Due to the characteristics of short shelf life and perishable, fresh products require high accuracy and reliability of short-term sales forecast. This paper carried out feature engineering analysis involving time, pricing, competitive product pricing, freshness and other micro-level factors. On this basis, we proposed the ARIMA-NARX combination forecasting model of fresh product sales. This combination forecasting model took advantage of the ARIMA to capture the linear rule in the sales time series, and adopted the NARX to describe the nonlinear relationship in the ARIMA residual with the feature matrix processed by feature creation and dimension reduction. Then, the NARX residual prediction result was used to correct the predicted sales of ARIMA. Finally, the prediction results of the combined model were compared with the real observation values and prediction results of ARIMA, NARX, ARIMA-NAR, SVM and Regression Tree models. The MSE/MAPE value and DM test verified the rationality and effectiveness of the ARIMA-NARX model, which could improve the prediction accuracy of short-term fresh product sales significantly.