An extreme Gradient Boosting (XGBoost) Trees approach to Detect and Identify Unlawful Insider Trading (UIT) Transactions

Battle Count: 74

The paper is primarily focused on regulatory detection of unlawful insider trading rather than developing trading strategies. However, it is relevant to quantitative trading in several ways: (1) understanding insider trading patterns informs market microstructure research, (2) the features identified (Market β, Return, Price Operating Earnings, IsDirector) are relevant to factor models and alpha generation, (3) the methodology could be adapted for detecting market manipulation in trading systems, and (4) regulatory compliance is critical for institutional trading operations. The connection is more indirect than papers focused on prediction or portfolio construction.

XGBoost itself is well-documented and readily available via scikit-learn and xgboost libraries. However, the full pipeline adds complexity: (1) data merging from SEC Form 4, CRSP, and Compustat-CapitalIQ requires significant data engineering, (2) PCA integration for comparative analysis, (3) hierarchical clustering with Spearman rank correlation for feature decorrelation, (4) permutation importance computation, (5) 100 repetitions with random sampling, and (6) 5-fold cross-validation with hyperparameter tuning. The core model is straightforward but the complete analytical framework requires moderate-to-advanced ML expertise.

The paper uses publicly available libraries (scikit-learn, xgboost) and describes experimental settings in detail (100 repetitions, 5-fold CV, z-score normalization, one-hot encoding). However, the dataset construction involves linking SEC Form 4 filings with CRSP and Compustat-CapitalIQ data via personid, cik, and companyid, which requires subscription access. No code repository is explicitly mentioned. The random sampling of lawful transactions from a pool of 9.6 million introduces some variability.