Multi-step Sales Forecasting of Retail Merchandise Based on EEMD-HW-GBDT Model
SKU-level time series of retail merchandise sales have strong non-stationarity and nonlinearity,so the multi-step sales forecast is difficult.To solve the above problems,a novel sales forecast model based on ensemble empirical mode decomposition(EEMD),Holt-Winters(HW)and gradient boosting decision tree(GBDT)was proposed.The model was divided into three stages.Region-all SKU level sales were predicted by the HW model in the first stage,and.the prediction results were allocated to SKU-level through the ratio-to-moving-average method.Then EEMD was used to process the original time series and the HW predicted value series to reduce the non-stationarity of the data and expand the input features.The final stage was to integrate multiple GBDTs based on HW prediction components,and internal and external features to obtain prediction results.The sales data from a domestic retail company were used for verification.The results show that the EEMD-HW-GBDT model has good predictive performance for multi-step sales forecast of retail merchandise.It is better than seven benchmark forecating models in terms of MAE,RMSE and WMAPE.
万方数据
维普
