Previsão de vendas de alta frequência com modelos de séries temporais: o caso de uma rede de supermercados

Abstract

Recent research on sales forecasting at the individual product level is done with weekly granularity for supermarkets (FILDES et al., 2019). The main objective of the present research is to compare daily sales forecasting models with weekly ones in the context of supermarket retailing. Secondarily, it is intended to discuss static and dynamic models and present the impact of the forecast horizon and training frequency on the performance of forecast models. In this research, a database of one store of a Supermarket Chain in Espírito Santo was used. We chose time series models (HoltWinters and ARIMA) and TBATS because it considers multiple seasonalities in the high frequency (daily) forecast to perform the sales forecast. As a result, it is noted that the daily forecast provides better predictive performance than the weekly forecast for the context in question. In addition, the dynamic models provided more accurate forecasts than the static models. And when comparing the performance of the HoltWinters, ARIMA, and TBATS models, the latter showed greater assertiveness, which can be explained by the presence of complex seasonality in supermarket retail sales data.