Forecasting Demand Using ARIMA Model and LSTM Neural Network: a Case of Detergent Manufacturing Industry

Abstract

Generating reliable and meaningful product demand predictions is an open challenge in the industrial environment. Demand forecasting is still an active avenue of research since it significantly affects business profitability because of uncertainties related to demand predictability, high product variety, and supply fluctuation. This paper deals with a practical real-life case study of a leading international company. Particularly, we investigate the demand forecasting for the industrial production of household detergents. To tackle this challenging problem over medium long-term prediction horizons, we propose two different techniques: (i) the traditional statistical approach namely the AutoRegressive Integrated Moving Average model (ARIMA), and (ii) the artificial neural networks based on Long Short Term Memory algorithm (LSTM). We empirically assess and compare these approaches to real data sets. Numerical experiments attest to the competitiveness of the results obtained for household detergents and cleaning products. Furthermore, the results reveal that deep learning models have a better overall performance than traditional statistical techniques. Precisely, the average percentage errors provided by the LSTM algorithm is 22% when compared to ARIMA (34%) showing better forecasting accuracy of the LSTM prediction model.