Inteligência artificial para previsão de séries temporais com modelos de base.

Abstract

This study aims to evaluate the efficiency of Foundation Models (FMs) in time series forecasting applied to the revenue of the thermoplastics sector in the Manaus Free Trade Zone, comparing them with traditional statistical methods such as Moving Average and Exponential Smoothing. A quantitative and experimental approach was employed, using public data from SUFRAMA covering the period from 2020 to 2025. To enhance the robustness of the models, Monte Carlo Simulation was applied to expand the historical series. The analyzed models included Amazon Chronos-T5 and Google TimesFM, both based on the Transformer architecture. The performance metrics adopted were MSE, MAE, and MAPE. The results demonstrated that the foundation models outperformed classical methods in terms of accuracy and stability, with TimesFM being the most efficient, followed by Chronos-T5. It is concluded that the use of FMs significantly improves forecasting accuracy, providing more reliable support for decision-making and industrial planning in the Amazon region.