A systematic equity fund wants to build a signal from SEC Form 4 insider transaction filings. Their data science team is debating whether to use raw dollar transaction amounts or a normalized metric. Which approach to signal construction produces the most robust predictive alpha, and what confound must be controlled for?

The most robust approach is to normalize insider purchases as a percentage of total insider holdings, often called the 'conviction ratio', cluster transactions by type, especially open-market purchases versus pre-scheduled 10b5-1 plan trades, and control for option expiration-driven sales so the signal isolates discretionary buying. Normalizing by existing holdings adjusts for the wealth effect: a $500,000 purchase means far more when it represents 15% of an insider's net worth in company stock than when it represents 0.1%. Academic research by Lakonishok and Lee (2001) and Jeng, Metrick, and Zeckhauser (2003) consistently finds that the most predictive insider transactions are those where the purchase represents a meaningful fraction of existing holdings. The 10b5-1 plan distinction is equally important. Pre-scheduled trading plans allow insiders to trade on a fixed schedule established months in advance during an open window, mechanically executing regardless of subsequent developments. These trades carry no informational content about current business conditions. In contrast, open-market discretionary purchases in the absence of an existing 10b5-1 plan represent genuine expressions of insider optimism and have significantly stronger predictive power for future returns. Option expiration-driven sales, meaning conversions of in-the-money options into shares and immediate sale, are mechanically motivated and should be excluded entirely from the signal. Failure to exclude them inflates the apparent sell-signal count and suppresses the historically observed asymmetry where purchases outperform sales as predictors. The raw-dollar approach that excludes purchases below $1 million is an imprecise proxy for conviction that penalizes legitimate high-conviction buys by insiders at small-cap companies. The breadth-count approach that counts distinct insiders buying in a month and excludes all sales has some validity in the literature but ignores the normalization problem: ten insiders each buying trivially small amounts in relative terms do not create the same signal as one insider making a concentrated open-market purchase. The shares-outstanding ratio approach that weights all transaction types equally is precisely what erodes the signal, as extensively documented in the insider trading literature.