A computational economics framework for rare-event prediction: evidence from cryptocurrency price shocks using smote and ensemble learning

This study develops a computational economics framework for modeling and predicting rare financial events, using cryptocurrency price shocks as an illustrative case. We treat shock detection as a binary classification problem, where extreme returns are defined by the 90th percentile of absolute log returns. To address the rarity of such events, we integrate the Synthetic Minority Over-sampling Technique (SMOTE) with ensemble learning methods (Random Forest, XGBoost, Light GBM) and benchmark against logistic regression. Using daily Bitcoin data from August 2024 to August 2025, our results show that ensemble models—especially Light GBM—achieve strong predictive performance (AUC = 0.919) and substantially higher recall for shock detection. Feature importance analysis highlights the predictive role of short-term volatility memory and liquidity variation, consistent with theoretical constructs such as EVT, EMH deviations, and the Mixture of Distributions Hypothesis. Beyond crypto currencies, the proposed framework provides a generalizable computational approach to rare-event forecasting in economics, with direct applications to crises, systemic liquidity shocks, credit defaults, and exchange rate collapses. By combining imbalance correction with ensemble learners, this study contributes methodologically to the computational economics literature on tail-risk prediction and systemic stability.