (1.11)

where, TS is distributed F [1, (n-2M-2)] under the assumptions of normality.

Test results based on Keenan model predictions are presented in Table 3 . In

Table 3. Other Nonlinearity Test Results

Test Test Statistics p-values

where, ηˆ = ηˆ 0 ( SSE )1 / 2 η̂ = η 0^ (SSE2)1/2. η̂ 0 is the regression coefficient, which is

8. The test statistic for RESET2 is approximately distributed χ ( p * ) when Ho is true

3.2. Tsay Test

TS = T ( SSE0 − SSE1 ) / SSE0 (1.12)

3.3. Ramsey Test RESET

TS = ((SSE0 − SSE1 ) /(k − 1)) /( SSE0 /(n − k )) (1.14)

where, in Equation 1.14 , k is the number of polynomials of forecasts, and n is

This linearity test incorporates principal component analysis of the

Finally, the test statistics are computed using Equation 1.16 .

TS = ((SSE2 − SSE1 ) /( p* )) /( SSE2 /(n − k (1.16)

where, in Equation 1.16 , k is the number of polynomials of forecast, n is the

RESET2 is a different version of the RESET linearity test that also