Insider Cluster Re-Test — We Extended It to 45 Quarters, and Found a Look-Ahead Error in the Process

⚠ This article was corrected on 2026-09-04. The conclusion as published on 9/3 ("still passes, 60-day holding +4.04%p") was based on figures that contained the look-ahead error described in §3 below. What was originally written stays in §1–2, and §3–5 record what was wrong and what remains. Not deleting a wrong article but appending the correction is the rule of this series.

1. What we originally set out to do

Part 12 was the only pass among five copy-trading tests, but its sample was 11 quarters. We downloaded the SEC dataset back to 2015Q1 and re-ran it over 45 quarters, 271,293 open-market purchases. The definition is the same — cluster = transaction-code P purchases by two or more different officers, directors, or 10% holders on the same ticker and same filing date, versus SPY, 0.2% round-trip.

2. The results published on 9/3 (contains look-ahead — do not cite)

On the basis of this table we wrote "half of it survives out of sample" and "the half-life is three months or more." We retract both.

One thing we found in this process does hold up: a single ticker's "2-person $760M purchase" appearing three days in a row turned out to be an investment company and its owner as an individual jointly reporting a single purchase under two names on one filing. A definition that counts by number of reporters mistakes such cases for clusters. We changed the definition to "2 or more separate accessions & 2 or more reporters" (8% of the total are joint filings), and all corrected figures below use that definition (n=2,279).

3. What was wrong — we were buying at the closing price on the filing date

It came out when we reopened the pipeline to compute the out-of-sample figure for the 60-day holding separately. The default behavior of searchsorted(filing date), the code that finds the entry point, is "the first trading day on or after the filing date" — and if the filing date is a trading day (98%), that is that day's close. The documentation said "next trading day's close"; the code did not. It was inherited as-is from the shared function of the 8/20 research, so Part 12 has the same error.

Why this is fatal: measuring actual EDGAR acceptance times, out of 100 cluster filings, 68% are after the close (16:00–22:00 ET), 28% are intraday, and 4% are pre-open. Buying at the filing date's close is physically impossible for two-thirds of the signals, and for the remaining third, if the second filing arrives after the close you cannot know that day whether it is a cluster.

And that one day was most of the signal:

Overnight return from the filing-date close to the next day's open is +0.76% (t 9.9), and the next day's intraday return is +0.26% (t 2.8). Three-quarters of the reaction to the filing is over before the open. The proposition this series has confirmed repeatedly ("the half-life of information is everything") here came out to under one day.

4. What the 60-day effect really is — the single year 2020

The +3.57%p at 60 days with next-session-open entry still looks large. Split by year:

• 2020: +19.1%p (n=385 — officers bought during the March crash and a V-shaped rebound followed)
• The 11 years excluding 2020: +0.68%p, t 1.4. The sign swings between -3.05 and +5.09
• Looking only at Part 12's sample (2023H2–), it is -0.47%p

Yesterday's conclusion that "60 days is the strongest" was a case of being fooled by the single year 2020. The original observation in the 8/20 note ("weakens after 60 days") was right.

5. What remains

• Next-session open · 20 days · clusters of $1M or more in value: +1.48%p, t 2.5, n=472 (~43 per year). Only the top quartile by value is positive; the bottom half is 0. Weak, but not zero
• However, in the universe for which we have price data (large and mid cap), this signal occurs 1–2 times per month. Six months of tracking gives a sample of 6–12 — statistically nothing can be decided

6. Decision

We are removing it from the live-candidate list. The collector (daily EDGAR → cluster detection) costs almost nothing, so we leave it running and just log the signals to the ledger. At whichever comes first — after March 2027, or 60 independent clusters within the universe — we will score the live sample on next-session open · 20 days · versus SPY close. We will not change the definition at that point. If it is under +1%p or negative, we terminate.

7. Korea did not work

We applied the same idea to Korean filings (OpenDART reports of ownership by officers and major shareholders). 2,236 currently listed tickers, 32,274 rows from 2024-09 to 2026-09. "Holding increase" clusters, 60 days: -4.16%p (t -3.6). The cause is the data — the Korean API does not distinguish the reason for the holding increase (on-market purchase / stock options / bonus / bonus issue), so what we tested was not "insider buying" but "changes in officer holdings." There is no filter corresponding to transaction code P in the U.S. Form 4.

8. What this episode leaves the series

We wrote look-ahead as the first item in Part 11, and two days after writing that article we found our own look-ahead. One line of entry-point code differed from the sentence "next trading day," and nobody read the two of them together. We are adding one item to the backtest methodology list: print the dates directly from the code and check the entry point with your own eyes. Do not trust the sentence in the documentation.