판례 기반 토픽모델링 및 머신러닝을 활용한 가상자산 불공정거래 유형 분석 및 예측 연구

The rapid expansion of the cryptocurrency market has transformed financial systems but has also introduced new forms of misconduct that threaten market integrity. This study proposes a quantitative framework for detecting and predicting unfair trading practices in the crypto-asset market by combining topic modeling and machine learning using U.S. court case texts. Using BERTopic, eight key types of unfair trading were identified: price manipulation, insider trading, false disclosure, unauthorized advisory, money laundering, platform liability, regulatory evasion, and general disputes. Network analysis revealed that price manipulation, insider trading, and false disclosure form central clusters of precedent, while regulatory evasion and unauthorized advisory remain isolated, reflecting weak legal formalization. Machine learning models were trained to classify unfair trading cases, with XGBoost achieving the best performance (accuracy 0.90, F1-score 0.68) and Random Forest following closely, aided by SMOTE to address data imbalance. The classification model showed high prediction accuracy in detecting price manipulation and insider trading, but its performance was inconsistent in platform responsibility and regulatory evasion types, which reflect the heterogeneity of these types and the low standardization level of norms. These research results suggest that machine learning models can play an important role in early detection of unfair trading in the virtual asset market, strengthening risk assessment, and supporting accounting procedures such as impairment loss testing and contingent liability estimation.