A Comparative Study of Demand Forecasting Models for a Multi-Channel Retail Company: A Novel Hybrid Machine Learning Approach

Abstract

Demand forecasting has been a major concern of operational strategy to manage the inventory and optimize the customer satisfaction level. The researchers have proposed many conventional and advanced forecasting techniques, but no one leads to complete accuracy. Forecasting is equally important in manufacturing as well as retail companies. In this study, the performances of five regression techniques of machine learning, viz. random forest (RF), extreme gradient boosting (XGBoost), gradient boosting, adaptive boosting (AdaBoost), and artificial neural network (ANN) algorithms, are compared with a proposed hybrid (RF-XGBoost-LR) model for sales forecasting of a retail chain considering the various parameters of forecasting accuracy. The weekly sales data of a US-based retail company is considered in the analysis of the forecasts undertaking the attributes affecting the sale such as the temperature of the region and the size of the store. It is observed that the hybrid RF-XGBoost-LR outperformed the other models measured against various metrics of performance. This study may help the industry decision-maker to understand and improve the methods of forecasting.