Research
The backtest behind the score.
Everyone repeats the same rule: when several insiders buy at once, follow them. We wanted to know whether that is actually true, so we measured it. Every number on this page comes from public SEC filings and market prices, and the method is described in enough detail to be checked.
What we tested
- Every Form 4 filing from January 2023 to March 2026, about 340,000 open-market trades.
- A cluster is two or more insiders buying the same company within a short window, counted only once per company per cooldown period, and only when the filing arrived within 45 days of the trade so it could actually have been acted on.
- 2,465 clusters in total, 2,452 with a full 90-day outcome.
- Outcome: the stock's return from the first trading day after the filing to 90 days later, minus the S&P 500 over the same days.
What we found
| Group | Cases | Avg vs market, 90 days | Beat the market |
|---|---|---|---|
All clusters Two or more insiders buying the same stock within a short window. | 2,452 | −0.6% | 41% |
Buyers with a strong prior record The specific people buying had beaten the market by 15% or more on their earlier group buys. | 131 | +10.8% | 49% |
Everyone else Buyers with a known record that was average or worse. | 919 | 0.0% | 42% |
- The group buying itself is close to worthless. Across all clusters the average result was slightly below the market and fewer than half beat it.
- The buyers' own history is what predicts. When the specific people buying had been right before, the next buy averaged more than ten points above the market.
- The usual "strength" signals added almost nothing. Deal size, the number of insiders, and whether the CEO was among them told us little once the buyers' records were known. Very large clusters of seven or more insiders did worse than average.
- Our first score was wrong. The original version, built from common-sense factors, was slightly anti-predictive on this data. It was thrown away and rebuilt from these results.
What this doesn't prove
- Three years is one market regime. The pattern may weaken or vanish in a different one.
- 90 days is one horizon. Results at 30 and 180 days were directionally similar but noisier.
- The proven-buyer group is small, 131 cases. A strong average from a small group deserves suspicion, which is why the live results page exists.
- Companies that were delisted may be under-represented in the price data, which flatters every group a little.
- Past results, simulated or real, do not guarantee anything. This is research, not investment advice.
How it is used
Every new filing is read as it lands. When a cluster forms, the score is built mostly from the buyers' own records, which are kept for 5,428 insiders and shown on the public leaderboard. The records are rebuilt each time the SEC publishes a new quarter of data, and every live signal's own 90-day outcome is added once it is known, so the method keeps being tested on data it has never seen.
The live results page runs a simulated $1,000 account that takes every strong signal, holds it 90 days, and shows every trade, wins and losses, against the S&P 500. Nothing is deleted or cherry-picked.
Found a flaw? Tell us.
We would rather fix the method in public than defend it. Write to [email protected].
Start free