Testing and comparing the performance of dynamic variance and correlation models in value-at-risk estimation

Abstract

This study addresses and examines certain advanced approaches for value-at-risk (VaR) estimation. In particular, we employ a multivariate generalized autoregressive conditionally heteroskedastic (MVGARCH) model involving time-varying settings and multivariate Markov switching autoregressive conditionally heteroskedastic (MVSWARCH) model with regime-switching techniques and compare them with a conventional linear regression-based (LRB) model. Our empirical findings are as follows: First, while the LRB VaR model behaves reasonably well in tranquil periods, it significantly underestimates actual risk during unstable periods. Second, in comparison with the LRB VaR model, MVGARCH- and MVSWARCH-based VaR models do better under unusual conditions, whereas better models are needed to estimate VaR. Third, dynamic variance settings improve the accuracy of VaR estimates. However, the effect of dynamic correlation designs on VaR is marginal.This study addresses and examines certain advanced approaches for value-at-risk (VaR) estimation. In particular, we employ a multivariate generalized autoregressive conditionally heteroskedastic (MVGARCH) model involving time-varying settings and multivariate Markov switching autoregressive conditionally heteroskedastic (MVSWARCH) model with regime-switching techniques and compare them with a conventional linear regression-based (LRB) model. Our empirical findings are as follows: First, while the LRB VaR model behaves reasonably well in tranquil periods, it significantly underestimates actual risk during unstable periods. Second, in comparison with the LRB VaR model, MVGARCH- and MVSWARCH-based VaR models do better under unusual conditions, whereas better models are needed to estimate VaR. Third, dynamic variance settings improve the accuracy of VaR estimates. However, the effect of dynamic correlation designs on VaR is marginal.