Uso de séries temporais e aprendizado de máquina na predição de vendas diárias de produtos em supermercados de médio porte do paraná

Abstract

This study investigates the use of Machine Learning techniques to predict the daily sales volume in medium-sized supermarkets, comparing traditional and deep learning approaches. The objectives were: (i) to evaluate the performance of different algorithms on datasets of different scales (250 and 2,000 products), (ii) to analyze the impact of atypical events, such as the COVID-19 pandemic, on predictive performance, and (iii) to build and provide an unprecedented experimental database for future research. The methodology comprised two case studies using real data from two supermarkets in southwestern Parana, covering the period from 2019 to 2024. In the first case study, with 250 products, Recurrent Neural Networks outperformed linear regression, although all models showed performance degradation during the pandemic. In the second case study, with 2,000 products, XGBoost achieved the best results in most instances, while LSTM showed superior performance for high-turnover products. The findings indicate that no single algorithm is universally superior; the choice depends on product profile and dataset size. As a main contribution, in addition to the comparative analysis, this work provides an experimental database containing six years of daily sales records for 2,000 products in medium-sized supermarkets, thereby expanding the literature on retail demand forecasting.