The performance of the Markov-switching model on business cycle identification revisited

Abstract

This study examines the performance of Markov-switching model on business cycle by applying the model to various economies. Specifically, three comparison groups are used: (1) the USA and Japan serving as the representatives for the industrialized economies (or IEs hereafter); (2) Taiwan and South Korea serving as the representatives for newly industrialized economies (or NIEs hereafter); and (3) Malaysia and Indonesia serving as the representatives for the developing economies (or DEs hereafter). The empirical results are consistent with the following notions. First, the Markov-switching model serves well to depict the business cycles for IEs and DEs. Nevertheless, the model is ineffective for the two NIEs, which underwent structural economic shifts to slower growth during our sample period of 1970–1998. Second, the two-period Markov-switching by dividing the sample periods into two sub-periods thus more effectively measures the two NIEs’ business cycles.

Acknowledgements

The authors are grateful to James D. Hamilton for his invaluable comments and suggestions.

Notes

The IEs’ per capita GDP are significantly greater than the NIEs’, which in turn are significantly greater than the DEs’. For instance, the 1999's per capita GDPs of the US and Japan (the two IEs) are 34 047 and 34 360 dollars, respectively. The 1999's per capita GDPs of South Korea and Taiwan (the two NIEs) are 8684 and 14 024 dollars, respectively. The corresponding measures of Malaysia and Indonesia (the two DEs) are 3242 and 1055 dollars, respectively.

Goodwin (1998) provides evidence supporting the notion that the MS models incorporating a quasi-Bayesian prior effectively depict the business cycles for UK, Japan and France, but not for Italy. Goodwin (1998), nevertheless, does not provide any explanations or remedies as to the ineffectiveness of the MS models for Italy.

Prior studies on the ineffectiveness of the MS models do not specify the particular types of the economies that may encounter such problems. Nor do they provide any explanations or remedies to mitigate the ineffectiveness problems.

In such a case, state 1 is an absorbing state and the Markov chain is reducible.

In a sensitivity test, we use higher order settings, which yield insignificantly improvements over the lower order settings. On the other hand, we think that the effectiveness of identifying business cycles does not significantly relate with the lag order numbers.

, if r = 0: filtering probabilities; r<0: predicting probabilities; r>0: smoothing probabilities.

Boyden, Fletcher, Goldfarb, and Shanno (BFGS) algebra is effective in deriving the estimators to maximize the value of the likelihood functions for non-linear models. See Luenberger (1984).

IP is used as opposed to GDP in order to obtain a large enough sample of the post-1990 business cycles for the two NIEs. Specifically, IP (GDP) measures are reported monthly (quarterly). Because of the need to separate data into two periods to control the structural shifts of South Korea and Taiwan, it appears to be desirable not to over-parameterize the model. Furthermore, the production functions most typically incorporate labour as an input variable and can thus reflect the level of unemployment of an economy.

The months surrounding the time when the MS models fail to depict the business cycles of the two NIEs to partition the South Korea and Taiwan IPs data have also been adopted. None of the alternatives appears to be more effective in depicting the business cycles.

The potential reasons that the two NIEs’ economic growth and volatility dropped substantially include exchange rate factors, the change of product modes, the more mellow industry's incorporating, the more stable price for input factors and output goods and so on.