Analyzing the trend and fluctuation of time series and making the interval forecast of USD/EUR exchange rate. It is of great value to improve the current accuracy of the forecasting method based on the trend. A BP neural network was used to extract the trend and the volatility was analyzed by using a auto-regressive moving average model and a generalized auto-regressive conditional heteroscedasticity model. Finally, the trend and volatility were combined to give the forecast. By the study of the USD/EUR exchange rate from July 2001 to October 2017, it is found that the BP neural network has a good non-linear characterization ability, but only the appropriate prediction accuracy can lead to a better prediction interval and the regression conditional heteroscedasticity model is superior to the auto-regressive moving average model for the analysis of volatility. The accuracy of the combined model can be improved by adjusting the parameters, errors and prediction accuracy of the BP neural network.