Abstract Given the importance of base metals prices for mining companies and producing countries, price forecasting tools are essential for sound decision-making and risk management. This study empirically investigates the performance of two newly proposed foundation models, namely MOIRAI and Chronos, in forecasting base metals prices by comparing their performance to the performance of traditional econometric (i.e., Naïve and ARIMA), and machine learning models (i.e., feed-forward neural networks). It also explores whether the combination of forecasts of the aforementioned models with the ones obtained from the univariate models can improve the forecast accuracy. Using a monthly dataset of aluminum, copper, lead, tin, nickel, and zinc prices from 1990 to 2025, we evaluate the six forecasting models across short (3-month), medium (6-month), and long (12-month) horizons. We find little to no evidence that the two foundation models perform better than the classical univariate models, as none of the models generated competitive forecasts for any commodity and horizon under study, apart from the long-term forecasts of zinc prices, for which Chronos showed the lowest out-of-sample error, based on all the evaluation metrics we have used. However, we found evidence that combining the forecasts of these models with those generated by classical univariate models can enhance the final accuracy, as in many cases the ensemble outperformed the other individual models. These outcomes challenge the perspective that the recently developed foundation models perform better than traditional econometric techniques. At the moment, it seems that foundation models need further development, therefore corporations, policy makers and investors should use them in a complementary rather than substitutive way, integrating them with established forecasting approaches rather than relying solely on their zero-shot capabilities.
Oikonomou et al. (Tue,) studied this question.