Professional Documents
Culture Documents
17 (2005) 480–495
www.elsevier.com/locate/econbase
Abstract
This paper studies the sectoral mechanism of shock propagation in business cycles by performing a
case study. After we transform business survey indices of different industries into comparable metrics,
we trace and compare them before and after reference dates. From these observations, we show that
manufacturing industry leads the peak and the trough of the most recent business cycle in Korea; and six
sectors are major players in manufacturing industry, i.e. ‘food products and beverages,’ ‘chemicals,’
‘machinery,’ ‘motor vehicles, etc.,’ and ‘basic metals,’ and ‘fabricated metals.’ In conclusion, the heavy
and chemical sectors may be the key to understanding recent business cycles in Korea.
# 2004 Published by Elsevier B.V.
JEL Classification: E3
Keywords: Business survey index; Generalized birth and death process; Leading industries
1. Introduction
As Zarnowitz (1992, p. 316) put it, there is no single proven and accepted cause of all
observed business cycles. The corollary of this statement would be that each business cycle
Corresponding author. Tel.: +82 33 760 2356.
E-mail addresses: yangjm@dragon.yonsei.ac.kr (J.-M. Yang), taewan@deu.ac.kr (T.-W. Kim).
1
Tel.: +82 51 890 1413.
has its own cause and characteristics. This paper is an attempt to find out the specific
characteristics of the recent business cycles in Korea. The most recent business cycle in
Korea experienced a major impact, that was, the foreign currency crisis in 1997. We believe
that the analysis of the time before and after the crisis will give us a pattern of propagation
of the impact. According to Korea National Statistical Office (2004), the peak of that recent
business cycle was the first quarter of year 1996. After the peak, Korean economy observed
a slight up-turn in the first quarter of year 1997, which lasted until right before the crisis.
However, during the period from the first quarter of year 1997 to the crisis, the banking
sector showed symptoms of collapse. The trough of the recent business cycle was the third
quarter of year 1998. In contrast to the movements near the peak, Korean economy showed
a relatively fast recovery from the trough compared to the other Asian economies.
In investigating these phenomena, we believe that it is important to study the sequence
of shock propagation in each particular phase of the business cycle; which industries lead
and which industries lag. Generally speaking, policy makers have to find out which
industries are causing the problem rather quickly in order to meet the demand of the general
public and fight against the down-turn of business condition.
Most countries, either officially or unofficially, survey business conditions and
announce business survey indices. These indices may be a good indicator for business
conditions. They are available in a relatively short period of time compared to other
economic data, and carry qualitative information without scale effects. Therefore business
survey indices are strong candidates for analyzing the sequence of the shock propagation.
The indices, however, have their own problems. As Piatier (1979) pointed out, the main
problem in comparing different business survey indices is that they are not comparable.
Moreover, business survey indices have bandwagon effects and systematic bias among
respondents. Bandwagon effects tend to exaggerate the true state of business condition and
systematic bias is also one of the major problems that undermine their usefulness. In order
to search for leading industries by using business survey indices, it is essential to transform
the indices into comparable ones after we filter out such noise from the indices. After we
calculate the comparable metrics from the business survey indices, we can search for
leading sectors in business cycles.
In the section immediately following, we discuss a method that filters out interaction
effects and adjusts the differences in information precision among the business survey
indices. In Section 3, we trace the sectoral paths before and after the turning points, and
search for leading sectors. We also discuss the major players in the leading industry.
Section 4 rounds off with some comments and conclusions.
In this section, we provide a model for filtering out both bandwagon effects and
systematic bias from survey indices. The model assumes that fundamental probability
1
This section is based on Yang (2004).
482 J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495
determines the way that respondents answer interviewers, and that certain stochastic
dynamics driven by the fundamental probability determine the survey results. Our task is,
therefore, to extract the fundamental probability.
Suppose that there are N firms, and that k firms answer business conditions are good. A
business survey index (BSI, I) is calculated by using Eq. (1).
k ðN kÞ
I¼ 100 þ 100 (1)
N
Let us assume that firms sequentially change their response so that the state may change
in an iterative process step by step. A step is defined as an imaginary short period of time,
which allows only one sequential move. Hence it is assumed that counterfactual dynamics
can reach an equilibrium instantly after infinite steps. In an equilibrium, there is no change
in the probability distribution over the states.
If n firms answer good at step t þ 1, then this state can be realized from the following
situations at step t.
k ¼ n 1; n; n þ 1;
where k is the number of the positive responses at step t.
At step t þ 1, the probability of k ¼ n, Ptþ1 ðnÞ, can be defined as Eq. (2).
Ptþ1 ðnÞ ¼ wn1;n Pt ðn 1Þ þ wn;n Pt ðnÞ þ wnþ1;n Pt ðn þ 1Þ; (2)
where wn1;n is the transition probability that the state of k ¼ n 1 is changed into the state
of k ¼ n.
Note that the cases of k ¼ 0 and k ¼ N are different from Eq. (2). In case of k ¼ 0,
Ptþ1 ð0Þ ¼ w0;0 Pt ð0Þ þ w1;0 Pt ð1Þ. In case of k ¼ N, Ptþ1 ðNÞ ¼ wN1;N Pt ðN 1Þþ
wN;N Pt ðNÞ.
The equilibrium probability of state n, Pe ðnÞ, is defined as follows:
Pe ðnÞ ¼ limt ! 1 Pt ðnÞ: (3)
In equilibrium, the outgoing force is equal to the incoming force. Therefore, the
necessary condition for equilibrium is Eq. (4). See Aoki (1996, p. 140) for more details.
Keeping these in mind, we assume that the transition probabilities can be defined as
Eqs. (5) and (6).
n n
wn;nþ1 ¼ z1 h ; (5)
N 1 N
n n
wn;n1 ¼ z2 h ; (6)
N 2 N
where h1 ðÞ is a transition ratio that represents a probability that business conditions are
good; h2 ðÞ is a transition ratio that represents a probability that business conditions are bad
instead; z1 ðn=NÞ and z2 ðn=NÞ are interaction factors.
Y
n
z1 ðk=NÞ Y
n
h1 ðk=NÞ
Pe ðnÞ Pe ð0Þ (8)
z ðk=NÞ k¼1 h2 ðk=NÞ
k¼1 2
It is well known that the equilibrium probability converges to a Gibbs distribution under
certain conditions (Aoki, 1996, p. 140). Therefore, this equilibrium probability will take
the following functional form as in Eq. (9).
Pe ðnÞ ¼ Z 1 exp½bN Uðn=NÞ; (9)
where b is a constant and Z is a normalization constant. Z is given as Eq. (10).
X
Z¼ exp½bN Uðk=NÞ (10)
k
By construction, the equilibrium probabilities in Eqs. (8) and (9) should be equal, and by
using this, the unknown function, Uðk=NÞ, can be calculated as follows.
First, take logarithm on both Eq. (8) and Eq. (9), and equate the right hand sides of both
equations because these equations should be equal. Then, Eq. (11) follows.
Xn X n
n z ðk=NÞ h ðk=NÞ
bNU ln Z lnðPe ð0ÞÞ þ ln 1 þ ln 1 (11)
N k¼1
z2 ðk=NÞ k¼1
h2 ðk=NÞ
Note that Pe ð0Þ ¼ 1=Z under the condition Uð0Þ ¼ 0. Define DGðk=NÞ as Eq. (12) for the
convenience
of notation.
h ðk=NÞ k
ln 1 2bDG (12)
h2 ðk=NÞ N
484 J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495
n X
n 2 X k 1 n
1 k
U DG f (14)
N N k¼1 N N k¼1 b N
We can have various interaction factors. The bandwagon effect may be easily modeled
by assuming that the transition probabilities can increase as the share ðk=NÞ increases. Not
only bandwagon effects but also asymmetry of business cycles concern us, because the
asymmetry is a well known phenomenon. For the stylized facts of business cycles, see
Lucas (1977). This can be incorporated into interaction factors by assuming that the
interaction factors are asymmetric across the transitions. If the transition probabilities
asymmetrically increase as the share ðk=NÞ increases, then interaction factors, for example,
can be changed to the following:
n n
z1 ¼ (15)
N N
2
n n
z2 ¼ (16)
N N
From these specifications, we can handle the asymmetry of business cycles.
If N is large enough to make the variables continuous, both n=N and k=N can be
approximated as continuous variables y and z, respectively. Eq. (14) can be approximated
by Eq. (17).
Z y Z y
1
UðyÞ 2 DGðzÞ dz f ðzÞ dz; (17)
b
where f ðyÞ ¼ lnðyÞ.
After we calculate the equilibrium probabilities, we would like to have the first order
condition for maximizing the distribution across the states. Note that the value of function
UðyÞ is a key to determine the value of equilibrium probability in Eq. (9). The equilibrium
probability is maximized at the state where UðyÞ is minimized. The first order condition for
minimizing UðyÞ is given in Eq. (19).
dUðyÞ 1
2DGðyÞ f ðyÞ (18)
dy b
dUðyÞ
¼0 (19)
dy
J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495 485
At the state where the equilibrium probability is maximized, we can have the following
relationship as Eq. (20).
I
y¼ (23)
200
By plugging the values of business survey indices (I) in Eq. (23) into Eqs. (20)–(22), we
can calculate the transition ratios h1 ðyÞ and h2 ðyÞ.
After we extract the fundamental force (transition ratios) by filtering out interaction
factors, we can calculate the probabilities of each economic condition. However, we have
another problem. The problem here is that the data (BSIs) can have informational bias across
the different survey groups. The informational bias here means that some survey groups have
a tendency to exaggerate conditions, while others do not. If this is true, then the informational
bias is the major problem that keeps us from comparing the indices and studying cross
sectional investigation of the shock propagation. Therefore it is important to filter out the bias
and to adjust the information precision by the criteria that have been confirmed in the past.
Suppose that the state as to whether the economy is in an expansion phase or not can be
represented by the state variable (z). The state variable (z) is set to have discrete values of 0
and 1. If the state is an expansion period, then z is set to be 1, and otherwise, 0. Remember
that we define the transition probability as the multiplication of the probability of being in
an expansion phase (h1 ðÞ) and an interaction factor that will be filtered out. By definition,
we can establish the following relationship:
expðbDGðyÞÞ
Prfz ¼ 1g / (25)
expðbDGðyÞÞ þ expðbDGðyÞÞ
information precision of bDGðyÞ is not perfect and differs across survey groups. In order to
take the information precision into consideration, Eq. (24) can be changed into Eq. (26).
expðgbDGðyÞÞ
Prfz ¼ 1g ¼ (26)
expðgbDGðyÞÞ þ expðgbDGðyÞÞ
g is a nonnegative scalar, which represents the information precision. If the informational
precision of bDGðyÞ is high, g should be large. That is, if g is large, then small increment in
values will result in a big change in the probability of being in an expansion phase. If the
informational precision of bDGðyÞ is low, g should be small so that it may reduce the
change in probability. In the extreme case of g ¼ 0, the information does not play at all. In
the other extreme case of g ¼ 1, the slightest change in information can indicate the
situation for sure or with probability 1.
Now let us discuss the estimation method of the precision measure g. Before we
estimate the precision measure, we should have a criterion based on the data. In this paper,
we suggest reference dates as a criterion. By the reference dates, we define the state
variable (z). If a date is in an expansion phase, then we set z ¼ 1, otherwise z ¼ 0: The
mean of z is as follows.
EðzÞ ¼ Prfz ¼ 1g 1 þ Prfz ¼ 0g 0 (27)
It is easy to apply the proposed method to the real data. We can calculate particular
values of bDGðyÞ from Eqs. (20) and (23) at each time by using a real data set. Therefore,
there is no problem to calculate L. g can be estimated by maximizing the likelihood
function (L) with respect to g.
Note that we can easily transform Eq. (26) into logistic distribution function without a
constant term as follows:
expðg2bDGðyÞÞ
Prfz ¼ 1g ¼ (30)
expðg2bDGðyÞÞ þ 1
expðg lnðyÞÞ
Prfz ¼ 1g ¼ (31)
expðg lnðyÞÞ þ 1
See Greene (1990, p. 662) and Amemiya (1985, p. 268) for estimation details.
Regarding estimation of the state probabilities, there are other approaches such as
Hamilton (1989).
J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495 487
Korea National Statistical Office (NSO) announces the reference dates of Korea after
they analyze GDP, industrial production and other economic indicators. Because the
reference dates are official and announced long after the actual dates, they are reliable in
analyzing business cycles. We will use these dates as guidelines for assessing the different
business survey indices. In the previous six cycles, we can easily find out that the duration
of the expansion phase is consistently longer than that of the contraction phase. The
average expansion duration is 34 months, while the average contraction duration is only 19
months. This asymmetry is well handled in our set-up of interaction factors (Table 1).
The business survey indices by sectors have been announced by the Bank of Korea every
quarter since 1994. The business survey index of ‘all industries’ is from the third quarter of
1991 to the first quarter of 2003. Other business survey indices are from the first quarter of
1994 to the first quarter of 2003. Due to the data limitation, our analysis about sectoral
movements has to be mainly based on case studies. We recalculate the state by quarters
based on the monthly reference date data. The expansion phase starts from a trough and
ends in a peak. We coded the state according to the following rule. That is, a trough is not
included in an expansion phase, but a peak is included in the phase. On the other hand, in
the contraction phase, a peak is excluded, and a trough is included.
The relationship between the transformed business survey indices and GDP’s cyclical
component is very strong. We measure GDP’s components by applying Hodrick–Prescott
filter to seasonally adjusted GDP. It turns out that the transformed indices represent cyclical
Table 1
Reference dates of Korea
Cycles Trough Peak Trough Duration of Duration of
expansion contraction
Cycle 1 March 1972 February 1974 June 1975 23 months 16 months
Cycle 2 June 1975 February 1979 September 1980 44 months 19 months
Cycle 3 September 1980 February 1984 September 1985 41 months 19 months
Cycle 4 September 1985 January 1988 July 1989 28 months 18 months
Cycle 5 July 1989 January 1992 January 1993 30 months 12 months
Cycle 6 January 1993 March 1996 August 1998 38 months 29 months
Cycle 7 August 1998 August 2000 N.A. 25 months N.A.
Note: ‘N.A.’ denotes ‘not available.’
488 J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495
Table 2
Estimated information precisions
Industry g S.E. P -value
All industries 0.12 0.35 0.73
Fishing 0.17 0.40 0.67
Mining and quarrying 0.05 0.33 0.89
Manufacturing 0.18 0.38 0.64
Electricity, gas and steam supply 0.14 0.54 0.79
Construction 0.11 0.32 0.73
Wholesale and retail trade, repair of motor vehicles 0.19 0.36 0.58
Hotels 0.19 0.38 0.61
Transportation 0.23 0.34 0.50
Real estate 0.10 0.39 0.80
Recreational, cultural and sporting activity 0.18 0.37 0.63
components well. The results of Granger causality test for the whole period showed that the
GDP cyclical components Granger cause the transformed business cycles indices; and the
test for the period after the foreign currency crisis shows that the transformed indices
Granger cause the cyclical component at the 10% significant level at almost all lags.2
We calculate the transformed business survey indices by sectors. The transformed
business survey indices are strongly correlated with the original counterparts. For example,
the correlation coefficient of business survey index of ‘all industries’ and the transformed
one is 0.99. Therefore, we can argue that the information loss due to transformation is very
small.
By using Eq. (28) and applying logit estimation method, we can estimate the
information precision (g). Table 2 shows the information precision of business survey
indices of each industry. Unfortunately, the results are not strong enough, but they do give
us a broad picture about the informativeness of the indices. As we expected, the most
informative indices are those of ‘transportation’, ‘wholesale and retail trade, repair of
motor vehicles’ and ‘hotels’. In other words, those industries are very sensitive to business
conditions.
3.2. Two episodes of the recent business cycle: transformed business survey index
analysis
that R2 s are above 0.90 for all cases. From this VAR estimation, we calculate the variance
decomposition and the result for the ‘all industries’ case, ‘manufacturing’ industry explains
15.5% after 20 quarters, which is the largest portion among the explanatory variables. In
order to answer the second question, we will trace the movements before and after a peak
and a trough. We believe that survey indices carry qualitative information of market
sentiments. From the data limitations, we have only two episodes; one is the peak of year
1996, and the other is the trough of year 1998. Figs. 1 and 2 show all industries’ paths
before and after the reference dates. From Fig. 1, ‘construction’ industry and
‘manufacturing’ industry show the contraction three quarters before the peak, while
‘construction industry’ bounced back one or two quarters after the trough. ‘Electricity, gas
and steam supply’ shows up and downs. Both ‘hotels’ and ‘wholesale and retail trade,
repair of motor vehicles’ show an increase for one quarter and a continued decrease
afterwards. Here, ‘manufacturing’ industry is the leading sector in the business cycle
around the peak of 1996. A similar phenomena can be confirmed during the episode of the
trough of year 1998. As shown in Fig. 2, ‘hotels’ industry, ‘wholesale and retail trade,
repair of motor vehicles’ industries, and ‘transportation’ industry under-performed during
1 year before and after the trough. However, ‘hotels’ industry, and ‘wholesale and retail
trade, repair of motor vehicles’ industries showed a rapid recovery, and ‘electricity, gas and
steam supply’ showed dramatic changes in the trend before and after the trough of year
1998. Industries such as ‘recreational, cultural and sporting activities,’ ‘fishing,’
‘construction,’ and ‘real estate’ showed slightly delayed responses. ‘Mining, gas and
steam supply’ industry did not show any changes either in the peak or in the trough, which
make us believe that that industry did not play any significant role over the business cycle.
‘Real estate’ industry showed very sluggish changes near the peak, while it showed one
quarter delay but a rapid response near the trough. In two episodes, ‘real estate’ and
‘mining, gas and steam supply’ industries played a role in stabilizing the volatility of
business cycles. ‘Transportation,’ ‘hotels,’ and ‘wholesale and retail trade, repair of motor
vehicles’ industries have under-performed both at the peak and at the trough. Another
contrasting phenomenon is the asymmetric responses of industries. Near the peak of year
1996, there were no dramatic changes, while most industries showed a quick turn from the
trough of year 1998. These transformed business indices show the co-movements toward
an expansion phase are strong. Even though we filtered out the asymmetric effect by
contagion, the impact of the 1997 foreign currency crisis on the economy was so deep that
the reaction against it might be stronger.
As the conventional way of thinking suggests, it turns out that ‘manufacturing’ industry
lead the reference date. The transformed BSI of ‘manufacturing’ industry showed
contraction by two quarters before the peak of 1996, and it saw a rapid recovery from one
quarter after the trough of 1998. Table 3 shows that manufacturing industry maintained the
consistent movements of ‘all industries’ and lead the reference date.
From the correlation coefficients analysis, we have strong co-movements between
‘manufacturing’ and ‘all industries’, whose correlation coefficient is 0.996. As we
expected, co-movements between ‘all industries’ and other industries are not relatively
strong. The correlation coefficients of ‘all industries’ with ‘wholesale, etc.’ and ‘hotels’ are
around 0.95, and others are much less than this. Summarizing all the movements, it is
J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495 491
Table 3
Changes in transformed business survey indices at the turning points
3 2 1 0 1 2 3 4
Near the peak of 1996
All industry + +
Fishing 0 + +
Mining and quarrying 0 + + + +
Manufacturing + +
Electricity, gas and steam supply + +
Construction + +
Wholesale and retail trade, repair of motor + 0
Hotels + + + +
Transportation + + +
Real estate 0 +
Recreational, cultural and sporting activity + + +
Near the trough of 1998
All industries + + + +
Fishing + + + + +
Mining and quarrying + + + + +
Manufacturing + + + + +
Electricity, gas and steam supply + + +
Construction + + +
Wholesale and retail trade, repair of motor + + + + +
Hotels + + + + +
Transportation 0 + + + + +
Real estate + + +
Recreational, cultural and sporting activity + + + +
Note: The number represents the number of quarters before and after the turning point. ‘ ’ means a decrease, ‘0’
means no change, and ‘+’ means an increase.
reasonable to believe that ‘manufacturing’ industry has been leading the business cycle in
recent business cycles.
Table 4
Principal component analysis of manufacturing sectors
Component 1 Component 2 Component 3
Eigenvalue 17.383 0.932 0.520
Variance prop. 0.828 0.044 0.025
Cumulative prop. 0.828 0.872 0.897
Variable Eigenvectors
Vector 1 Vector 2 Vector 3
Food products and beverages 0.233 0.101 0.175
Textiles 0.187 -0.530 0.026
Wearing apparel and fur articles 0.221 0.011 -0.202
Leather, luggage, handbags, saddlery, etc. 0.207 0.088 -0.380
Wood and products of wood, etc. 0.217 -0.038 -0.230
Pulp, paper and paper products 0.216 0.066 0.363
Publishing, printing and reproduction of 0.215 0.371 -0.047
Coke, refined petroleum products, etc. 0.201 0.438 0.287
Chemicals and chemical products 0.233 -0.072 -0.023
Rubber and plastic products 0.230 -0.064 0.102
Non-metallic mineral products 0.220 -0.145 -0.180
Basic metals 0.233 -0.042 -0.022
Fabricated metal products, except machinery 0.231 0.049 -0.069
Machinery and equipment n.e.c 0.236 0.011 -0.098
Office, accounting and computing machinery 0.220 0.179 0.013
Electrical machinery and apparatus n.e.c 0.223 -0.148 -0.086
Radio, television and communication equipment 0.224 -0.157 0.005
Medical, precision and optical instrument 0.185 -0.478 0.319
Motor vehicles, trailers and semi-trailers 0.231 0.102 -0.025
Other transport equipment 0.209 0.081 -0.337
Furniture and manufacturing n.e.c. 0.209 0.086 0.483
3
Even though ‘foods and beverages’ shows a high coefficient, we disregard it because this is categorically too
different from the others.
J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495 493
Fig. 3. Five sectors in the manufacturing industry around a peak: the first quarter of 1996.
Fig. 4. Sectors in the manufacturing industry around a trough: the third quarter of 1998.
trough, but after the turning point, it under-performed among five sectors. In connection with
the previous discussion, we can say that all these five sectors played active roles in the most
recent business cycle, judging from the fact that they led the cycle of the aggregate
manufacturing industry.
494 J.-M. Yang, T.-W. Kim / Japan and the World Economy 17 (2005) 480–495
Acknowledgements
We would like to thank anonymous referees for helpful comments. Needless to say, all
remaining errors are our own.
References