Considering that the financial time series usually exhibit leptokurtosis, fat-tail, bias, clustering and leverage in reality, a quantile regression approach is applied to estimate the VaR (Value at Risk) in this paper. We build the quantile regression VaR model to compute VaR of Shanghai Composite Index, based on GARCH family models and different distributions of the residual. The back-testing is used to evaluate the performance of VaR estimations. The comparisons under various assumptions indicate that the proposed quantile regression VaR model is insensitive to both the model and the distribution of residual, which signifies that the model has excellent applicability. In addition, this paper suggests that the longer the holding period is, the better the performance is.