This project aims at applying Machine Learning (ML) methods in order to obtain forecasting gains in comparison to traditional models. The Machine Learning methods work with a high-dimension of variables even when there are more variables than historical data available, which is not possible for linear models. A big number of models were tested with metal commodities, such as the iron ore, steel and copper, in a monthly forecasting horizon up to two years. Motivated by that, we gathered a set of time series with a considerable amount of observations, including financial instruments, exchange rates and macroeconomic indicators, which were classified and displayed in tables. The methodologies used in each model are also explained in this work. Finally, we discuss the performance of each model in different forecasting horizons for the copper spot price in the London Metal Exchange (LME). As we will see, the LASSO and the Elastic Net models, in comparison to the others, presented the best performance among different forecasting horizons. Even further, the LASSO obtained a significant gain in comparison to the benchmark models Random Walk and Autoregressive, as well as the more traditional models for forecasting heteroskedastic financial time series, ARCH and GARCH. Its method of variable selection proved to be robust in different windows for each forecasting horizon, presenting smaller and more stable performance metrics, such as the RMSE, MAE and MAD along the forecasting windows.