Measuring the Chinese business cycle

Abstract

This article documents the business cycle characteristics of the Chinese economy by adopting both nonparametric and parametric methodologies. The two approaches are applied to relevant macroeconomics indicators – Gross Domestic Product (GDP) and Industrial Production (IP) indices – aiming to investigate the growth cycle (deviation cycle). We provide a clear chronology of the Chinese growth cycle. One significant characteristic of the Chinese growth cycle is the relatively direct influence of government policies. However, recently these policy effects have become less significant when compared to global economic influences. Our study provides an enhanced understanding of the properties of business cycle dating algorithms and as such contributes to future Chinese business cycle research.

Keywords:

• business cycle
• Markov switching
• nonparametric

JEL Classification::

• E32
• C22
• O53

Notes

¹ One usually uses the GDP to characterize the business cycle as it is seen as the most comprehensive statistic to describe economic activity, but one could indeed use other series, such as industrial production, which is less comprehensive but has better data availability.

² See Harding and Pagan (2005) for more details, but briefly this can be seen by setting P _t equal to y _t−1.

³ Social demand includes government taxes, profits of enterprises and income of residents

⁴ Source: Data are downloaded from the NBSC Website, macroeconomic climate indices.

⁵ They use disaggregated annual Chinese GDP for the period from 1979Q1 to 1994Q4, which were taken from Abeysinghe and Rajaguru (2004) using Chow–Lin disaggregation.

⁶ Source: Data from 1952 to 2004 come from the China Compendium of statistics 1949–2004, Tables 1–9, Indices of Gross Domestic Product; data from 2005 to 2007 come from the China Statistical yearbook 2008, Tables 3–4, Indices of Gross Domestic Product. These data were converted to the same base year (base year 1978=100).

⁷ Source: DataStream, CH GDP (Year to date% change) CONN.

⁸ Source: World Development Indicators (December 2008), Industry, value added (annual% growth).

⁹ Source: Data from 1990M02 to 1991M01, and 1993M02 and 1993M09 come from the China Economic Information network; data from 1991M12 to 2002M10 come from the International Monetary Fund (IMF), International Financial Statistics (IMF Website); data from 2002M11 onwards come from the NBSC website (the table called the value added of industry, which is consistent with the data from the IMF). Data in 2006M01, 2007M01, 2008M01 and 2009M01 are missing.

¹⁰ IP measures are available at a higher frequency, which is especially important for parametric dating methods. As industrial production is considered to be a coincident indicator and in many countries constitutes a very significant proportion of GDP, it is commonly seen as a good proxy for overall economic activity. See also Agénor et al. (2000) and Rand and Tarp (2002).

¹¹ This is based on authors' calculation from the annual GDP index.

¹² For this reason unit-root tests are not reported here, but are available upon request.

¹³ For simplicity, we did the above adjustment to deal with the missing data problem.

¹⁴ We find that this very simple approach delivers a plausibly smooth series. We also attempted a somewhat more complicated approach, assuming that the Chinese Spring Festival impacts economic activity for roughly 40 days: 20 days before the Festival day and 20 days after the Festival day (including the Festival day). Roughly, the more effective days fall in a particular month, the smaller IP index for that month should be. Hence, we generated a dummy variable to represent the days affected by the Chinese Spring Festival. However, this method does not remove the spikes in the original data series as well as the simple methodology described.

¹⁵ The algorithm ensures that there is a minimum time distance between peaks and troughs. Refer to the application for a discussion of this point.

¹⁶ ‘Contraction’ in this context does not imply negative growth, but rather decreasing growth.

¹⁷ The BBQ rule is proposed by Harding and Pagan largely for two reasons. First, much of the parametric dating work deals with quarterly data. Second, there exists some limitations in the B&B procedure, such as the repeated smoothing of the data which can lead to data being detrended. The intention of smoothing in the B&B procedure is the removal of some idiosyncratic variation during the identification process of peaks and troughs. The smoothing process is ignored once applied to quarterly data. See Harding and Pagan (2002) for more details.

¹⁸ We use Gauss code downloaded from Harding's Web Page. A STATA version is available for download from Phillipe Bracke's webpage at the LSE.

¹⁹ The IP index is a slightly shorter series than the GDP, starting in 1961. The BBQ method has a rule of eliminating turning points from the beginning and the end of the series, hence failing to identify a trough in the IP series around 1961/1962.

²⁰ We use Gauss code written by Watson (1994) and downloaded from his website (http://www.princeton.edu/~mwatson/).

²¹ From this it is apparent that the minimum phase restriction parameter acts very much like a bandwidth parameter that controls the amount of local information that enters the final dating result. We are grateful to a referee for pointing this out.

²² We use the OX code written by Hans-Martin Krolzig, package of MSVAR 1.32a for Ox 3.4 (Krolzig's Workpage). As we are only concerned with the univariate case, the MS Vector Autoregression (MSVAR) model has been reduced to an MS Autoregression (MSAR) model. We do not have exogenous variables either.

²³ Statistical inference in three regime models with moderate sample size is difficult, but in any case, not of primary interest for the work at hand, which is why we only report the average growth rates, transitional probabilities and smoothed regime probabilities.

²⁴ Depending on the data frequency, MS models categorize quarters or months into the different regimes. In this table, we only illustrate annual categorizations and hence there is some loss of information. For more accurate regime classification, refer to the smoothed probabilities plotted in Fig. 10.