What 50 Years of Insider-Trading Research Has Found
A descriptive synthesis of the Form 4 and 13F literature, its strongest distinctions, and its limits.
What 50 years of research has found
The durable finding is conditional. Legal corporate-insider transactions contain information in historical samples, but the average Form 4 is not a uniform research object. Results vary by transaction direction, insider role, firm size, information environment, disclosure delay, portfolio weighting, benchmark, and trading cost. Jaffe (1974) established the event-study foundation. Seyhun (1986) showed why firm size and outsider costs constrain the interpretation. Lakonishok and Lee (2001) made purchase–sale asymmetry and agreement across multiple insiders central. Jeng, Metrick, and Zeckhauser (2003) then separated insiders’ own historical portfolio returns from the result available to an outside observer.
The literature does not support translating a filing into a statement about what one security will do. Insider Atlas therefore grades the published evidence for a class of patterns, never a company or person. A-grade papers most directly support non-routine open-market purchases and multi-insider purchase clusters. B-grade work supplies strong but conditional evidence about costs, firm size, R&D, governance, disclosure, and role. C captures market-wide aggregation. D marks weak or heavily confounded sale and routine-calendar interpretations. E marks raw institutional holder counts and publication-sensitive anomalies whose meaning is especially ambiguous without controls.
Routine versus opportunistic is a first-class split
Cohen, Malloy, and Pomorski (2012) is the decisive modern refinement. Their historical-calendar classifier assigns insiders with recurring monthly patterns to a routine group and places the residual activity in an opportunistic group. More than half of their transaction universe is routine, and the documented return and firm-news associations concentrate in the residual group.
The current site exposes a deliberately narrower approximation. Codes associated with grants, issuer dispositions, tax withholding, exercises, and other mechanical activity (A, D, F, I, and M) are marked routine. Open-market P and S transactions remain eligible for cluster detection. That code rule cannot identify a recurring annual open-market sale, and it can differ from the paper’s person-history classifier. Routine-only windows remain visible in the cluster drill-down but do not enter headline counts. The LOW routine-only event demonstrates the distinction; it is an observable mechanics label, not a judgment about the insiders.
Purchases and sales are not mirror images
Purchases require an insider to commit capital and are less commonly explained by taxes, vesting, option exercise, or diversification. Sales can serve all of those purposes. Lakonishok and Lee find that purchase activity supplies the information in their sample while average sales do not. Jeng, Metrick, and Zeckhauser find significant abnormal performance for the insiders’ purchase portfolio and no significant abnormal performance for their sale portfolio. Cohen and colleagues add the routine-calendar dimension.
This asymmetry is why the interface never renders a sale as the negative image of a purchase. Reported dollars, direction, and transaction code remain factual fields; their colors are not recommendations. A large sale total can reflect compensation and portfolio concentration, and one block can dominate a dollar-weighted measure. Transaction counts, distinct-insider counts, roles, and source filings remain visible beside the total.
The dead-cat-bounce caveat belongs beside every sales view
Sales often follow earlier price strength because insiders diversify appreciated employer stock. Reading every sale as a bearish view can therefore manufacture a misleading “dead-cat” story: the observed sale is real, but the inferred directional meaning is not established. Blackout windows and Rule 10b5-1 plans add more timing motives.
Every sales visualization on this site carries the same caveat: sales often reflect liquidity, diversification, or tax needs; average Form 4 sales are not a reliable negative indicator. The SUNB sales cluster places it directly below the event metrics. Insider and company profiles keep it beside transaction histories, and company pages retain it even when sells dominate reported net dollars. The caveat does not dismiss a sale; it states what the filing alone cannot establish.
Where the current data and literature agree
The current five-day Form 4 sample contains 37 detected cluster windows: 23 open-market windows and 14 routine-only windows. That large routine share is directionally consistent with Cohen, Malloy, and Pomorski’s central warning that mechanical activity is common, though the site and paper use different classifiers. The data also show why sales need special handling. Fifteen of the 23 open-market clusters are sales, and their reported dollars exceed purchase-cluster dollars in this small window. Frequency and dollar prominence do not supply the return evidence that the literature finds weak on average; they make the caveat more important.
The member structure likewise validates the need for transparent definitions. The BGDE event contains distinct insiders within one ISO week, while each page preserves the underlying EDGAR filings. That is consistent with the measurement emphasis in Lakonishok–Lee and Alldredge–Blank (2019): colleague clustering is different from one person splitting a transaction across dates.
Where the current data and literature disagree—or cannot yet be compared
The site’s recent sample is sell-heavy, while the strongest academic evidence concerns purchase-side events. That is a difference in what is common versus what has the clearest historical support, not evidence against either result. The current dataset also contains no security-return series, factor model, delisting returns, R&D measure, firm-policy history, or general-counsel approval data. It cannot reproduce the literature’s abnormal-return estimates, small-firm splits, high-asymmetry tests, or governance mechanisms.
The 13F module has a similar boundary. Lewellen (2011) shows that institutions in aggregate closely resemble the market. Calluzzo, Moneta, and Topaloglu (2019) show that institutions respond to published anomalies, making overlap consistent with attention and crowding. Frazzini, Kabiller, and Pedersen (2018) demonstrate why manager results need factor and leverage attribution. A delayed long-equity snapshot cannot settle those questions. Insider Atlas reports the snapshot, lag, overlap, and provenance; the literature tells readers where inference must stop.