Title: Portfolio optimization based on multivariate GARCH copula models
Authors: Maziar Sahamkhadam - Linnaeus University (Sweden) [presenting]
Abstract: Multivariate GARCH models in combination with vine copula are tested to forecast one-step ahead stock index returns and compute optimal portfolio weights. The common marginal distributions in multivariate GARCH models are multivariate normal and Student $t$, which are not able to capture the tail dependency structure required in modeling fat-tailed financial returns. To model tail behavior, the Clayton canonical vine copula is used, which captures the lower asymmetric tail dependence with uniform marginal obtained from dynamic conditional correlation and generalized orthogonal GARCH models. Moreover, we test extreme value theory in modeling the downside risk in multivariate GARCH-Copula forecasting settings. Three portfolio strategies including minimum Conditional Value-at-Risk (CVaR), maximum reward-to-risk (SR) and Markowitzs mean-variance (MV) are obtained and back-tested over an out-of sample period, which contains financial crisis. The results show out-performance based on DCCGARCH and GOGARCH models comparing to univariate GARCH. In particular, the results of VaR back-testing and risk-adjusted performance show improvement in forecasting the downside risk for CVaR portfolio strategy obtained by GOGARCH canonical Clayton vine model with semi-parametric marginal distribution (based on EVT). This model also leads to a better economic performance for SR portfolio strategy over a long-horizon investment period.