A Comparative Study of Univariate Time-series Methods for Sales Forecasting

Abstract

Sales time-series forecasters, data scientists and managers often use timeseries forecasting methods to predict sales. Nonetheless, it is still a question which time-series method a forecaster is best off using, if they only have time to generate one forecast. This study investigates and evaluates different sales time-series forecasting methods: multiplicative Holt-Winters (HW), additive HW, Seasonal Auto Regressive Integrated Moving Average (SARIMA) (A variant of Auto Regressive Integrated Moving Average (ARIMA)), Long Short-Term Memory Recurrent Neural Networks (LSTM) and the Prophet method by Facebook on thirty-two univariate sales time-series. The data used to forecast sales is taken from time-series Data Library (TSDL). With respect to the Root Mean Square Error (RMSE) evaluation metric, we find that forecasting sales with the SARIMA method offers the best performance, on average, relative to the other compared methods. To support the findings, both mathematical and economic reasoning on the drivers of the observed performance for each method are provided. However, a decision maker or an organization need to evaluate the trade-off between forecasting accuracy and the shortcomings associated with each method.