[1K5-ES-2] Machine learning: Social application (2)

[1K5-ES-2-02] Forecasting Item-level Retail Sales Demand Using Combination Model

〇YINGSHA YANG¹, KAZUHIRO KOIKE¹, YASUYUKI MITSUI¹ (1. ASKUL)

Keywords:Retail sales, Combined forecasts, Machine learning, Classical time series forecasting methods

The accuracy of sales demand forecast at item-level is a big issue for inventory management in retail industry, especially in e-commerce domain where forecasting becomes more difficult because of vast array of items, frequent promotions and unexpected event. In this paper, we propose a combination forecast model which integrates some cutting-edge forecast methods, such as decomposition model like Prophet and machine learning methods like XGBoost and CatBoost as well as some classical statistical methods. In order to handle massive data efficiently, we use a new light-weight weighting method to combine single models. We test the model based on real sales data set on long time span, and results show the combination method is superior to any single method in both stability and accuracy.