What Form 4 AI is

The problem

Roughly 200–600 SEC Form 4 insider transactions are filed every trading day. Most are mechanical — option exercises, tax-withholding sales, scheduled grants, 10b5-1 plan executions — and carry no directional view from the insider. The few that are informative get drowned in the noise. Reading every filing by hand takes hours; skipping them means missing the signal.

What we do

Form 4 AI ingests every Form 4 from SEC EDGAR within ~60 seconds of publication, filters out the mechanical noise, and applies a two-layer analytical pass:

• Rule-based conviction score (0–100). A deterministic, transparent score derived from the filing's characteristics — role of the insider, transaction size relative to compensation, cluster behaviour, recent corporate context, and insider trading history. Documented end-to-end on the methodology page.
• LLM thesis. A short analyst-style narrative that grounds the score in the specific facts of the filing — what the insider did, why it might matter, and what could be wrong with the read. Generated by a large language model and constrained to return structured output (paragraph, conviction, key factors, risks).

The two layers run independently. We're letting the data decide which one is the better predictor over a 60–90 day window — see the public track record for forward-returns vs SPY across both methods.

What we don't do

• We don't claim alpha. We're not a hedge fund. We're a research and triage tool.
• We don't recommend trades. Every conviction score is a signal in a noisy domain, not a buy or sell call.
• We don't predict prices. The score reflects how informative the filing looks; the market still does what the market does.
• We don't hide behind a black box. The rule-based score is fully documented, the methodology version is pinned per-row in the database, and the public track record is rebuilt nightly from raw forward returns.

How we're different

We deliberately publish our track record before claiming it works. Every score we generate gets matched against forward returns vs SPY at 5 / 15 / 30 / 60 / 90 / 180-day horizons. The numbers are on /track-record — methodology, sample sizes, and the honest caveats included. When the data tells us a particular signal isn't holding up, we'll say so.

Plans

The free tier shows the last 30 days of filings, a weekly digest, and the conviction score. Paid plans add the full LLM thesis on every filing, watchlist alerts, and (on the higher tier) SMS delivery for cluster buys. See /pricing.

Contact

Questions, feedback, or "you got the methodology wrong": support@form4ai.ai.