The Only Copy-Trading Strategy That Passed — U.S. Insider Cluster Buying

⚠ Correction (2026-09-04) — The original figures in this article (+2.78%p, t 5.6) reflect a look-ahead error in which the backtest code entered at the closing price on the filing date itself. Since 68% of Form 4 filings are submitted after the market close, that price was not available to buy at. Recomputing with the actually investable next trading day's open, the same sample (2023H2 onward) gives 20-day +1.22%p (t 2.6), and the full 2015–2026 period gives +0.57%p (t 2.0). The overnight reaction from the filing-date close to the next day's open was +0.76% (t 9.9), meaning most of the original figure was what the market had already priced in immediately after the announcement. Details in Article 17. The body below preserves the original text, with corrected figures shown alongside in the table.

In this series, following foreign investors (Article 1), institutions (Article 2), Congress (Article 3) and Buffett (Article 4) were all rejected. The fifth target, tested with the same methodology, passed: open-market buying by insiders at U.S. companies.

Test Method

• Data: SEC insider transactions data sets, 11 quarters (2023Q3–2026Q1), 50,031 open-market purchases (TRANS_CODE=P) — only trades where officers, directors, or 10% holders bought with their own money (compensation and option exercises excluded)
• Signal aggregation: by (ticker, filing date). U.S. insider filings have a lag of 2 business days, the shortest of any disclosure tested in this series
• Entry: closing price on the trading day after the filing date → Correction: the code entered at the closing price on the filing date itself (look-ahead). The struck-through figures in the table below are that result; the corrected figures use the next trading day's open

Results

By year (cluster, 20 days): 2023 +3.94 / 2024 +2.07 / 2025 +2.63 / 2026 +3.52%p — positive in all four years.

Why Only This One Passed

This is consistent with the common conclusion of the series: what determines whether copy trading works is not "whose trades you follow" but the disclosure lag and the half-life of the information. The fact that the effect weakens at 60 days and disappears at 120 days tells us the half-life of this information is about one month — and the disclosure lag is 2 days. Information with a one-month half-life can be caught with a 2-day lag. Congress (lag of up to 45 days) and 13F (45-day lag) have half-lives shorter than the lag, so nothing was left by the time of disclosure.

The structure of the signal — "several officers each buying with their own money on the same day" — also makes sense: a single purchase can have many explanations (rebalancing, signaling), but simultaneous purchases are unlikely to coincide by chance.

So Why Hasn't It Been Deployed Live Yet (Reservations)

1. 18% price coverage — our U.S. price data (1,596 large- and mid-cap names) covers only 18% of the tickers in the insider signal. Excluding small caps is conservative in direction, but questions about sample representativeness remain
2. Cross-event market correlation may have inflated the t-statistics
3. Whether +2.5%p per month survives real-time collection lag and execution slippage is untested
4. A 2.5-year sample — it may depend on the 2023–2026 bull market

The plan is to extend the test back to 2015 with historical data and to fill in small-cap prices before judging. Attaching this many reservations even to a passing verdict — that is the same principle behind why this series trusts its rejections.

Data Sources

• SEC Insider Transactions Data Sets (Form 4), self-collected U.S. equity price database