Professional Documents
Culture Documents
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
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
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.
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
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
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.
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 .
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
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
Asymmetric Business Cycle Fluctuations via Time Series Models … 7
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,
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).
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
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
Asymmetric Business Cycle Fluctuations via Time Series Models … 11
fails to reject homoskedasticity in all countries, and the Lagrange Multiplier (LM)
test detects autoregressive conditional heteroskedasticity (ARCH) only in Japan.
Table 2. (Continued)
~ ' )' ,
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).
yt = π ' wt + ut (2d)
where,
u t ~ Nid (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
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 .
Asymmetric Business Cycle Fluctuations via Time Series Models … 15
n
ηˆ = ηˆ0 ( ∑v )
i = M +1
2 1\ 2
t (1.10)
Table 3. (Continued)
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)
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.
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.
18 Khurshid M. Kiani
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 .
3.4. RESET1
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)
3.5. RESET2
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
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.
Asymmetric Business Cycle Fluctuations via Time Series Models … 21
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.
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
22 Khurshid M. Kiani
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.
Asymmetric Business Cycle Fluctuations via Time Series Models … 23
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.
24 Khurshid M. Kiani
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
Econometrics, 10, 109-125.
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
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.
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).
30 Khurshid M. Kiani
⎡ 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.
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)
Neural Networks Forecasts Evaluations … 31
where,
u t ~ N (0, σ 2 ), ~ ' )' ,
wt = (1, w ~ = ( y ,..., y )'
w
t t t −1 t− p
where
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,
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.
where,
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.
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.
3. EMPIRICAL RESULTS
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).
36 Khurshid M. Kiani
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.
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
Neural Networks Forecasts Evaluations … 37
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.
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,
Neural Networks Forecasts Evaluations … 39
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.
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
40 Khurshid M. Kiani
Table 5.1. AGS TEST: In-sample Performance of Linear vs. ANN Model
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.
42 Khurshid M. Kiani
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?
Journal of Monetary Economics, 31, 149-163.
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
Econometrics, 3, 191-200.
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.
Neural Networks Forecasts Evaluations … 43
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
Econometrics, 10, 109-125.
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
Political Economy, 97, 1255-1260.
Terasvirta, T., Lin, C.F. & Granger, C.W.J. (1993). Power of neural network test,
Journal of Time Series Analysis, 14, 209-220.
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
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.
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
2. EMPIRICAL MODEL
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
A Robust Evidence of Business Cycle Asymmetries … 49
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.
y t = π ' wt + u t
(3.1)
where
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)
1
This test statistic is approximate because of the nuisance parameter that appears under the
alternative hypothesis (Davies, 1977; Andrews, 2001).
A Robust Evidence of Business Cycle Asymmetries … 51
3. EMPIRICAL RESULTS
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).
52 Khurshid M. Kiani
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
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.
54 Khurshid M. Kiani
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.
A Robust Evidence of Business Cycle Asymmetries … 57
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.
Axelrod, R. (1987). The evolution of the strategies in the iterated prisoner’s
dilemma. In Davis, L. D. (Eds.), Algorithm and Simulated Annealing, Morgan
Kaufmann, Los Altos, CA: pp 32-41.
Beaudry, P., & Koop, G. (1993). Do recessions permanently change output?
Journal of Monetary Economics, 31, 149-163.
58 Khurshid M. Kiani
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
countries. Oxford Bulletin of Economics and Statistics, 66, 333-351.
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
exchange in an economy with artificially intelligent agents. Journal 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?
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.
60 Khurshid M. Kiani
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
Econometrics, 10, 109-125.
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.
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, 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
Political Economy, 97, 1255-1260.
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.
Journal of Time Series Analysis, 14, 209-220.
Tuckey, J. (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.
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.
A Robust Evidence of Business Cycle Asymmetries … 61
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.
64 Khurshid M. Kiani
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
66 Khurshid M. Kiani
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
Asymmetric Business Cycle Fluctuations and Contagion Effects… 67
where,
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.
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.
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 .
3. EMPIRICAL RESULTS
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.
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.
72 Khurshid M. Kiani
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. (Continued)
Test statistics shown in columns 2 and 5 are constructed respectively using in-
sample and jackknife out-of-sample forecasts that are distributed
Asymmetric Business Cycle Fluctuations and Contagion Effects… 75
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
76 Khurshid M. Kiani
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
Asymmetric Business Cycle Fluctuations and Contagion Effects… 77
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.
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.
78 Khurshid M. Kiani
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-
Asymmetric Business Cycle Fluctuations and Contagion Effects… 79
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.
80 Khurshid M. Kiani
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?
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.
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.
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.
Asymmetric Business Cycle Fluctuations and Contagion Effects… 81
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.
Estrella, A., & Mishkin, F. (1998). Predicting U.S. Business Cycle Regimes.
Journal of Business and Economic Statistics, 17, 313-323.
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.
Goldberg, D. (1989). Genetic Algorithm in Search, Optimization and Machine
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.
82 Khurshid M. Kiani
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.
86 Khurshid M. Kiani
Key phrases: Real GDP growth rates; fat tails; stable distributions; neural
networks; out-of-sample forecasts; long memory; nonlinearities; business
cycles
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).
88 Khurshid M. Kiani
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.
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.
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)
90 Khurshid M. Kiani
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.
ε t | I t −1 ~ zt ct z ~ iid Sα (0,1)
In Regime 2:
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.
⎡ n ⎛ k ⎞⎤
f ( x) = sig ⎢α 0 + ∑ α j sig ⎜⎜∑β ij xi + β 0 j ⎟⎥ + ε
⎟⎥
(5.3)
⎣⎢ j =1 ⎝ i =1 ⎠⎦
yt = π ' wt + u t (5.4)
⎡ 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 5.6 , 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.
Business Cycle Asymmetries in Asian Economies… 93
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 .
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. EMPIRICAL RESULTS
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.
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.
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.
Business Cycle Asymmetries in Asian Economies… 95
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.
Table 1. (Continued)
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.
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.
Business Cycle Asymmetries in Asian Economies… 101
9
Armenia, Indonesia, Iran, Malaysia, Philippines, Japan, Kazakhstan, Russia, Singapore, Thailand,
and Turkey,
102 Khurshid M. Kiani
Table 6. (Continued)
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,
Indonesia, Iran, Japan, Kazakhstan, Malaysia, Philippines, Russia, Singapore,
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.
Business Cycle Asymmetries in Asian Economies… 105
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.
Axelrod, R. (1987). The evolution of the strategies in the iterated prisoner’s
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
Econometrics, 11, 23-40.
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., & 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.
106 Khurshid M. Kiani
Terasvirta, T., Lin, C. & Granger, C. (1993). Power of neural network test.
Journal of Time Series Analysis, 14, 209-220.
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
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
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