Methodology
How the signals are scored
Everything on Insider Intent is computed from public filings, shown with its evidence, and held to a statistical bar before we claim anything. The claim itself: About to move — not “goes up.” Here is the whole method, in plain language.
The two-axis read
Every signal carries two separate scores, and they answer different questions.
Notability asks: how unusual is this move? A CEO's first-ever open-market buy, an executive selling most of their stake outside any scheduled plan, five insiders buying the same name in the same week — these are measurably out of the ordinary. Notability is deterministic and auditable: it's computed from the filing itself, so it exists the day the filing lands.
Evidence asks a harder question: how have trades of this TYPE turned out before? A claim about outcomes takes three things: enough similar trades to measure, a correction for multiple testing (we test many signal types, so some will look good by pure luck — we correct for that), and evidence on data the rule never saw.
We are re-running that measurement now. Until it finishes, every evidence grade you see is provisional, and none of them is a claim that a signal type predicts returns. When a signal card says “not enough evidence yet,” that is not a verdict that the signal is worthless — an unusual move can be worth your attention on its unusualness alone.
Where the data comes from
Public regulatory and government sources, spanning every category of informed-money activity — not one feed, and not a pile of unrelated datasets, but the full picture of who is positioned where:
- Insider trades — SEC Forms 3/4/5 (corporate-insider transactions), Form 144 (proposed sales), and 10b5-1 plan adoptions.
- Institutional positioning — 13F holdings and their quarter-over-quarter changes, 13D/13G stakes (activist and passive, as a share of class), and NPORT-P monthly fund books (including shorts and options).
- Congress — Senate eFD and House Clerk member-trade disclosures.
- Corporate actions & catalysts — buyback 8-Ks, share issuance, exec/board changes, M&A, PIPEs, tender offers, Form D, and HSR filings.
- Short-side positioning — FINRA daily short volume and bi-monthly short interest, plus fails-to-deliver.
- Fundamentals & context — SEC XBRL company facts back to 2009, plus government context: federal contract awards, lobbying, clinical trials, and SEC enforcement.
Prices are a multi-source composite (Tiingo primary, with automatic failover). Every figure on the site links back to the filing it came from — we call these the receipts. If you can't click through to the source, we don't show the number.
Source credits: news headlines are found via the GDELT Project. Lobbying figures come from the Lobbying Disclosure Act database, Office of the Secretary of the Senate (lda.gov), accessed on the dates shown beside them; the Senate Office of Public Records cannot vouch for the data or for analyses derived from it. Clinical-trial and drug-approval entries are condensed from ClinicalTrials.gov and openFDA records — shortened to one line per event, with no ownership of the underlying data claimed.
How signals are scored
Each filing is checked against a library of signal patterns — the recurring shapes that make informed-actor activity meaningful. Among them:
- Out of character — a trade unusual for this specific actor: first buy ever, ten times their normal size, a break from years of routine.
- Meaningful exits — sales that reduce a real fraction of a stake, outside any pre-set trading plan. A scheduled plan sale is the anti-story; a discretionary exit is the story.
- Clusters of buyers — multiple insiders at the same company buying in a tight window.
- Buying the dip — insiders stepping in while their own stock is drawn down.
- Crowded-short setups — insider buying into a heavily shorted name.
- Cross-source agreement — different kinds of filing converging on the same name, weighted by how surprising the agreement is (below).
When a pattern fires, the card shows its evidence in words — who did what, how much, and the specific fact that makes it unusual — not a bare score you have to take on faith.
What each figure label means
Every computed figure in the product wears one of three labels, so you always know what kind of number you are looking at. Each label links back to this section.
- Measured — a count of real things: filings read, trades disclosed, days elapsed. Nothing is modelled; the number comes from counting stored records, and its receipt names the records that were counted.
- Simulated — what following each filing from the day it became public would have earned over the stated time frame, measured in trading days. It is a back-test, not money anyone made. Known limits: it uses the disclosure day (never the earlier trade date, so there is no hindsight), it ignores trading costs and whether you could have bought at that day's price, and overlapping filings are not compounded into a portfolio. Insider-return measurement is experimental and labelled as such wherever it appears.
- Too few cases — the figure is shown, but it rests on fewer cases than our floor of 12 graded trades, so we do not treat it as evidence either way. We show it with this label rather than hide it: a thin sample is a fact worth seeing, not a verdict.
Cross-source convergence
When different kinds of filing — say, corporate insiders and an institutional 13F — agree on the same name, that is worth a look — it describes the filings, it is not a claim about returns. Agreement is only interesting when it's surprising. So we measure how often that agreement would happen by chance, given each source's own lean: if a fund holds half the market and insiders buy broadly, overlap is expected and means little. Convergence is scored by how far the agreement sits from what chance alone would produce.
The Insider Intent Index
The index on our public page is the weekly bullish share of directional informed-money disclosures filed that week — corporate insiders (Form 4 and related notices), members of Congress, institutional position disclosures (13F/13D/G), and issuer actions (buybacks, offerings). It counts breadth of filings, not dollars: insider dollar volume is structurally sell-heavy (compensation flow), so the informative signal is the shift in the buy/sell mix. A week's reading is final the moment the week ends — we bucket by filing date, never by trade date, so a number you cite today cannot quietly change tomorrow. The current reading is framed against the index's own trailing-year range because the level alone means little; the deviation is the read. Weeks with too few filings or incomplete source coverage are excluded rather than guessed. Only public-domain filings are counted, no price data is involved, and no model touches the number — it is arithmetic over disclosures, reproducible by anyone.
What we refuse to do
This product exists because most of this category is hype dressed as signal. The refusals are the method.
- No forecasts or price targets. We tell you what informed actors did and how unusual it is. The claim is that a name may be about to move — never that it 'goes up'. We show which way the insiders leaned, a filed fact; we do not predict which way the price goes.
- No AI-generated numbers. Every figure on the site is computed deterministically from filings and linked to its source. The AI writes the words around the numbers; it never produces one.
- No copy-trading or position sizing. We are an information tool. We will not tell you what to buy, how much, or when to sell.
- No cherry-picked win rates. “442% in 30 days” is the genre we exist against. A track record that only shows the winners is marketing, not evidence.
- No blended magic score. How unusual a move is and how much past evidence stands behind its signal type are different questions. We show the two axes separately, always — collapsing them into one number hides exactly what you need to know.
- No edge claims before the statistical bar. We do not publish a performance number for a signal type until it survives our full statistical process — enough samples, multiple-testing correction, out-of-sample evidence. That is also why you don't see one today.
Our own track record
The standard applies to us too. The product scores its own surfaced signals by the same method it applies to everything else, and we will publish that record — hits, misses, and the honest nulls — once it clears the same statistical bar we hold every signal type to. That surface is being built; it ships when it's earned, not before.
Corrections to past scores
30 September 2026 — pre-planned trades. We corrected how we read whether an insider trade was made under a pre-arranged trading plan (a Rule 10b5-1 plan). Filings before April 2023 have no checkbox for this, so we had been treating almost all older sales as unplanned; we now also read what the filing's footnotes say, and we read the checkbox's actual answer where there is one. 240,811 past trades changed — all but six now correctly marked as pre-planned. For those trades the notability score is lower, the trade is labelled routine rather than opportunistic, and flags that only apply to discretionary trades (such as “selling into strength”) no longer appear. Each trade's answer comes from its own filing; nothing was estimated.
4 October 2026 — returns against the market. We recalculated the “beat the market” return (how a stock moved after a filing, compared with what its usual link to the overall market would predict) for 265,475 past filings, so every such figure that was off by a tenth of a percentage point or more now follows the method we use today; smaller differences were left as they were. Most had been calculated by an earlier version of our method: some were measured from a starting day later than the right one, some predate improvements to how we compare a stock with the market, and some had less price history to work from. For 90,122 of them the figure had drifted with no single cause we could pin down, most likely because our price data was corrected after it was calculated. A handful that fail today's data checks were withdrawn rather than recalculated. Plain returns (“did the stock go up”) were corrected only where the starting day was wrong. Track records, leaderboards and signal statistics have been rebuilt, so if an insider's record or rank looks different from last week, this is why. Every figure comes from actual prices, and we kept a copy of every earlier value.
5 October 2026 — plain returns. We recalculated the plain return (“did the stock go up”) for 1,736 past figures that were off by a tenth of a percentage point or more from the method we use today. For 1,065 of them, on 482 filings, we found the exact cause: we could rebuild the old figure to the last digit from our price history, and for all but 41 we confirmed it with a second, independent check. Most had been measured from the wrong starting day, almost always a later one, because our price history at the time was missing the days around the filing; others had been worked out on prices rounded to the cent, usually from before a stock split, or across a split our price history did not yet reflect. The other 671 had no single cause we could prove (many most likely because our price history was adjusted, for example for dividends, after they were calculated); they were recalculated to today's method too. Smaller differences, in plain returns and in returns measured against the market, are left as they are for now; a later recalculation of every figure the same way replaces them. Track records and leaderboards that use plain returns have been rebuilt; site-wide figures barely moved, but if an insider's record looks different, this is why. Every figure comes from actual prices, and we kept a copy of every earlier value.
Insider Intent is an information tool. Nothing here is a recommendation to buy or sell any security.
See the method on live filings: the public signal board.