DEMAND FORECASTING FOR LAST-MILE LOGISTICS IN THE BRAZILIAN POSTAL SECTOR: A COMPARATIVE ANALYSIS ORIENTED TO FLEET SIZING

Accurate demand forecasting is a central requirement for last-mile logistics planning, especially in the allocation of vehicles, teams, and resources. In this work, four demand forecasting models are comparatively evaluated: SARIMA, Holt-Winters, LightGBM and XGBoost, applied to real time series of a company in the Brazilian postal sector, covering four operational units. The models were trained with daily data from January 2020 to January 2024 and evaluated outside the training sample, in the period from February 2024 to January 2025. The results indicate superior performance of the \textit{gradient boosting} algorithms. XGBoost presented MAPE between 1.81\% and 3.68\%, while LightGBM obtained values between 1.82\% and 4.12\%. SARIMA and Holt-Winters showed greater variability, with MAPE between 6.54% and 21.70%. XGBoost was selected as a reference model for application and for future studies of stochastic fleet sizing.