Predicting Rare Earth Supply Shocks Using a Machine Learning Classifier Built from Open Geopolitical Data

Abstract

China is responsible for some 90% of the global processing capacity for rare earth elements (REEs) and can therefore use its export quotas as a geopolitical tool (International Energy Agency, 2024; OECD, 2023). These bans have an impact on markets, as the gaps in supply and demand can be measured, and news of these restrictions is made available in Chinese but not released in English until 48 to 72 hours later on official Chinese outlets (Kuo, 2021). In this paper, I explore the possibility of a machine learning classifier being able to anticipate a REE supply shortage—an event that has yet to be obvious from the market—in all cases for which I can find publicly available data. Three open-source datasets were combined, the GDELT global event database (Leetaru & Schrodt, 2013) with Chinese state media coverage indicators and daily stock prices for two Indian downstream companies and the REMX (rare earth) ETF for 2020–2025 to create 1,556 daily observations. The 10 features were designed and trained in a logistic regression classifier using a strict time-ordered train/test split without data leakage. If applied to unseen data on 2025 the model's recall was 0.87 and its ROC-AUC was 0.732, and the first alert was made an average of 28.5 days before the two real restriction events of 2025, when the REMX ETF was flat or declined. The largest single indicator was the quantity of Chinese state media coverage, not the sentiment. The model triggers about one false alarm every two trading days, showing some limitations in practice. However, the findings show that freely available geopolitical signals are actually predictive in a meaningful way of rare earth price shocks, with lead time.