GARCH―GPD for Financial Risk Measure Based on Hybrid Genetic Algorithm.docVIP

GARCH―GPD for Financial Risk Measure Based on Hybrid Genetic Algorithm 【Abstract】The shape parameter and scale parameter of generalized Pareto distribution are estimated by hybrid of genetic algorithm and pattern search. The volality of the return is obtained by GARCH model. VaR and CVaR are computed respectively under GPD model and GARCH-GPD model. The experimental results show that VaR and CVaR based on GARCH-GPD are more effectively measure the financial risks. 【Key words】GPD； Genetic algorithm； Pattern search； Risk measure 0 Introduction Extreme value theory especially peaks over threshold model （POT） is a safe financial risk measure method[1-2]. The key and difficulty in POT model are estimating the parameters of generalized Pareto distribution（GPD）. It has been confirmed value at risk（VaR） or conditional value at risk （CVaR） based on GPD model is more effective than normal distribution[3-5]. GARCH model can describe the clustering feature of the financial return and hence it is also a good tool for measuring VaR and CVaR[6-7].If the error distribution in GARCH is not normal distribution but GPD， then GARCH-GPD model is obtained [8-10]. Genetic algorithm （GA） has good global optimal performance. Pattern search（PS） is a good local optimization method but it needs an initial point. This paper proposes that we can combine the advantages of GA and PS in the parameter estimation of GARCH-GPD model. 1 Peaks Over Threshold Model Table 3 shows us that CVaR is higher than VaR in each model （GPD， GARCH and GARCH-GPD） under the same confidence level. This indicates that VaR is not enough sufficient for risk measure. At the same time， it can be seen clearly that VaR and CVaR computed by GARCH-GPD is the largest； GARCH is the second and GPD is the smallest. Though GPD can effectively fit the tail of the return， it ignores the heteroskedasticity which results in the lowest VaR and CVaR. GARCH can describe the clustering feature of the return and has higher VaR and CVaR. GARC