The short-term export volume forecast is of great significance for international trade decision-making and macroeconomic regulation, but the time series of export volume usually contains significant seasonal, trend and nonlinear fluctuation characteristics at the same time, so it is difficult to obtain ideal results with a single forecast model. In order to improve the prediction accuracy and stability, this paper proposes a hybrid prediction method combining seasonal autoregressive moving average model (SARIMA) and attention based memory network (attention based LSTM). Firstly, SARIMA model is used to describe the linear structure and seasonal components in the export volume series, and its prediction residual is modeled; Then, the attention LSTM is used to learn the nonlinear dynamic characteristics in the residual sequence, and finally the prediction results are obtained by additive fusion. The experimental results based on the monthly export volume data of Poland show that compared with the traditional SARIMA, LSTM and other comparative models, the MAE, RMSE and MAPE of the SARIMA-Attention-LSTM (Proposed) on the test set are reduced, on average, by about 20%–35%, respectively. The prediction residual fluctuation converges significantly, and shows better stability and generalization ability in repeated experiments. The results show that the effective integration of statistical model and deep learning model can significantly improve the short-term export forecasting performance, and provide a feasible and efficient solution for the prediction of complex economic time series.