OUR OWN STUDY · AS OF 20 OCT 2026
Do insider buys really work? We recalculated 301,904 of them.
A lot is claimed about insider buying and very little is measured. So we measured it: 301,904 open-market purchases from the official Form 4 filings with the US Securities and Exchange Commission, 2016 to 2026, 8,319 stocks. Plus daily prices, a control test and a strict split into a search and a test period.
The result is uncomfortable: almost everything you read about insider buying does not survive the check. One pattern does – and it is not the one people advertise.
An insider purchase is a signal when the stock is liquid and already rising: at least +10 % over the 20 trading days before the filing. Such cases returned +2.11 % against the S&P 500 over 20 trading days after costs. The same days, the same filter, but without an insider purchase: −0.62 %.
How we calculated
The method decides the outcome, so here it is in full:
- Real purchases only. Transaction code P – an open-market purchase. Grants, option exercises, tax withholdings and gifts do not count.
- One event per stock and filing day. A filing often consists of dozens of individual lines. What counts is what an observer sees: “insider purchases were reported for this stock today”.
- Entry only once you could have known. The open on the day after the filing – not the insider’s own trade date, which only becomes visible days later.
- Holding period 20 trading days, measured as excess return against the S&P 500 (SPY) over the same window.
- 0.20 % trading costs per trade, round trip – realistic for liquid US stocks.
- Search and test period separated. Rules were searched in 2016–2022 and checked in 2023–2026. Never the other way around.
- Confidence bands by bootstrap. Anything whose band does not sit entirely above zero counts as chance.
The trap that nearly ruined everything
The first run looked spectacular: CEO purchases +45.9 %, large purchases +59.7 %. Only the median did not fit – it was negative, with a hit rate of 43 %. Here is why:
| Stock | Excess return, 20 days | Daily turnover of that stock |
|---|---|---|
| ETST | +129,895 % | $4,634 |
| AIVN | +6,572 % | $461 |
| IGSC | +5,042 % | $149 |
These are penny stock shells with a few hundred dollars of daily turnover. Nobody can buy there – a single order moves the price by percentage points. 318 of 73,125 events, that is 0.43 %, carried the entire positive average. Without them it was negative.
So everything that follows requires a minimum of tradability: one million dollars of average daily turnover. That is the difference between a pretty number and a usable one.
What does not work
All figures in tradable stocks, 20 trading days, after costs, against the S&P 500:
| Popular pattern | Measured | Verdict |
|---|---|---|
| Any insider buys | +0.11 % | worthless |
| Cluster buy – several insiders of the same company | +0.26 % | worthless |
| The CEO buys | +0.34 % | worthless |
| Large amount – above $1m (after 60 days) | −2.08 % | actually hurts |
| Insider buys into a falling price | −0.34 % | failed in the test |
| Filtering out routine buyers | +2.23 % | no added value |
One case is particularly striking: buying into weakness – an insider stepping in while the stock falls – looked like the best find of all in the search period at +7.65 %. In the test period it collapsed completely. That is exactly what the split is for.
What does work
The rule: an open-market purchase + a liquid stock (≥ $1m daily turnover) + the stock has already risen ≥ +10 % over the previous 20 trading days.
Entry at the open on the day after the filing, hold for 20 trading days.
| Avg. per trade | Median | Hit rate | Cases | |
|---|---|---|---|---|
| With an insider purchase | +2.11 % | −0.63 % | 47.5 % | 5,879 |
| Without – same days, same filter | −0.62 % | −1.70 % | 43.3 % | 29,105 |
The difference is +2.74 percentage points, with a confidence band from +2.10 to +3.41 – entirely above zero. And it is positive in all eleven years, including 2019, 2022 and 2024, when both groups were negative in absolute terms.
That is the decisive point: a stock that has already risen does not beat the index on its own. The control group earns nothing. The entire advantage comes from the insider.
The bigger the run-up, the stronger the effect
| Run-up of the stock (20 days) | Avg. per trade |
|---|---|
| ≥ +5 % | +1.22 % |
| ≥ +10 % | +2.11 % |
| ≥ +20 % | +3.82 % |
| ≥ +30 % | +5.20 % |
That steady increase is the best sign this is not a chance finding. Random patterns do not behave so neatly.
One refinement also holds: the buyer’s own record
If the buying person averaged better than +3 % on their earlier purchases, the result improves clearly – and in both periods:
| Search 2016–22 | Test 2023–26 | Hit rate (test) | |
|---|---|---|---|
| Base rule | +2.10 % | +2.26 % | 45.6 % |
| + buyer has a good record | +4.62 % | +4.45 % | 53.5 % |
Here the median turns positive and the hit rate rises above 50 %. That changes the character: now the majority of trades win, not just a minority with large outliers. To be honest: in the test period insiders with a poor record also ran positive. So the rule reads good record = bonus, not “poor record = avoid”.
What argues against it
These points belong here even though they weaken the result:
- The hit rate is below 50 % and the median negative. Most individual trades lose slightly; the average comes from the winners. This only works across many positions at once – around 40 on average – never across a single one.
- Survivors distort the picture. Of 5,463 stocks checked, 2,089 no longer had price data – acquired, delisted, bankrupt. Those cases are missing. Since the event and control groups are drawn from the same pool this affects both, but it is not perfectly clean.
- Costs are estimated. 0.20 % round trip is realistic for liquid US stocks. Slippage beyond that is not included.
- Ten years are ten years. They contain the 2020 crash and the weak 2022, but no long bear market like 2000–2003.
- A measured edge is not a promise. Once a pattern becomes well known it can disappear.
Why we publish this
Because we used to claim the same thing as everyone else. Our own insider page said cluster buys were “the most reliable insider pattern”. The measurement says +0.26 %. We changed that.
We applied the same check to our own chart pattern – with a negative result that is published as well: chart patterns in the backtest (in German).
The current hits under exactly this rule are listed in the Insider Radar under “What stands out”.
Source of the raw data: SEC, Insider Transactions Data Sets (Forms 3/4/5), 2016–2026. Price data: daily closes, adjusted for splits and dividends. Past measurements are no guarantee of future results. This is not investment advice.