Abstract: In view of the peaked and fat-tailed characteristics of financial return data distribution and its effect of clustering fluctuation and especially the 'leverage effect' of fluctuation on VaR (value at risk) estimates and some efficiencies when estimating VaR with various assumptions of return data distribution, a semi-parameter approach based on EGARCH-VaR model is developed. This model is then compared with the approach based on to measuring VaR on the basis of GARCH model which is assumed to be normal/t distribution. An empirical analysis in combination with posterior testing is done on Chinese stock market risk, which shows that during VaR measuring the semi-parameter approach is superior to the approach based on GARCH model which is assumed normal/t distribution.