A hybrid machine learning model for sales prediction

Abstract

Accurate sales forecasting models can help retailers develop more appropriate business plans. This paper is based on the LightGBM framework and the XGBoost framework to build a sales forecast model. First, two models were built separately based on these two frameworks. Then we assigned weights based on the prediction results of these two models and performed model integration. The integrated model has the characteristics of the two models at the same time, and shows better predictive ability. Before training the model, a large amount of data needs to be preprocessed first, so feature engineering is required in this article. First, we delete some functions that are not related to model input. Then the features are extracted and classified to obtain the mean, standard deviation and other statistics of some features. Experimental results show that the RMSE of this method is 2.07, which is significantly better than the two models before the integration. The RMSE of the model based on LightGBM is 2.09, and the RMSE of the model based on the xgboost framework is 2.11.