Title: Multivariate extensions of the ACD peaks-over-threshold method for forecasting value at risk
Authors: Katarzyna Bien-Barkowska - Warsaw School of Economics (Poland) [presenting]
Abstract: A new dynamic peaks-over-threshold (POT) model is proposed for extreme events in financial markets. The random times when the sizes of negative financial returns exceed given threshold are modeled in line within the marked point process theory, where the marks correspond to the magnitudes of extreme losses. We develop a multivariate version of the autoregressive conditional duration (ACD) model, where the conditional intensity of extreme negative returns has not only the self-exciting structure, but also the cross-exciting structure, since it can instantaneously react to the time-varying covariates such as large positive returns or volatility peaks. In our approach the observed times of all these intervening events can accelerate or decelerate the awaited occurrence of extreme losses. We apply the extended multivariate version of the ACD model to six major stock indexes and show that it outperforms the standard ACD-based POT methods for forecasting value-at-risk and expected shortfall.