Developing Predictive and Explainable Models for Cryptocurrency Delistings: A Case Study of Binance Exchange

초록

This study develops an explainable machine learning model to predict cryptocurrency delistings using Binance data. It combines quantitative indicators (price, volume) with qualitative data from real-time news and Reddit. Latent Dirichlet Allocation (LDA) is used to extract topic trends and community reactions, which are transformed into time-series features. XGBoost, LightGBM, and CatBoost are compared, with SHAP applied for model interpretability. Results show that sharp price drops, repeated risk-topic exposure, and Reddit responses strongly predict delisting. XGBoost achieves the best performance, offering practical insights for early warning systems and investor protection.

키워드