Retail sales forecasting for a Brazilian supermarket chain: an empirical assessment

Abstract

Time series forecasting is a consolidated, broadly used approach in several fields, such as finance and industry. Retail can also benefit from forecasting in many areas such as stock demand, price optimization, and sales. This study addresses retail sales forecasting in Nordestão, a large Brazilian supermarket chain that respectively ranks 3rd and 27th in sales regionally and nationally. The data considered spans five years of daily transactions from eight different stores. Knowingly effective machine learning techniques for forecasting are adopted, namely linear regression, random forests, and XGBoost. We further improve their performance with features we engineer to address seasonal effects. The best algorithm varies per store, but for most stores at least one of the methods proves effective. Importantly, the models display effective performance across multiple testing weeks, and improve over the current approach of Nordestão by a significant margin. Besides the traditional relevance of sales forecasting, our work is a means for Nordestão to evaluate the impact of the COVID-19 pandemics on sales.