The global financial crisis and its transmission to Asia Pacific

Abstract

By virtue of the US markets closing later than Asia-Pacific (AP) markets, returns observed in the US include information not reflected in AP until the next day. Provided there is enough integration among markets, this asymmetry should then generate Granger causality from the US returns towards returns in the AP region. It is thus obvious that when testing for Granger causality among the Western and AP markets, interdependence due to non-synchronicity should be clearly identified and factored out. Our novel method is to measure contagion, usually defined as excessive market integration during crises, as excess interdependence beyond the non-synchronicity induced. We test for contagion of the 2008 Financial Crisis from the US to AP. The unmistakable emergence of a new cointegration relation that intertwines the US and AP markets only during the crisis, but not before, presents evidence of long-run contagion, that is a new channel that transmits crisis-related info. We find strong contagion from US to Korea, no contagion to Hong Kong, and Australia is a mixed case. Our explanation is that the Korean economy is the most vulnerable due to its export orientation, Hong Kong becomes disentangled due to its relation to China, which is consistent with the theory of ‘decoupling’ of China, and the Australian case is somewhere in the middle, probably due to its close ties to the US but also its strong reliance on natural resources and gold.

Keywords:

• Financial market integration
• non-synchronicity
• Granger causality
• cointegration
• contagion

Notes

1. In principle, we could include all countries in a single five-vector system. This would not be a parsimonious approach, and it would diminish the power of the tests. Furthermore, it is difficult to imagine that the Australian markets carry extra systematic info for Korea beyond what is already included in the US and Japan returns.

2. In the second period we use a VECM model because markets become cointegrated.

3. The critical values of a model with more deterministic parameters than needed are negative and smaller than the critical values of the right model specification. Thus, adding more deterministic parameters increases the critical values (in absolute value) making it harder to detect stationarity.

4. In principle, we could include all countries in a single five-vector system. This would not be a parsimonious approach, and it would not enhance the power of the tests. It is difficult to imagine that the Australian markets carry extra systematic info for Korea beyond what is already included in the US and Japan returns.

5. We add a constant term in equation 3 in order to represent the drift component as it derives by the unit root analysis. Similarly, we do not include any drift for the NC VECM in equation 4 as it does not follow by our unit root tests.

6. Cheung and Lai (1993) based on Monte Carlo tests find that the trace test is more robust to both skewness and excess kurtosis than the maximum eigenvalue test. For an analytical discussion see Harris and Sollis (2003, p. 175).

7. We denote by r the number of cointegrated relationships and by n the number of markets included in a system.

8. With AP we designate any of the other three markets AP ∊ {HK, KR, AU}.

9. We estimate VAR systems for the BC period because the markets do not formulate any long-run equilibrium, so there is no sense in estimating VECMs. All systems in the VAR/VECM framework are estimated based on single differenced data (returns) and any restriction is tested based on the lagged first difference data (lagged returns).

10. Johansen et al. (2000) suggest that we use the HQ criterion if different information criteria suggest different lag values.

11. Additionally, there is one more condition that has to be satisfied before we proceed with the standard (TY) analysis, namely that ‘…the order of integration of the process does not exceed the true lag length of the model…’ (Toda & Yamamoto, 1995). Clearly, that is the case for this paper since the true lag length (k) is always greater than one for all models in both periods.