Abstract: CAPM model cluster assume that the stochastic error term is homoscedasticity, which can not effectively characterize the conditional heteroscedasticity of financial asset return volatility. In order to use this kind of model correctly in empirical research, two improvements are proposed:GARCH model is introduced to describe the fluctuation law of disturbance term, instead of the traditional assumption of normal distribution; Conditional heteroscedasticity is combined into the model as a risk factor. Empirical research is also developed on the actual data set of American banking industry, and the results show that the improvement of this paper is reasonable, which provides a useful reference for empirical analysis of using such models.