The Information Dynamics of Insider Intent: How Reporting Inversions (Form 144) Mask Informational Rents in Insider Sales (Form 4)

Battle Count: 50

Highly relevant for quantitative trading in several dimensions: (1) The 1.3% pricing differential between Form 144 intent and Form 4 execution provides a measurable alpha signal for event-driven strategies. (2) The 52.4% opacity rate and 90-day non-execution event create a predictable negative drift (H1) and subsequent reversal (H2) exploitable by systematic strategies. (3) The large-cap significance paradox (14.49 bps, p=0.021) suggests a tradeable signal concentrated in liquid, institutionally-covered stocks. (4) The illiquidity jump of up to 2.63x at high signal magnitudes informs execution strategy and market impact modeling. (5) The cross-sectional finding that prior idiosyncratic volatility amplifies the signal (p=0.021) provides a conditional trading rule. However, the absolute magnitudes (14-32 bps) are modest, and the 90-day holding period introduces significant execution and risk management challenges.

High implementation complexity due to: (1) Multi-database linkage requiring LSEG, BoardEx, CRSP, Compustat, and Fama-French data with a three-stage identifier resolution protocol (personid → directorid → companyid → PERMCO). (2) The three-phase analytical pipeline spanning univariate OLS, Carhart four-factor event studies, 10+ ML architectures with SMOTE augmentation and cost-sensitive learning, and causal ML ensembles (DML, GRF, X-Learner). (3) DGTW-adjusted Calendar-Time Portfolio construction with 125-cell characteristic grids. (4) Handling extreme class imbalance (3.08% minority class in 2.6M observations). (5) The causal inference framework requires satisfying unconfoundedness through high-dimensional controls and honest inference. (6) Proprietary data access (WRDS) is a significant barrier. However, individual components (event studies, XGBoost, GRF) are well-documented in open-source libraries.

The paper relies on proprietary databases (LSEG Insider via WRDS, BoardEx, CRSP, Compustat, Fama-French factor database) that require institutional subscriptions. While the methodology is described in detail with equations, the exact data construction pipeline, matching algorithms, and hyperparameter tuning are not fully specified. No code repository is explicitly provided. The author's website (krishpn.github.io) is referenced but no specific GitHub repo with code is linked. The paper is a preprint under review, suggesting potential revisions.