Comparing classical time-series models and machine learning for demand forecast on the beverage industry in COVID-19 pandemic

Due to the growth of competitiveness in the market, demand forecasting has become a fundamental tool to manage production and identify new opportunities for the company. The fundamental goal of a series analysis is to make predictions from historical data to support decisions accurately. During the COVID-19 pandemic, the market has undergone numerous changes, and consumer needs have changed, directly affecting beverage sales. In this work, classic models of time series - Holt-Winters and ARIMA - and machine learning - support vector machines and random forests - were used to perform demand forecasts from several historical data series of a real beverage direct distribution centre located in Brazil. The data used were stratified into nine data series: 1) the total volume of beverages sold by the operation; 2) separated by type of beverage (beer and non-alcoholic beverage); 3) in six sales channels. Indeed, as the comparison considers demands before and after the pandemic (including pre-and post-vaccination), the predictions were challenging. The comparison of models considers predictions up to 15 steps (months) ahead using the RMSE and MAPE error metrics. Here, the models with the best-aggregated performances were ARIMA and SVM; however, no model was strictly better than the others.