Questa pagina è disponibile solo in inglese.
Methodology
The rules this site applies, with the numbers currently in force. Nothing here is hidden behind the score.
Which trades count
Only open-market purchases and sales reported on SEC Form 4 (transaction codes P and S) in common stock. Awards, option exercises, gifts and other transfers are left out. A purchase below $1,000 does not count towards dollar totals. The same trade reported twice (an amendment, or co-filers of one purchase) is counted once.
What is a buying cluster
A run of purchases at one company by 3 or more distinct insiders. A pause longer than 30 days between two purchases ends the run and starts a new one. A cluster stays in the active list while 3 or more distinct insiders bought in the last 180 days. A company can therefore have several runs; the list shows the latest.
What is flagged
- Offering: IPO allocations, private placements and similar subscriptions, found from footnotes, from 8-K filings, or because 3 or more insiders bought at exactly the same price on one day near the listing or outside the day's traded range.
- Routine and plan: purchases under a trading plan (Rule 10b5-1), a dividend reinvestment or compensation plan, or with a regular cadence.
Flagged purchases stay visible and are discounted in the score. The list can hide clusters that are mostly IPO or plan buying.
The score shown
The Insider Score on every list and cluster page is the match score, plus the points of the AI review when there is one. The match score adds up seven inputs, each times a fixed weight, starting from a base of -26.24, and is kept between 0 and 100. The weights were calibrated on past clusters. Each cluster page shows the points of each input.
- +12.397 per tenfold of the dollars bought, counted up to $5,0M
- +6.205 times the natural log of the number of distinct insiders: each extra insider adds less
- +10.586 when a CEO is among the buyers
- -18.570 times the change from the 1-year high to the close before the buying: a 30% fall adds 5.6. A fall beyond 30% adds nothing more: the points are counted up to that fall, and the cluster page still says how far the stock really fell
- -0.112 per day between the first and the last purchase
- -5.594 times the share of the dollars in routine or plan purchases
- -12.673 times the share of the dollars bought by the largest buyer
On the cluster page the base has no line of its own: it is taken into the dollars line, which counts the dollars from $131 instead of $1 (at that amount the dollars and the base cancel out). The lines then add up to the match score by themselves.
Beside each input, a bar from 0 to 100% shows how much of the points that input can give, or keep, it has: a long bar is always good, also for the inputs that take points off. The dollars run from $10,0K (empty) to $5,0M (full); the insiders are full at 10; the fall from the 1-year high is full at 30%; the days from the first to the last purchase leave nothing at 60; routine buying and the largest buyer's share are full at none and at an even split. These limits are display choices, not part of the score.
The studies behind the score
| Study | Found | Input here |
|---|---|---|
| Jeng, Metrick & Zeckhauser, Review of Economics and Statistics, 2003 | Insider purchases were followed by returns above the market; sales were not | Only purchases are scored |
| Piotroski & Roulstone, Journal of Accounting and Economics, 2005 | Insiders tend to buy after the price has fallen and when the stock looks cheap | The fall from the 1-year high |
| Cohen, Malloy & Pomorski, Journal of Finance, 2012 | Purchases made in the same month every year carried almost no information; the others did | Routine and plan purchases count for less |
| Alldredge & Blank, Journal of Financial Research, 2019 | Purchases by several insiders close together were followed by higher returns than single ones | The number of distinct insiders |
| Kang, Kim & Wang, 2018 | Clusters with both executives and directors did best; clusters of directors only showed no advantage; the clusters studied lasted at most five days | The days from first to last purchase |
These studies cover years up to 2016 and returns before costs. What followed the clusters listed on this site is on the Track record page: since 2016 the score has not ordered returns beyond the first days.
The research score (not used for ranking)
A weighted blend of five factors, each from 0 to 100%, times the discounts for routine and offering buying, minus a penalty for insider selling in the 90 days before.
- 25% Commitment: the size of the purchases against the company and against what the buyers already held
- 20% Breadth: how many distinct insiders bought
- 20% Seniority: the rank of the buyers
- 20% Context: buying after the stock fell against the market
- 15% Timing: purchases close together
The AI adjustment is a separate, visible number: a model reads the filings and the news around the cluster, classifies the buying, and adds or subtracts points. It is never used in the track record.
Where there is no AI review, the pipeline flags can add their own line: points taken off when the filings themselves (Form 4 footnotes, 8-K filings, prices) show the buying was an offering, a plan or a 10b5-1 programme. It is shown apart from the match score and never used in the track record.
What happened after past clusters, measured from the next market open, is on the Track record page.
Missing data
A missing input counts as neutral and the cluster is marked "partial data"; it is never replaced with an estimate. A share count filed more than 400 days ago, or dated before the company's first trading day, gives no market cap.
Returns and windows
Every past return on the site is shown one way: the stock's change minus its benchmark's over the same days (a small-cap index for companies under $2B, a large-cap index otherwise), with the stock's own change beside it. A ticker page sums up the last 90 days, the trade tape the last 7 days and the insider history the last 5 years; each heading names its window.
Trades under $10,000, 10b5-1 plan trades in the rankings, and filings more than 30 days late count for less on the pages: hidden by default or marked, never removed from clusters, scores or totals.
Timing
What was known is decided by the SEC's acceptance time of each filing, in New York time. New filings are picked up during business days and prices are end-of-day; the site warns when its data is out of date.