Desenvolvimento de modelos de previsão de preço do minério de ferro

The main objective of this work was to develop and comparatively evaluate forecasting models for the price of iron ore, a key raw material for various industries such as civil construction and automotive. Accurate forecasts in this scenario can bring economic stability and health to companies in the sector. To this end, traditional statistical approaches (ARIMA), machine learning models (XGBoost), and hybrid strategies (Ensemble and Residuals Method) were implemented. The analysis also investigated the impact of incorporating exogenous variables on the models’ performance. The results demonstrated the superiority of the Ensemble model without exogenous variables, which combined the predictions of ARIMA and XGBoost, achieving the lowest Root Mean Square Error (RMSE of 6.97) and the highest coefficient of determination (R2 of 0.67). This performance is notable as it was the only one to outperform the baseline model (RMSE of 7.05 and R2 of 0.65), which in turn surpassed the other complex approaches, highlighting the intrinsic difficulty of the forecast. It was found that the addition of exogenous variables was, in general, detrimental to performance, especially in linear models. It is concluded that combining models of different natures is a robust strategy for forecasting complex time series, and that the proper handling of external variables is a critical factor for successful forecasting