Insider Purchase Signals in Microcap Equities: Gradient Boosting Detection of Abnormal Returns

Battle Count: 81

Highly relevant to quantitative trading, particularly for microcap equity strategies. The paper provides a concrete, backtested signal (Form 4 insider purchases) with quantified predictive power (AUC 0.70, mean CAR 6.3% in the best momentum bucket). The counter-intuitive momentum finding (purchases disclosed into price strength outperform those into weakness) directly challenges common mean-reversion heuristics used in trading systems. The feature importance analysis identifies actionable variables (distance from 52-week high, price deviation, volatility). However, practical relevance is tempered by microcap liquidity constraints, limited capacity, and the need for realistic execution modeling. The signal is statistically robust but economically constrained.

Moderate complexity. The core model (XGBoost binary classifier) is straightforward to implement using standard libraries. However, the full pipeline involves several non-trivial steps: parsing SEC EDGAR Form 4 filings, CUSIP-to-permanent-identifier mapping for ticker changes and mergers, computing Fama-French three-factor abnormal returns with 252-day rolling loadings, constructing point-in-time market condition features, handling class imbalance via threshold optimization, and implementing temporal train/validation/test splits. The data engineering (especially SEC filing parsing and market data alignment) represents the bulk of the implementation effort. The model itself is well-understood and widely available.

The paper describes a clear methodology with specific data sources (SEC EDGAR for Form 4 filings, institutional-grade market data vendor for prices/volume), explicit feature construction, temporal train/validation/test splits, and defined hyperparameter tuning via time-series cross-validation. However, the market data vendor is not named, code is only 'available upon request' rather than publicly hosted, and the exact XGBoost hyperparameters are not fully enumerated. The Fama-French factor data source is also unspecified. These gaps reduce reproducibility despite the otherwise transparent pipeline description.