Tranquil and crisis windows, heteroscedasticity, and contagion measurement: MS-VAR application of the DCC procedure

Abstract

The key objective of this study is to show that two potential shortcomings of the Determinant of Change in Covariance (DCC) matrix procedure of Rigobon (2003), namely with the arbitrary determination of the windows, i.e. tranquil and crisis periods and the violation of its heteroscedasticity assumption under the null, can be simultaneously addressed via a simple incorporation of a Markov-switching vector autoregressive approach into the overall DCC procedure. To demonstrate this, we revisit the period around the time of the East Asian crises using daily stock exchange of Indonesia, Malaysia, Philippines, Thailand, Singapore, Korea, Hong Kong and Taiwan, and test whether there is a significant break or discontinuity in the stock exchange returns of the eight East Asian markets during crisis periods, especially around the time of the 1997 financial crises. In contrast to that of Rigobon (2003), our results show that the propagation of shocks shifted significantly starting with the onset of the sharp decline in the Hong Kong stock market.

Notes

¹ See Lee and Kim (1993), Calvo and Reinhart (1996), Frankel and Schmukler (1996), Valdez (1997), Baig and Goldfajn (1999) and Agenor et al. (1999). See Pericolli and Sbracia (2003) for the usage of the terms ‘correlation breakdowns’.

² In addition, Dungey and Zhumabekova (2001) emphasized that the adjustment in the correlation coefficient suffers from low power when typically dealing with a relatively small crisis sample. More recently, Corsetti et al. (2005) assert that the results obtained by Forbes and Rigobon (2002) of the lack of evidence of contagion in almost all cases can be attributed to arbitrary assumptions on the variance of the market-specific noise in the country where the crisis started. As they demonstrate, the adjustment in the correlation coefficient is biased towards the null hypothesis of interdependence, that is, the null hypothesis is erroneously accepted in a number of cases, when it should be rejected in favour of contagion.

³ See also Caporale et al. (2005).

⁴ The presentation of the DCC procedure in this section follows closely that of Rigobon (2003).

⁵ As also mentioned in Billio et al. (2003), the discussion of this caveat in Rigobon (2003) is not necessary since almost no rejections were found anyway in his paper.

⁶ This is also another way of saying that the common unobservable shock (z) is homoscedastic for the rejections to be interpreted as parameter instability.

⁷ For instance, Billio and Pelizzon (2003) clearly documented that inferences based on heteroscedasticity-adjusted conditional correlation coefficients, with the choice of the crisis and tranquil windows exogenously determined, are highly sensitive to varying lengths of the tranquil and crisis windows. See also Boyer et al. (1999), Dungey and Tambakis (2005) and Pericolli and Sbracia (2003) for recognition of this problem with previous studies on contagion.

⁸ The choice of these countries basically follows the same eight Asian countries that Rigobon (2003) analysed to conduct the DCC test on a multivariate (regional) framework.

⁹ The Kwiatkowski-Phillips-Schmidt-Shin (KPSS) unit-root tests have also been conducted and the results are consistent with those of the Augmented Dickey-Fuller (ADF) reported in Table 1. For the sake of brevity, we do not report the KPSS results.

¹⁰ In other words, as also clarified in Rigobon (2003), if the mass is too small or too large this implies the determinant is different from zero. Thus, conclusions can already be made by simply observing the mass either above or below zero.

¹¹ Note that the Korean crisis falls within that of the crisis in Hong Kong. Hence, we do not separate the two cases of crises.

¹² See Table 5, p. 275 of Rigobon (2003).

¹³ This is a rejection only at the 20% confidence.

¹⁴ We also tried different robustness test where we introduce interest rates as controls, changes in the starting sample period as well as in the end-sample period did not change the qualitative results of the DCC test version from an MS-VAR, wherein we find rejections of the null of stability either the 20% or stronger.

¹⁵ Caporale et al. (2005) makes the same conclusion, see p. 486.