October 12, 2026 · BriMindInvest Research Team · 17 min read
“When three executives buy their own stock in the same quarter, pay attention.” It is the single most repeated piece of advice in retail investing — and almost nobody who repeats it has measured it. We did. Across 929 cluster-buy events from October 2023 to July 2026, the signal beat the S&P 500 by less than one percent over three months, and lost to it by 4.44% over twelve. Here is the full dataset, including the three breakdowns that change how you should read a Form 4.
Backtests are easy to rig, so here are the rules before the results. Every one was fixed in advance and applied mechanically to the full universe — no stock picking, no hindsight filters.
An insider can file a Form 4 up to two business days after the trade, and the cluster is only visible once the third filing lands. By then the price has often already moved. The gap between “what the insider paid” and “what you could have paid” is where most of the published edge in insider studies lives. We measure only the second one.
Across all 929 events, here is the performance against a dividend-adjusted S&P 500 at three horizons. Note that the sample shrinks as the horizon lengthens — a cluster from March 2026 has a three-month window but not a twelve-month one.
| Horizon | Events | Avg Excess vs SPY | Median Excess | Avg Raw Return | Hit Rate |
|---|---|---|---|---|---|
| 3 months | 929 | +0.98% | -1.47% | +6.60% | 45.5% |
| 6 months | 841 | -1.44% | -5.14% | +9.48% | 43.4% |
| 12 months | 653 | -4.44% | -11.87% | +15.11% | 39.2% |
The centre line is the index. Anything left of it means the signal lost.
The intuitive refinement is to follow the money — ignore a director buying $20,000 of stock and pay attention when someone writes a seven-figure cheque. The data does not support it.
| Cluster size | 3m events | 3m excess | 3m hit | 12m events | 12m excess | 12m median | 12m hit |
|---|---|---|---|---|---|---|---|
| Under $1M | 694 | +1.88% | 46.8% | 503 | -4.76% | -10.09% | 39.8% |
| $1M – $5M | 155 | -2.53% | 41.3% | 103 | -4.75% | -15.08% | 34.0% |
| $5M+ | 80 | -0.05% | 42.5% | 47 | -0.31% | -23.61% | 44.7% |
At three months, clusters totalling $1M–$5M underperformed the index by 2.53% — worse than the sub-$1M bucket, which actually managed +1.88%. At twelve months the mid-size bucket has the lowest hit rate in the entire study (34.0%) and a median excess of −15.08%.
The $5M+ bucket looks respectable on average at twelve months (−0.31%) until you read the median: −23.61%. That is the widest average-versus-median gap anywhere in the data. Translation: a few giant purchases worked spectacularly and most were badly wrong. With only 47 twelve-month observations, this bucket is also the thinnest — treat it as suggestive, not settled.
The plausible explanation is selection, not stupidity. Large open-market purchases cluster in small, volatile, cash-hungry companies where an insider with conviction can move the needle — and where the base rate of failure is high regardless of who is buying. The purchase size is telling you about the company's risk profile, not about the insider's information.
If three insiders is a signal, four should be a stronger one. It was not.
| Breadth | 3m events | 3m excess | 12m events | 12m excess | 12m median | 12m hit |
|---|---|---|---|---|---|---|
| Exactly 3 buyers | 621 | +1.45% | 424 | -4.01% | -15.30% | 36.8% |
| 4 or more buyers | 308 | +0.02% | 229 | -5.22% | -3.82% | 43.7% |
Four-or-more-buyer clusters produced a worse average twelve-month excess return (−5.22%) than three-buyer clusters (−4.01%). The one place breadth helps is consistency rather than magnitude: the broader clusters had a much better median (−3.82% vs −15.30%) and a higher hit rate (43.7% vs 36.8%).
Three-buyer clusters are a lottery-ticket distribution: usually quite wrong, occasionally enormously right. Four-plus clusters are more tightly packed around a mildly disappointing outcome. Neither one beat the index. If you were going to trade either, the broader cluster is the less erratic of two losing propositions — which is not a recommendation.
This is the most interesting cut in the dataset. For every insider in every cluster, we looked up their earlier open-market purchases and asked whether those past buys beat the S&P over the following six months. Then we sorted the clusters by their buyers' historical record.
The standard advice is to follow proven winners. The data says something close to the opposite.
| Buyers' track record | 3m events | 3m excess | 3m hit | 12m events | 12m excess | 12m hit |
|---|---|---|---|---|---|---|
| Past buys beat the S&P | 187 | +0.17% | 46.0% | 112 | +2.73% | 44.6% |
| Past buys lagged the S&P | 265 | +2.57% | 50.2% | 167 | +1.09% | 43.1% |
| No earlier buys on record | 477 | +0.41% | 42.8% | 374 | -9.05% | 35.8% |
Clusters bought by insiders with a losing history returned +2.57% excess at three months with a 50.2% hit rate — the only bucket in the entire study to clear a coin flip. Clusters bought by insiders whose past purchases beat the index returned just +0.17%.
The honest reading is mean reversion plus small samples. An insider's “track record” is usually built on two or three prior purchases — nowhere near enough to separate skill from luck. Sorting on it mostly sorts on noise, and noise reverts. The practical lesson is not “follow the losers”; it is that insider track records carry no forward information, so any site ranking insiders by past performance is selling you a sorted random number.
The largest bucket by far — 477 three-month events and 374 twelve-month ones — is clusters where none of the buyers had any prior open-market purchase on record. These were catastrophic over a year: −9.05% average excess, a −16.52% median, and a 35.8% hit rate.
This is the only cut in the study that materially separates good events from bad, and it is a negative filter rather than a positive one. A board that has never bought its own stock before suddenly buying together is not a vote of confidence — in this sample it was most often a distressed company whose insiders were buying cheap after a collapse, with the collapse continuing.
Six individual events from the sample, all with a complete twelve-month window, show why a median and an average diverge so sharply here. These are not the best and worst of all 929 — they are drawn from the most recent 300 events the dataset stores in detail — but the spread is representative.
| Ticker | Signal date | Buyers | Cluster size | Track | 12m stock | 12m SPY | Excess |
|---|---|---|---|---|---|---|---|
| ABSI | 2025-09-24 | 3 | $0.4M | none | +268% | +18% | +249% |
| MRVL | 2025-09-25 | 4 | $2.1M | lagged | +204% | +17% | +187% |
| LBRX | 2025-09-16 | 4 | $75.0M | none | +158% | +17% | +141% |
| IMRX | 2025-10-02 | 4 | $0.1M | beat | -33% | +17% | -50% |
| BRCB | 2025-09-16 | 5 | $311.9M | none | -74% | +17% | -91% |
| FLUX | 2025-09-17 | 5 | $0.0M | none | -79% | +17% | -95% |
ABSI returned 268% while FLUX lost 79% — both were three-or-more-insider clusters inside the same three-week stretch of September 2025. Any strategy whose outcomes range from +249% to −95% excess is not a signal you can size a position around. It is a coin flip with a very long tail, and the tail is what makes it feel like skill when it works.
Note also that BRCB — a $311.9M cluster, five buyers, the kind of event that generates headlines — lost 91% to the index. The loudest version of the signal was among the worst outcomes.
The older academic literature on insider buying was genuinely encouraging, and we are not disputing it. Studies through the 1980s and 1990s repeatedly found abnormal returns following insider purchases, with the effect strongest in smaller companies and in clustered buying. Those findings are the origin of the conventional wisdom this backtest contradicts.
The most likely explanation for the gap is not that the old work was wrong. It is that the conditions that produced the edge no longer hold. Four changes stand out, and together they are enough to account for it.
Before 2003, Form 4 was filed on paper, due by the tenth day of the month after the transaction. An insider could buy on the second of the month and the filing might not be readable until the tenth of the next — nearly six weeks later, and then only by someone who physically obtained the document. Sarbanes-Oxley cut the deadline to two business days and mandated electronic filing on EDGAR.
That is the single most important change. An edge that rested on a six-week information advantage available to a handful of diligent readers cannot survive a two-day deadline and a public, machine-readable feed. Anything in the filing is in the price almost immediately, and our entry rule — the first close after the completing filing — is deliberately built to measure the world after that change rather than before it.
Insider-buying screens are a standard feature of every retail brokerage platform and every professional data terminal. Cluster buying in particular is a well-known, widely published heuristic. A signal that is free, mechanical, and universally monitored is exactly the kind that gets arbitraged to zero — and a signal that is more than arbitraged away, because of crowding into the same names at the same moment, can produce the negative excess returns we actually measured at longer horizons.
Executives today receive far more of their pay in equity than their 1980s counterparts did. An insider who already holds a large, concentrated position has a different reason to buy more, and the pool of insiders whose purchase reflects a pure, well-informed valuation judgement — rather than optics, board policy, or a share-ownership guideline — has shrunk. Many companies now require directors and officers to hold a multiple of salary in stock, which mechanically generates open-market purchases that contain no view at all.
Management understands perfectly well that the market reads insider buying as bullish. A modest, well-publicised purchase after a share-price decline is a cheap and entirely legal way to express confidence. The cost is small, the headline is favourable, and the information content is close to nil. Our track-record breakdown is consistent with this: the insiders with the best apparent prior records were not the ones whose purchases subsequently performed best, which is what you would expect if a meaningful share of buying is communicative rather than predictive.
Our sample runs from October 2023 to July 2026, which is a single market regime — and a strong one for large-cap indices in particular. Insider buying is structurally tilted toward smaller companies, so a period in which mega-caps led the index is a period in which almost any small-cap-tilted signal will show negative excess return against SPY. Part of what we measured is almost certainly that tilt rather than the signal itself.
We would not claim this backtest proves insider cluster buying never works. It shows that over 929 events in this window, it did not beat the index after costs you could actually have paid — and that the conventional refinements for improving it made it worse, not better. That is a narrower claim, and it is the one the data supports.
If you want to check this yourself — and you should — these are the errors that will make your version look far better than reality. Each one has produced published results that did not survive contact with a real portfolio.
Add transaction costs and the picture gets worse again. Insider clusters concentrate in illiquid names with wide spreads, and a backtest executing at the close ignores the slippage you would actually incur entering and exiting a thinly traded small cap.
A historical backtest is only half of an honest record. The other half is a forward log that cannot be revised, so here is ours as it currently stands.
Since 28 September 2026 the detector has identified 86 insider clusters, graded by conviction:
We want to state this plainly rather than let you discover it. Every forward performance statistic on the live tracker currently has a sample size of zero. Not a small sample — zero. The tracker began on 28 September 2026, which is two weeks ago. The shortest forward window we measure is three months, so the first resolvable results cannot appear before late December 2026, and the twelve-month figures will not exist until autumn 2027.
Any site showing you a live insider-signal hit rate two weeks after starting to record one is showing you a backfilled number. Backfilling a forward log defeats its only purpose, so ours is blank and will stay blank until the windows close.
There is a subtler reason the conviction grades are worth watching rather than trading. The grades are assigned by the same logic whose components the backtest just examined — purchase size, number of buyers, buyer track record — and none of those three improved outcomes in the historical sample. Two of them inverted. So the honest expectation going into this forward period is that the high-conviction bucket will not outperform the low-conviction bucket, and if it does, that will be the genuinely surprising result worth writing about.
That is an uncomfortable thing to publish about your own scoring system, which is roughly why it is worth publishing. The grades are a description of what the filings look like, not a claim about what the stocks will do.
None of this says Form 4 filings are worthless. It says cluster buying is not a buy signal you can trade mechanically. The data is still genuinely useful in three narrower ways.
If you already like a company on fundamentals and the insiders are quietly selling into strength outside their scheduled plans, that is worth knowing. The asymmetry matters: insiders buy for one reason and sell for a dozen, but unscheduled, clustered selling narrows the list considerably. We cover how to isolate that in our companion piece on reading insider selling.
The “no prior buys on record” bucket at −9.05% over twelve months is the most actionable number in this study, and it tells you what to avoid. A first-ever cluster buy at a company whose stock has already collapsed is a value trap far more often than a bottom.
Knowing that a CEO holds a large unhedged position tells you something durable about alignment. That is a governance observation with a multi-year shelf life, not a trade with a three-month horizon.
We publish the figures that argue against our own tooling because a signal study you cannot check is worth nothing. Every number above is regenerated weekly and rendered live on the insider dashboard.
The backtest above is regenerated on a schedule from the same pipeline that powers our insider dashboard: a 3,200-ticker universe, Form 4 filings parsed directly from SEC EDGAR, insider roles resolved, and scheduled 10b5-1 sales separated from discretionary ones.
You can see the current cluster buys, the insiders behind them, and — importantly — each buyer's measured track record, so you can judge the signal rather than take our word for it.
Insider cluster buying is the most widely repeated bullish signal in retail investing, and across 929 measured events it did not work. It added less than one percent over three months, lost 1.44% over six, and lost 4.44% over twelve — with a hit rate that never reached a coin flip and a median outcome substantially worse than the average at every horizon.
Every intuitive refinement failed too. Bigger cheques were not better. More buyers were not better. Sorting insiders by their past success was worse than useless, because those track records are built on samples far too small to mean anything.
The one durable finding is a negative screen: when a board that has never bought its own stock suddenly buys together, the stock lost 9.05% to the index over the following year and was down more often than not. That is a warning, not an invitation — and it is the opposite of how the signal is usually sold.
Use insider filings to understand incentives and to check a thesis you already hold. Do not use them as a reason to buy something you have not otherwise researched.
Get AI prediction signals, unlimited stock comparisons, portfolio analytics, and personalized watchlists — free for 14 days, no credit card required.
14-day free trial · No credit card required · Cancel anytime
Not ready to sign up? Get one free email a week instead.