Measuring Contagion Risk in High Volatility State between Major Banks in Taiwan by Threshold Copula GARCH Model

This paper aims to study the structural tail dependences and risk magnitude of contagion risk during high risk state between domestic and foreign banks. Empirically, volatility of stock returns has the properties of persistence, clustering, heteroscedasticity, and regime switchs. Thus, the threshold regression model having piecewise regression capability is used to classify the volatility index of influential foreign banks as “high” and “low” of two volatility states to further estimate Kendall taus i.e. structural tail dependences between banks using three models: Gaussian, t, and Clay copula GARCH. Using fuzzy c-means method, both domestic and foreign banks are categorized into 10 groups. Through the groups, 5 domestic and 7 foreign representative banks are identified as the research objects. Then, the in-sample data of daily banks’ stock prices covering 01/03/2003 ~06/30/2006 are used to estimate the parameters of threshold copula GARCH model and Kendall taus. The out-of-sample data covering 07/01/2006~03/25/2014 are used to estimate the Kendall taus gradually using rolling window technique. Several research findings are described as follows. In high state, the tail dependences are two times much larger than in low state and most of them have up-trend property after sub-prime crisis and reach the peak during Greek debt. It implies that the volatility is high in risk event and the contagion is high after risk event. In high state, HNC has the highest tail dependences with foreign banks but its value at risk is the lowest. It can be considered as an international attribute bank with lower risk. On the contrary, YCB and FCB have the lower tail dependences with foreign banks but their value at risks are quite high. They are viewed as a local attribute bank with higher risk. The Bank of American, Citigroup, and UBS AG have the relatively higher value at risk. Citigroup has been tested to Granger cause ANZ and all domestic banks. It is necessary to beware the contagion risk from Citigroup. Among three models, in low state, Gaussian and t copula models have the better significance of estimation than Clay copula model. However in high state, Clay copula model has the same acceptable estimation and more capability to uncover the instant nonlinear jumps of tail dependences while Gaussian and t copula models reveal the linear changes of tail dependences as a curve.