Machine-Learning Model Flags China Rare Earth Supply Risk Weeks Before 2025 Export Controls

Highlights

• Student researcher Ayaan Syed Imam built a logistic-regression classifier using GDELT, Chinese state-media, and market data across 1,556 daily observations from 2020–2025.
• The model achieved 87% recall and 0.732 ROC-AUC on held-out 2025 data, with alerts preceding two Chinese restriction events by an average of 28.5 days.
• Chinese state-media volume outweighed sentiment as a predictive signal, suggesting policy formation leaves a detectable information trail before official announcements.
• A false-alarm rate of roughly one every two trading days and only two confirmed 2025 test events limit the model's stand-alone reliability for investors or procurement officers.
• REEx analysts argue a robust early-warning system must integrate Chinese-language policy signals, export licensing data, customs flows, inventory levels, and physical oxide prices.