LONG-TERM PREDICTION OF DAILY TIME SERIES USING MODELS BASED ON REGRESSION TREES

This work aims to evaluate the application of regression models based on decision trees, such as Random Forests and XGBoost, in the long-term prediction of daily time series. Classic approaches such as Seasonal Naïve, SARIMA and ETS were compared with machine learning methods, using real sales data from an e-commerce. The study included exploratory analysis, engineering and selection of characteristics, and optimization of hyperparameters. The results showed that tree-based models performed better in terms of root mean squared error (REQM), outperforming classical statistical models, especially in series with complex seasonalities. The analysis reinforces the importance of using well-constructed explanatory variables to improve long-term forecasts of time series.