Methodology
Signals are computed from public regulatory filings and end-of-day market data. Statistics describe the historical record; AI explains it. They do not predict future returns.
How to read the colors
Three colors, three meanings — fixed on every screen, and colorblind-safe by default (no red/green axis anywhere).
The machine found something
Chartreuse marks computed signals and detected patterns.
You can act on it
Green text means interactive — links, buttons, the active page. It never encodes data.
Money in, money out
Cobalt = buys, vermilion = sells — always with a written label, never color alone.
The signals
- Cluster buy / sellCluster
- At least 3 distinct insiders trading the same company in the same direction within a 90-day window. This describes reported activity, not independent investment decisions or a forecast.
- Unusual size (z-score)Unusual size
- The transaction's USD value is at least 2 standard deviations above that insider's own trailing history (up to 20 prior same-direction transactions; requires at least 3).
- First buy after 12 monthsFirst in 12mo
- The insider’s first purchase after at least 12 months without buying. Check the source filing and the earlier activity before interpreting the gap.
- Routine vs informative
- Buys recurring at near-fixed intervals (coefficient of variation ≤ 0.25 across intervals) are flagged routine — typically plan-driven and less informative.
- Forward returns & alpha
- Returns start at the transaction date and use adjusted closes over 21, 63, 126 or 252 trading days. Alpha subtracts the source-market benchmark return (SPY for the US, FCHI for France, DAXK for Germany, AEX for the Netherlands, OMXSPI for Sweden). A missing benchmark leaves alpha unavailable.
- Track record (win / hit rate)
- Per insider, over their nonderivative purchases (minimum 5): win rate = share of buys with positive forward return; hit rate = share with positive alpha; plus median returns per horizon.
- Conviction score
- A score from 0 to 100 for a nonderivative purchase. It combines unusual size, buying clusters, the buyer’s role, a first buy after 12 quiet months, the change in their holding, drawdown and recent price context. It describes the recorded activity, not the probability of an investment outcome.
- Proven insider
- A historical badge for natural persons with at least 10 matured purchases and a Wilson lower bound of at least 60% for a positive 63-trading-day stock return. It does not establish outperformance against the market or predict future results. Funds and holding entities are excluded.
- Signal backtests
- Historical returns grouped by signal, with sample sizes and confidence intervals. The starting date is the transaction date, so some performance can precede public disclosure. This is not a simulation of a strategy executed after publication and does not account for trading costs.
- Increase in holdings
- For a purchase, quantity bought divided by holdings before the trade (holdings after minus quantity bought). A value of 100% means the position doubled. Unavailable for sales, unknown holdings or an initial position with no positive prior balance.
- Filing delay
- Calendar days between the transaction date and the filing date in UTC, with a minimum of zero. This measures elapsed time, not whether a legal filing deadline was met.
- Backtest hit rate
- Share of events with a strictly positive market-adjusted return at the selected horizon. Only events with an available return enter the calculation. Zero excess return is not a hit. This is a historical proportion, not a forecast.
- Median market-adjusted return
- Middle value of the event returns minus their benchmark returns at the selected horizon. Only events with an available return enter the calculation. Cells with fewer than 100 events or 24 distinct filing months are marked as insufficient. This is not an annualized return or a forecast.
Plain vocabulary
The words a filing uses, in ordinary language. Nothing here is a signal — it is what the raw record means.
- Form 4
- A US disclosure of changes in beneficial ownership by a reporting insider. It can contain purchases, sales, grants and derivatives. Read both securities tables and the footnotes.
- Nonderivative purchase
- Our purchase filter uses code P and excludes derivatives. For SEC filings, P covers both exchange and private purchases; the code alone does not establish where the trade took place. Grants, option exercises and gifts have different codes.
- Rule 10b5-1 plan
- A trading schedule set in advance, so the trades execute automatically at pre-set dates. It exists to shield an insider from insider-trading claims, which also means a trade under such a plan carries little signal about what they think today.
- Transaction date vs filing date
- The day the trade happened, and the day it was declared. They differ, sometimes by weeks. Returns are measured from the transaction date; the feed is ordered by filing date, because that is when the information became public.
- Director, officer, 10% owner
- The three categories that must declare. A director sits on the board, an officer runs part of the company, a 10% owner holds enough capital to be presumed informed. Roles are shown as declared by the regulator.
- Filing deadlines
- How long an insider may wait before declaring, and it varies by regulator: two business days in the US, three in the EU. A late filing is visible here as a wide gap between transaction and filing date.
Sources: SEC EDGAR (US), AMF (FR), BaFin (DE), AFM (NL) and FI (SE). Amendments supersede original filings. Values are converted to USD at the transaction-date FX rate.