Bivariate tail risk analysis for high-frequency returns via extreme value theory

Quantifying the nature of extreme value dependence in high-frequency fluctuations of asset prices is an important yet difficult problem. In this work, we propose a two-stage estimation procedure for conditional joint distribution of high-frequency extremes, given past information on returns. The mo del combines an intraday volatility component and GARCH model for marginal time dependence with a tail dependence model for extreme values which is based on the framework of regular variation. Examining 15-second returns of four banking sector securities, we find that there exists tail dependence in the detrended residuals. The proposed model outperforms a benchmark Gaussian model in predicting conditional value-at-risk and expected shortfall, as well as in predicting the probability of jointly extreme returns.