Aiming at the limitation of ARIAM model in predicting stock prices with relatively complex fluctuation trends, this paper proposes an improved ARIMA method (BMA-ARIMA) based on B-spline basis expansion and model averaging. The proposed method uses intraday prices as auxiliary information and considers its functional characteristics. By combining the two methods of parallel model averaging and recursive model averaging, the intraday price function features extracted based on the B-spline expansion method can be used to the greatest extent to fit the residuals predicted by the ARIMA model, thereby improving the prediction accuracy of ARIMA. In addition, since the regression prediction method in the model averaging stage is model-free, when the nonlinear regression model is selected, it can capture the linear and nonlinear information of the intraday price function characteristics. Specifically, this paper uses CART as the base model, and the actual data analysis results show that BMA-ARIMA can improve the prediction accuracy of the ARIMA model and has certain robustness. Finally, the method can theoretically be extended to time series forecasting in the fields of medicine, water conservancy, electric power and environmental science.