Leakage-Aware Post-Event Classification of Strong Earthquakes in a Global USGS Feed Snapshot

This publication, titled “Leakage-Aware Post-Event Classification of Strong Earthquakes in a Global USGS Feed Snapshot” presents a reproducible machine-learning study using 10,464 cleaned earthquake records obtained from the United States Geological Survey (USGS). It compares logistic regression, random forest, extremely randomized trees, and histogram-based gradient boosting for classifying already-recorded strong earthquake events. To improve methodological reliability, earthquake magnitude and USGS significance scores were excluded from the input features to prevent target leakage. Logistic regression achieved the strongest overall performance, recording a ROC-AUC of 0.986, 95.2% accuracy, and 99.4% recall. The study demonstrates how transparent, leakage-aware machine-learning methods can support earthquake-data analysis, scientific communication, and reproducible research. It focuses strictly on post-event classification and does not claim to predict future earthquakes.