0.87
recall on unseen 2025 data
Published · Grade 12
Predicting Rare Earth Supply Shocks Using a Machine Learning Classifier Built from Open Geopolitical Data
A classifier built only from free data that flagged both 2025 Chinese rare-earth restrictions about four weeks before markets moved.
Enterprise Development and Microfinance, Vol. 36, No. 3 (ABDC-indexed). Mentor: Dr Loonkar, visiting faculty, NMIMS.
- recall on unseen 2025 data
- 0.87
- recall on unseen 2025 data
- ROC-AUC
- 0.732
- ROC-AUC
- average lead time before the two real 2025 restrictions
- 28.5 days
- average lead time before the two real 2025 restrictions
- false alarms per trading day at the chosen threshold, reported honestly
- 1 in 2
- false alarms per trading day at the chosen threshold, reported honestly
01 Why I did it
The question.
The surprising result was that the volume of state-media coverage predicted the shock better than its sentiment. A government talks more before it acts. I liked that the whole thing runs on free data: an Indian electronics firm could run it on a laptop.
02 Abstract
What the paper says.
China controls roughly 90 percent of rare-earth processing and uses export restrictions as a geopolitical instrument. Official announcements appear in Mandarin on state channels 48 to 72 hours before English-language outlets report them. Three open datasets, the GDELT event database, Chinese state-media coverage indicators and daily prices for the REMX rare-earth ETF plus two Indian downstream equities, were merged into 1,556 daily observations from 2020 to 2025. Ten features were engineered and a logistic regression classifier trained under a strict time-ordered split with no leakage.
03 Method and results
How, and what came out.
- GDELT global events, Chinese state-media coverage volume and sentiment, REMX ETF and two Indian electronics stocks
- Ten engineered features, logistic regression, chronological train and test split
- Evaluated on unseen 2025 data by recall and ROC-AUC
- Spillover traced to Indian downstream equities
0.732
ROC-AUC
28.5 days
average lead time before the two real 2025 restrictions
1 in 2
false alarms per trading day at the chosen threshold, reported honestly