A Hybrid Model for Demand Forecasting Based on the Combination of Statistical and Machine Learning Methods

Abstract

Demand Forecasting (DF) is nowadays a key component of successful businesses in retailing field. In fact, accurate customer’s demand forecasts and insights into the reasons driving the forecasts may increase confidence, assist decision-making and therefore boost’s the retailer’s profit. It is then crucial for an accurate DF model to not only understand the retail time-series repeated patterns but also the impacts of factors such as the promotions on the data behavior. The literature review of existing research works has shown that statistical models gave good results in accurately detecting time-series components such as seasonality or trend but they fail when it comes to detecting external factors or causal effects compared to machine learning models. Moreover, the combination of both models either focused only on the trend component and neglected the seasonality or considered both of them but used sophisticated neural networks, which are computationally expensive. To this end, in this paper, we propose an approach that combines statistical and machine learning models to take advantages of their aforementioned properties. We used first Multiple Linear Regression (served as the baseline model as well) and linear interpolation to remove the promotions effect from the data and compute promotional multipliers. Then, each resulting data was fed to two statistical models (Prophet and Exponential Triple Smoothing). Finally, the combination step consisted in reintegrating the promotions effect into the forecasting results of each statistical model. Quantitative and qualitative evaluations of the hybrid models’ performance showed that the hybrid models outperformed the baseline model.