Since there are some drawbacks with the existing backtesting methods based on VaR and expected shortfall, in this paper we apply a new backtesting method, the saddle-point technique, to the daily returns of three stock indices in China. The new and traditional backtesting methods are implemented on three different models, and the results show that the simple GARCHNormal model can't capture the market risk of stock index in China, and this serves as a warning for risk modeling based on normal distribution which is quite common in practice. Moreover, the advantage of the saddle-point technique lies in that it allows us to carry the backtesting with the annual share index yields.