Pre-registered 2026-10-14 (commit 90cc39b) · 1 run · data snapshot 2026-09-19T090800Z
The idea
Not every insider purchase means the same thing. Some insiders buy in the same month every year, for bonus timing or fixed trading windows; those trades carry no news. Others buy off their usual pattern, and those are the ones that should know something. Cohen, Malloy and Pomorski (2012) showed that separating the two makes the opportunistic purchases predict abnormal returns over the following month. The counterparty is whoever sells to an insider with better information.
Three risks were accepted in writing: a published anomaly that may since have been arbitraged away, an effect concentrated in small caps with wide spreads, and a lab base rate in which most ideas die for lack of alpha.
How it was tested
Pre-registered on 2026-10-14 (commit 90cc39b) on a dataset new to the lab: the SEC Insider Transactions Data Sets, 7.35M Form 4 transactions, calibrated against the original filings. Each insider-year was labeled routine or opportunistic using only the three previous years. The event was an open-market purchase by an opportunistic insider, dated by the day the filing was published, with the issuer required to match the current ticker's company. US stocks at $5 or more and $1M of daily dollar volume. Buy at the next open, hold 21 sessions with no stop or target, largest purchases first, up to 20 a day. Threshold t 1.65. The same construction on routine insiders ran alongside as a control that would be reported but would not decide.
Where it died
On 4,476 trades, alpha was −0.585% per trade with t −1.96 and a beta of 1.28. The control came out almost identical: t −2.16, −0.678% per trade, beta 1.13. The thesis predicts a gap between the two groups, and there was none. Gate 8, Gate F and stability failed as well; the only faint resemblance to the paper was 2014-2019 at t +1.25, which turned into −2.52 in 2020-2026.
The record
Thesis
Insiders who buy on a fixed calendar pattern (same month every year — bonus timing, fixed windows) carry no information; insiders who trade without that pattern buy when they know something. Their open-market purchases should predict abnormal returns over the following month.
Prior: Risks accepted in writing: a published 2012 anomaly that may be arbitraged away, concentrated in small caps where spreads are wide, and a lab base rate in which most measured ideas die for lack of alpha.
Cohen, L., Malloy, C. & Pomorski, L. (2012). Decoding Inside Information. Journal of Finance 67(3).
Specification
data
SEC Insider Transactions Data Sets (Form 4), 7.35M transactions, calibrated against the original filings
classification
Insider-year classified as routine or opportunistic using only the three previous years (no look-ahead)
event
open-market purchase by an opportunistic insider, dated by the filing's PUBLICATION date
universe
US stocks, price ≥ $5, 20d dollar volume ≥ $1M; issuer identifier must match the current ticker's company
entry
next open after publication
exit
21 sessions, no stop or target
ranking
purchase value, largest first, max 20 new positions/day
control
same construction on routine insiders' purchases (reported, does not decide)
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Gate 7 — Jensen alphafail
After regressing out market exposure, is there alpha left?
t ≥ 1.65
t (HAC)-1.96alpha / trade %-0.58beta+1.28trades4,476
The routine-insider control came out the same: t −2.16, −0.678% per trade, beta 1.13. The thesis predicts a difference; there is none.
Gate 8 — Capture-preserved alphafail
Does the subset of trades a real portfolio can actually take keep the alpha?
scaled t_cap and ratio ≥ 0.50
t captured-0.61capture ratio+0.00
Gate F — Frictionfail
Is the alpha per trade large enough to pay for its own execution?
alpha/trade ≥ 0.30%
alpha / trade %-0.58
Stabilityfail
Is the alpha present in both halves of history?
t > 0 in both halves
t 2014 2019+1.25t 2020 2026-2.52
Why it died
Negative alpha and beta above 1 for both the opportunistic insiders and the routine control. The calendar-pattern split that separates information from habit in the original paper does not appear in this universe and period. The only resemblance is 2014-2019 (t +1.25, not significant), which reverses in 2020-2026 (t −2.52) — consistent with an anomaly arbitraged after publication, or with the post-2020 regime.
What it taught
The control decided the reading, not the headline t: a negative alpha alone could have been blamed on noise, but a control with the same sign and size says the construction measures small-cap beta, not insider information. The micro-caps where the literature places the effect were excluded by the pre-declared universe and were not added back after the result.
Verdict history
2026-10-14Pre-registered (commit 90cc39b) on a dataset new to the lab
2026-10-14Single run: t −1.96 vs routine control t −2.16 No alpha