Form 4 AI
Home

Methodology overview

How we read every Form 4.

A high-level overview of what Form 4 AI ingests, what it filters out before scoring, what it evaluates on the filings that pass the filter, and how the conviction score should be interpreted. We publish this overview so users can decide whether the approach matches their workflow — exact factor weights, thresholds, and model details are not disclosed publicly.

1. What we ingest

We poll SEC EDGAR continuously and pick up every Form 4 (and Form 4/A amendment) within roughly a minute of publication. For each filing we capture the underlying transaction lines, the reporting insider's role, the issuer, and the timestamps. Everything downstream is grounded in this raw, public data.

2. What we filter out before scoring

The first job, before any scoring, is separating noise from signal. A large share of Form 4 filings carry no directional view from the insider — they're mechanical compensation events. Filings whose transactions are entirely mechanical are tagged at parse time and never reach the AI analysis layer. Examples:

  • Option exercises that don't involve a discretionary decision.
  • Tax-withholding sales paired with vesting events.
  • Scheduled grants and equity awards.
  • Bona fide gifts and intra-family transfers.
  • Conversions of derivative securities.

This step is the largest concrete win the system can point to: roughly 70% of incoming Form 4 filings are skipped here. The current breakdown is published on the calibration log.

3. What we evaluate

For filings that survive the mechanical filter, two analytical layers run independently and are shown side by side:

Rule-based score (deterministic)

A transparent score that combines a set of well-known features from the academic insider-trading literature. We don't disclose the exact weighting, but the broad categories are:

  • Transaction type. Open-market purchases by an insider using personal cash carry the most weight; routine compensation events carry the least.
  • Insider role. Senior officers (CEO, CFO, etc.) generally carry more weight than passive holders. We treat sole-officer purchases as higher-information than large-holder trims.
  • Transaction size. Larger open-market purchases receive more attention than symbolic purchases. Small token purchases are treated cautiously.
  • Cluster context. Multiple insiders buying the same ticker in a short window is one of the most-cited insider patterns and is treated as stronger context.
  • Insider history. Each insider's prior open-market buy track record vs the broader market, where enough history exists.
  • Surrounding context. Recent corporate disclosures and price action, used as confounder checks.

AI thesis (narrative)

A separate AI analysis layer produces a short analyst-style paragraph for each surviving filing — what the insider did, why it might matter, key factors, and explicit risks. The AI is constrained to ground its output in the filing and a curated context bundle, and to lower conviction when an input is missing rather than invent it. The AI does not see market prices in real time and is not making short-term price forecasts.

4. How to interpret the conviction score

Both layers output a 0–100 conviction score. The score is a relative confidence measure, not a price target or a return forecast. Higher scores indicate cleaner directional setups — open-market purchases by senior insiders, with transaction size, history, and surrounding context lining up. Lower scores indicate routine, ambiguous, or mechanical activity.

We deliberately do not publish exact bucket cutoffs. The point of the score is relative ranking within the day's flow, not a tradable threshold. The public calibration log shows how scores have related to forward returns at a bucket level, with sample sizes flagged honestly.

5. How calibration works

Every filing we score gets matched against forward returns vs SPY at multiple horizons. The methodology pins entry to the first regular-trading-session close strictly after the filing timestamp (no same-day lookahead). Calibration runs nightly and the results are published — including the calls that didn't work — on the calibration log. The page is honest about sample size and labels small buckets as calibrating rather than reporting them as if they were stable.

6. What the system does NOT do

  • Does not predict short-term price moves. The thesis is a directional view on the informational content of the filing, not a "this stock will go up next week" call.
  • Does not invent data. If the bundle lacks an input — for example, the insider has too little history to score reliably — the system says so and lowers conviction.
  • Does not chase narrative. We don't pull in social-media sentiment, options flow, or news-feed surface. We deliberately stay grounded in regulatory data and insider history.
  • Does not give buy/sell advice. The output is interpretation, not a recommendation.

7. Known limits

  • Insider-buy backtests are notoriously path-dependent. Even published academic findings on cluster buying weaken out of sample.
  • Insider sales are noisier than buys (10b5-1 plans, taxes, diversification). The buy side is the cleaner signal in this dataset.
  • The historical-edge calculation needs enough prior buys to be meaningful; insiders without that history are flagged as cold-start and scored conservatively.
  • Delisted tickers can drop out of the calibration table when pricing data is unavailable. We're transparent about this in the limitations section of the calibration log.
  • The system uses a third-party AI analysis layer for the narrative thesis. Different AI systems would produce different theses on the same bundle.

8. Disclaimer

This is research and journalism over public regulatory filings. Not investment advice. AI-generated theses are pattern recognition on disclosed data, not stock picks. Past insider performance doesn't guarantee anything. Insider-buy backtests are notoriously noisy. Do your own work. Verify every filing on SEC.gov.

See the calibration log →Browse the live feed →