Sales forecasting for a supermarket chain in Natal, Brazil: an empirical assessment

Abstract

Time series forecasting is a consolidated, broadly used approach in several fields, such as financing and industry. Retail can also benefit from forecasting in many areas, such as price and sales optimization and stock demand. This study addresses retail sales forecasting in Nordestão, a large Brazilian supermarket chain. Though located in a state with a low gross domestic product, Nordestão ranks 3rd and 27th, respectively, in regional and national sales. The data considered here spans five years of daily transactions from eight different stores. Different machine learning techniques, knowingly effective for forecasting, are adopted, namely random forests and XGBoost. We further improve their performance with feature engineering to address seasonal effects. The best algorithm varies per store, but for most stores at least one of the methods is proven to be effective. Feature engineering had a great impact on the modeling, showing that the empirical analysis was largely responsible for achieving high scores, reaching score in range of 90%. Besides the traditional relevance of sales forecasting, our work is a means for Nordestão to evaluate the impact of the COVID19 pandemic on sales and improve on other operational tasks such as stock planning and distribution.