1. The four inputs
Every thesis is grounded in four hard inputs from the bundle our system assembles for each filing. The model is instructed to never assert anything outside this bundle, and to lower conviction when an input is missing.
(a) Trade size in context
Raw dollar amount, but more importantly: relative to the insider's recent compensation, prior personal holdings, and any prior open-market activity by the same insider. A $50k buy by a line-officer earning $200k/yr is a different signal than a $50k buy by a CEO earning $5M/yr.
(b) Insider role
Form 4 reportingOwner role flags — Director, Officer (with officerTitle), 10% holder, Other. We weight these very differently:
- Sole-Officer / sole-Director buys carry the most signal: an individual making a personal cash decision based on their internal information.
- Isolated 10%-holder sales carry the least: large holders trim positions for portfolio, estate, or tax reasons unrelated to their view.
- CEO / CFO buys sized comparably to one year of cash compensation are the highest- quality single-filing setup we score.
(c) Insider's historical edge
For every insider with at least 3 prior open-market buys on record, we precompute their 90-day, 1-year, and 3-year forward edge vs SPY. An insider who has beaten the index 3 of their last 4 personal buys is materially more informative than a first-time filer or one with a poor track record. New insiders are flagged as low information rather than treated as average.
(d) Surrounding context
For each filing we pull recent 8-K disclosures within ~90 days, a price-action snapshot (21- and 252-day returns, 30-day ATR), and recent news headlines. The model is instructed to call out confounders explicitly — a buy that follows an Item 5.02 CEO appointment is plausibly a routine alignment purchase rather than a price-view trade.
2. Transaction-code semantics
We treat SEC transaction codes with very different weights. The model has explicit instructions for each:
- P — open- market or private purchase. Strongest single bullish signal: the insider used personal cash.
- S — open- market or private sale. Bearish only after controlling for diversification, tax-driven sales, 10b5-1 plans.
- A — grant or award under Rule 16b-3(d). Non-economic; downweighted.
- D — disposition to issuer (typically buyback-related). Generally non-informative for the insider's view.
- F — payment of exercise price or tax via withholding. Mechanical, not a discretionary sell. Downweighted.
- M + same-day S — option exercise and sell paired. We treat as a cashless liquidity event, not a fresh bearish signal.
- G — bona fide gift. Non-informative for price view. Downweighted.
3. Cluster detection
Cluster buying — three or more distinct insiders at the same issuer purchasing on the open market within 7 days — is one of the most reliable bullish signals in the public-disclosure universe. It's also surprisingly hard to detect without a rolling-window pipeline, which is why most retail tools miss it.
Our cluster detector fires the moment the third filing lands. Premium subscribers get an SMS alert within seconds. The thesis on each clustered filing carries a +cluster boost — typically 15- 25 conviction points — and routes through our higher-tier model.
4. Calibration of the conviction score
The conviction score is a 0–100 integer. Distribution targets:
- 5–25 — a typical filing: routine grant, small officer trim, isolated sell. Most filings live here.
- 26–55 — somewhat informative; usually a sole-officer buy of meaningful size with no offsetting context.
- 56–69 — high-quality but missing one element (small sample on edge, recent confounding 8-K, etc.).
- 70–89 — clean cluster buy by senior officers, or single-officer buy with positive historical edge against a beaten-down stock and no confounding 8-K.
- 90+ — rare; requires overwhelming evidence across all four input categories.
The model is explicitly instructed to pull conviction toward 50 when in doubt. Overconfidence is the failure mode we guard most aggressively, since a high-conviction call that's wrong destroys trust faster than a low-conviction call that's right.
5. What the model 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 — e.g., the insider has no scoring history because they have fewer than 3 prior open-market buys — the model says so explicitly 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 + insider history.
- Does not give buy/sell advice. The output is interpretation, not a recommendation.
6. Where we publish results
Every conviction-≥70 call gets tracked through 7-day, 30-day, and 90-day forward returns vs SPY on our public track record page. No selection. The calls that don't work are on the same page as the ones that do. With fewer than 30 settled high- conviction calls, treat aggregate numbers as directional only — we publish raw counts, not just averages, so you can apply your own significance bar.
7. Known limits
- Insider-buy backtests are notoriously path-dependent. Even published academic findings on cluster buying weaken out of sample.
- We don't currently model option-grant vesting schedules, which means some Code A grants could be incorrectly treated as uniformly non-informative.
- The historical edge calculation requires ≥3 prior open-market buys; insiders with fewer get scored without it. This is a cold-start problem we don't yet have a fix for.
- We use Anthropic Claude Sonnet 4.6 as the analyst. Different models would produce different theses on the same bundle. We're transparent about this and may A/B against other frontier models later.
8. Disclaimer
This is 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.