Methodology

Insider trades: SEC Form 4 filings from EDGAR. U.S. corporate insiders (officers, directors, 10%+ owners) must disclose trades in their own company's stock within two business days. We ingest every Form 4 daily. Consequence: what you see here can lag the actual trade by up to two business days — that lag is a legal reality, not a data error.

Market beliefs: Polymarket's public APIs. Once a company is linked to a related prediction market, we snapshot that market's probability once per day. The probability history starts the day the link is made.

The link between them: software suggests company↔market pairs by matching names and tickers against market questions, but every pair shown on this site was approved by a human. Nothing is auto-published.

Only open-market purchases and sales (SEC transaction codes P and S) count toward the score — trades where an insider voluntarily spent or received their own money at market prices. Everything else is stored and shown, but excluded from scoring as noise:

• Compensation mechanics — grants (A), option exercises (M), tax withholding (F): the employer's schedule, not the insider's conviction.
• Gifts (G), conversions (C), other codes — not market bets.
• Derivative-table rows — the reported "price" on these is an exercise price, not money at risk, so their dollar values would mislead.
• Rule 10b5-1 plan trades — trades executed under pre-arranged plans. We exclude a filing when its 10b5-1 checkbox is set or when its footnotes disclose a plan (the checkbox alone under-reports).
• Implausible values — single transactions over $1B are treated as probable filing typos (they occur!) and excluded from scoring.

For each approved company↔market pair, over a trailing 14-day window:

insider_signal = Σ value of open-market buys − Σ value of open-market sells
market_move    = probability today − probability 14 days ago
score = log₁₀(1 + |insider_signal|) × (1 + |market_move| × 5) × cluster_bonus

The cluster bonus (×1.5) applies when three or more distinct insiders traded the same direction in the window — several people acting alike is historically more informative than one. The log means a $10M signal doesn't drown a $1M signal; the market-move multiplier rewards pairs where the prediction market is actually moving. Pairs need at least $250K of net insider flow to appear on Trending.

Agree vs. diverge: the verdict compares what each side believes about the company, which depends on what a "yes" means for that question. For a bullish question ("Will X beat earnings?"), insiders buying while the probability rises = agree. For a bearish question ("Will X hit (LOW) $140?"), rising odds are bad news — insiders selling into that is agree: both negative, even though the raw directions look opposite. Opposite beliefs = diverge — the interesting case. Questions with no bullish/bearish reading (KPI ranges like "operating margin 34–36%?") are marked no direct read instead of being forced into a verdict. We label the pattern; we never claim to know why anyone traded.

What's a high score? The score is not out of 10 or 100 — it has no ceiling, and its only job is to rank. Because of the log, its first digit roughly equals the number of zeros in the insider dollars: $1M of net flow ≈ 6, $10M ≈ 7, $100M ≈ 8, before the market-move and cluster multipliers scale it up. The $250K entry floor makes ~5.4 the lowest score that can appear. In practice most entries land between 5 and 15; anything above 20 is exceptional — large money, a large market swing, and several insiders moving together at once.

• Form 4 lags trades by up to 2 business days; daily processing adds up to one more.
• A market's probability history begins when we link it — young links show "market flat / young" until 14 days accumulate.
• Insiders sell for many reasons (diversification, taxes, houses); buys are generally more informative than sells.
• Filings occasionally contain filer errors; we flag the implausible ones but cannot catch everything. Every number links to its source so you can check.
• This is a data-journalism project — not investment advice, and nothing here implies wrongdoing by anyone named.