Research record
We fixed the question, event definition, benchmark and confirmatory horizon before observing the result. This page publishes the method, the failed coverage gate, the amendment and the finding — including the parts that did not support the original belief.
Confirmatory result · 21 trading days
−0.93%
Mean return minus the S&P 500
95% confidence interval
−1.89% to +0.08%
The interval includes zero. No reliable predictive effect was established, and the point estimate is negative.
Company insiders — directors, officers and holders of more than 10% of a stock — disclose their trades on SEC Form 4. A common belief is that several insiders buying around the same time signals favourable future returns. We tested that specific claim.
Most Form 4 activity is not an insider choosing to buy stock in the open market. Across 4.75 years, 52% of transactions were sales, 24% compensation grants, 11% option exercises and 8% open-market purchases. The event definition therefore required every condition below.
This produced 7,544 events, with a median cluster value of $396,702 and a median of three insiders.
The event date was the filing date, not the transaction date. Information cannot be evaluated as though it were public before disclosure. The measurement window opened at the close of the first trading day after filing.
The median disclosure lag was two days, but only 68% of filings arrived inside that window. Measuring from the trade date would have credited a hypothetical strategy with information nobody yet had.
The single confirmatory test was cumulative abnormal return over 21 trading days: the stock’s return minus the S&P 500’s return over the same period. Five-day and 63-day horizons were secondary and could not replace the named test after the result was known.
The preregistration required price coverage for at least 90% of events. Coverage was 80.2%, so the gate failed. The missing pattern initially resembled survivorship bias and the analysis did not simply proceed.
All 1,512 missing events were investigated. Only 4 — 0.3% — involved a company that disappeared during the 21-day measurement window; 89% were still filing more than a year later, and the median gap to a final filing was 912 days. The absence was principally a historical price-vendor coverage gap, not disappearance during the measured window.
Amended before outcomes were observed
The gate became at least 80% coverage and fewer than 2% of missing events disappearing inside the window. A high-coverage 2024–2026 replication with the same sign and overlapping intervals was also required.
Before real returns were analysed, the code was tested on synthetic data with a known 3.00% effect. It recovered 2.95% with an interval excluding zero. A second run with no planted effect correctly returned an interval containing zero.
| Horizon | Mean adjusted return | 95% interval | Status |
|---|---|---|---|
| 21 days | −0.93% | −1.89% to +0.08% | Confirmatory |
| 5 days | +0.87% | +0.52% to +1.22% | Secondary |
| 63 days | −4.47% | −5.89% to −2.98% | Secondary |
The positive five-day observation was not the confirmatory test and is not presented as evidence that clustered buying “works”. Selecting it after seeing all horizons would be the cherry-pick the design was built to prevent.
The preregistered benchmark was the S&P 500, while clustered insider purchases occur disproportionately at smaller companies. Small caps underperformed large caps over the period, so the measure may charge events for a size effect. We did not switch benchmarks after seeing that an alternative could flatter the result.
A size-adjusted study might find a small positive short-horizon effect. That remains a hypothesis for a future preregistered study, not a reinterpretation of this one.
The study establishes that clustered insider buying did not predict positive 21-day market-adjusted returns across 5,849 analysed events, to a precision of about ±1%. It does not establish that filings are useless; they remain factual disclosures of what people close to a company chose to do with their own holdings.
Ironclad’s insider pages describe disclosed activity. They do not predict prices, rank companies, generate alerts or tell anyone what to buy or sell.
The preregistration and amendment were sealed before any outcome was observed.
Return to the factual SEC Form 4 activity feed.