Sales forecasting has always been a hot research topic and has great significance for all enterprises. In recent years, with the rise of deep learning, there are more and more models for sales forecasting, and the performance of single models is often not ideal. Therefore, there are more and more combinatorial models. In this study, we use Stacking strategy to support XGBoost, Support Vector Regression (SVR), GRU neural network as the basic model, then lightGBM as the final prediction model, with new features are merged. The advantages of several models are condensed, which greatly improves the prediction performance of the model, good enough to be more close to the real sales data, and provide a new prediction method for regression prediction.