Detecting and Explaining Unlawful Insider Trading: A Shapley Value and Causal Forest Approach to Identifying Key Drivers and Causal Relationships

Battle Count: 66

The paper is primarily focused on regulatory detection and causal explanation of insider trading rather than direct trading strategy development. However, it has indirect relevance: (1) understanding information asymmetry and market microstructure factors (Market Beta, Price-to-Book, Return) informs alpha generation; (2) identifying features that predict informed trading can help quants avoid adverse selection; (3) the causal analysis of factors like director status and valuation metrics connects to factor investing literature; (4) the methodology (SHAP + Causal Forest) is transferable to other financial ML applications. The paper does not propose trading signals or portfolio construction directly.

The pipeline involves multiple stages: XGBoost training with 5-fold CV and early stopping, SHAP computation on test data, hierarchical clustering with Spearman correlation, iterative VIF filtering, and Causal Forest with AIPW estimation. Requires expertise in both ML (XGBoost, SHAP) and causal inference (CATE, propensity scores, honest trees). The econml library simplifies CF implementation, but proper interpretation of results, handling of multicollinearity, and understanding of causal assumptions add complexity. Data merging across SEC EDGAR, CRSP, and Compustat with Levenshtein matching adds engineering overhead.

The paper uses publicly available data (SEC EDGAR Form 4, CRSP, Compustat-CapitalIQ) and standard open-source libraries (scikit-learn, xgboost, econml). However, the exact preprocessing pipeline, Levenshtein matching threshold (85%), and specific hyperparameters for Causal Forest (1000 trees, max depth 10, honest fraction 0.8:0.2) are described but no code repository is explicitly linked. The paper is a preprint under review, and some methodology details reference prior work (Neupane and Griva 2024b) for replication.