Análise comparativa de modelos de previsão de demanda aplicados a múltiplas bases e níveis de granularidade no contexto do varejo

Abstract

The retail sector is highly dynamic and competitive, requiring effective strategies for inventory management, pricing, and logistics. Demand forecasting becomes an essential tool to optimize resources, reduce operational costs, and ensure product availability, minimizing losses and improving the consumer experience. Furthermore, in a market where customer preferences change rapidly and seasonal factors directly influence sales, accurately forecasting demand allows companies to anticipate market fluctuations and adjust their operations quickly and efficiently. This paper presents a comparative analysis of demand forecasting models applied to multiple retail databases, considering different levels of granularity, such as total sales, by store, and by product. Four widely adopted quantitative algorithms were used: ARIMA, Prophet, LSTM, and Convolutional Neural Networks (CNNs), evaluating their predictive performance on recognized databases, such as Favorita, M5, and Olist. Evaluation metrics considered include RMSE and MAPE, in addition to training and prediction time. The results highlighted the superiority of the LSTM model, which obtained an average RMSE 7\% better than the second best model. In the M5 database, LSTM stood out with an average RMSE of 1013, approximately 43\% better than ARIMA, which was the second best model in most general metrics. CNNs showed better performance in the Olist database, while Prophet showed limitations in capturing patterns in more complex time series. The study concludes that advanced techniques, such as neural networks, are essential to deal with the complexity of data in retail, providing greater accuracy and efficiency in demand forecasting, aiding in strategic decision-making for the sector.