APPLICATION OF MACHINE LEARNING ALGORITHMS FOR DAILY SALES FORECASTING IN A BRAZILIAN BREWING INDUSTRY

Abstract

Sales forecasting is a strategic element for the brewing industry, a sector characterized by high seasonality and significant fluctuations associated with weekends, holidays, and festive events. Traditional statistical models, although widely applied, present limitations in capturing nonlinear relationships and complex patterns inherent to demand time series. In this context, this study aims to evaluate the application of machine learning algorithms for daily sales forecasting in a Brazilian brewing industry. To this end, a historical dataset of daily sales covering the period from May 2024 to May 2025 was used, adopting a temporal split in which data from 2024 were employed for model training and data from 2025 were reserved for testing. Temporal feature engineering was performed, including calendar variables, holiday indicators, lagged variables, and moving statistics. Four tree-based algorithms were evaluated: Decision Tree, Random Forest, XGBoost, and LightGBM, using the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE) as performance metrics. The results indicate that ensemble models significantly outperform the single decision tree, with XGBoost achieving the best predictive performance, explaining approximately 89.9% of the variance in daily sales. It is concluded that the application of machine learning algorithms combined with temporal feature engineering constitutes an effective approach for demand forecasting in the brewing sector, providing relevant support for production, inventory, and logistics planning.

DOI: https://doi.org/10.56238/sevened2026.008-124