Forecasting critical mineral prices with machine learning and deep learning: Evidence from the Adaptive Market Hypothesis

Critical minerals such as lithium, copper, nickel, cobalt, manganese, and rare earths are central to the global energy transition, yet their prices are highly volatile and difficult to forecast. This study evaluates whether a suite of classical econometric, machine learning, and deep learning models can improve the predictability of monthly critical mineral prices. Using an expanding-window framework with models initially trained on pre-2022 data and evaluated out-of-sample over 2022–January 2026, model performance is assessed using Diebold–Mariano, Clark–West, and Pesaran–Timmermann tests. Forecastability varies significantly across minerals and over time, consistent with the Adaptive Market Hypothesis. The Temporal Fusion Transformer outperforms the Random Walk for nickel and rare earths, XGBoost is the best model for lithium, ARIMA works best for cobalt, N-HiTS leads for copper, while manganese prices are effectively unforecastable, with no model beating the Random Walk by a statistically robust margin. Directional accuracy is more consistently significant than point forecast accuracy across models and minerals. These improvements correspond to economically meaningful reductions in forecast error of up to 8% in certain markets—worth up to approximately $36 million per year under national-scale procurement scenarios—and a simple model-timed procurement rule lowers average purchase prices by 0.3–3.3% across minerals. These results indicate that predictability in critical mineral markets is episodic and regime-dependent rather than persistent. From a policy perspective, even modest gains in forecast accuracy can enhance procurement timing, risk management, and strategic stockpiling decisions under frameworks such as the EU Critical Raw Materials Act and the U.S. Energy Act of 2020.