Who is selling your stock at the top — and how often they've done it before.
A former NYSE Vice Chairman joined a quantum-computing company's board in 2021. In September 2025 he sold 15%, then 41%, then 60%, then 75% of his stake — in four consecutive days, inside a 12% price band, five days before the stock rolled over.
Every number in that sentence is from public SEC filings. Nobody reads them. This tool does.
Not investment advice. Not an accusation — everything here is legal and disclosed. All data is public SEC filings. We just read them. Someone should.
npx uncle-watch rate QUBT
or from source:
git clone https://github.com/PeachMo16/uncle-watch && cd uncle-watch
node uncle.mjs rate QUBT
No dependencies. No API keys. Node 18+ and the SEC's public EDGAR API. A weekly leaderboard of the tickers in WATCHLIST is rebuilt by GitHub Actions every Monday — send a PR to add a ticker.
The parser, amendment handling, rolling-window clustering, and score dimensions are covered by synthetic fixtures, so the tests never need the network:
npm test
Downloads the most recent Form 4s (EDGAR's recent list, capped at 200; coverage is printed rather than assumed), classifies each sell's plan status three ways — 10b5-1 indicated / no 10b5-1 indication / unknown (the SEC's mandatory checkbox only applies to filings from April 2023; before that, a missing flag proves nothing) — then scores six risk dimensions, each 0–100 with the evidence attached:
| dimension | what it smells for |
|---|---|
| No 10b5-1 indication | sells filed without the pre-scheduled-plan checkbox |
| Cluster density | different insiders reaching for the exit in the same week |
| Ladder acceleration | escalating %-of-stake sells day after day (whole run worth ≥ $25k — shape alone is not a signal) |
| Exit discipline | one insider unloading inside a tight price band |
| Dilution engine | shelf registrations and offering supplements |
| Shell history | name changes, fresh 10-12G registrations (reverse-merger tell) |
Open-market insider buying is reported as an independent counter-signal — shown in green, never scored. Real buying is positive information when present, but the literature gives no exchange rate against sell-side risk, and a token buy must not launder a boat. Absence of buying is not evidence of danger.
Composite = weighted average = the uncle rate. It also derives the uncle exit zone: the price band where insiders historically pulled the ripcord.
🐀 $QUBT · UNCLE RATE 83/100
No 10b5-1 indication 87 20 of 23 status-known sells filed with no 10b5-1 indication
Cluster density 100 4 strict 10-day windows with ≥2 insiders selling
Ladder acceleration 100 one insider escalated 53% → 100% of stake (~$573k) in 3 days
Exit discipline 74 5 sells inside $15.02–$16.88 (12% band)
Dilution engine 100 27 offering/shelf filings
Shell history 30 fresh 12(g) registration (a common reverse-merger route)
· Open-market insider buying: none observed (not scored)
uncle exit zone: $11.70–$15.62 (median $15.02) · last close $8.46
For calibration, the same rubric on a boring mega cap:
🐀 $GOOG · UNCLE RATE 17/100
No 10b5-1 indication 0 0 of 306 status-known sells filed with no 10b5-1 indication
Cluster density 0 no multi-insider sell clusters
Ladder acceleration 0 no escalating multi-day sell ladders worth ≥ $25,000
Exit discipline 0 no insider with ≥3 sells lacking a 10b5-1 indication
Dilution engine 100 30 offering/shelf filings (mostly bond 424B2s — see limitations)
Shell history 0 no name changes or fresh registrations on file
· Open-market insider buying: none observed (not scored)
Three hundred and six insider sells at Alphabet. Every single one under an indicated 10b5-1 plan. That's what a boring boat looks like.
Every reporting owner has a personal CIK. Feed it in and get every company
they've ever filed ownership forms on — their whole career of boats. who follows
the SEC's older submission-index files and downloads every supported ownership XML
by default, grouped by issuer CIK so ticker and name changes stay together and
two people with the same name stay apart. It covers what EDGAR still serves, not
every trade of a lifetime — and it prints what it couldn't get.
node uncle.mjs who 1215183
node uncle.mjs who 1215183 --as-of 2025-09-17 --recent-days 90
node uncle.mjs who 1215183 --limit 200 --json
The report compares the most recent 90 calendar days with the preceding 1,095 days at each issuer separately. It shows selling/buying days, sale months, plan-status counts and, when the record supports it, the recent median reported sale value per selling day divided by that person's baseline median. Multiple transaction rows on one day are combined so splitting an order does not create extra selling days. It does not pool prices or dollar baselines across companies.
The size ratio is withheld if the baseline has fewer than five selling days or less than 180 days between its first and last sale, if price/security/ownership basis is missing or mixed, if joint-owner filings obscure attribution, or if known coverage gaps affect the comparison period. Those minimums are display heuristics, not validated predictors. Values are reported units, without currency, inflation or stock-split normalization. A ratio is a descriptive change, not a probability of misconduct or a return signal. Calendar-month lists do not implement the Cohen/Malloy/Pomorski routine/opportunistic classifier.
--as-of excludes later filings before resolving amendments, so a later
correction cannot rewrite an earlier report. Transaction dates after the cutoff
are excluded too. This is a day-level reconstruction from currently retrievable
filings, not an archived intraday market-data snapshot. The default cutoff is
today in UTC. The JSON includes raw transaction timelines, accession numbers,
filing dates, acceptance timestamps, source links, excluded records and coverage.
For the dated Fagenson example, the observed issuers are:
FAGENSON ROBERT B (CIK 1215183) · as of 2025-09-17
57/57 supported XML filings loaded; no download failures or limit omissions
RWY RENT WAY INC 2003 → 2006 6 filings
DSS DOCUMENT SECURITY SYSTEMS 2004 → 2018 26 filings
TQ CASH TECHNOLOGIES 2007 1 filing
NHLD NATIONAL HOLDINGS 2012 → 2021 15 filings
QUBT Quantum Computing Inc. 2021 → 2025 6 filings
Five boats, all micro caps, one career. The only boat he ever cashed out on is the one that 40x'd during a hype wave. That list is a fingerprint.
The QUBT window contains five observed selling days and zero in its preceding baseline window, so the personal size ratio is unavailable, not infinity. The tool won't invent a prior routine to compare against — the list above is the comparison, and it's the reader's job to read it.
Reports save to data/people/<CIK>/history.json; an explicit cutoff saves to
history-YYYY-MM-DD.json. Use --data-dir <directory> to isolate a run. Name
lookup searches cached insiders and refuses ambiguous matches; a CIK is direct.
node uncle.mjs rate QUBT --history --limit all
node uncle.mjs actions QUBT --history --limit 500
who reads all advertised indexes and all supported XMLs unless --limit N
is supplied. rate/actions retain their recent-list, 200-XML default. --history
adds old indexes; --limit all removes the XML cap. The offering/shell dimensions
still use the recent index, preserving their existing definition. Changing the
ownership window can change the other dimensions; compare reports with the same
coverage, not just the same ticker.
Every fetch records loaded XMLs, unsupported non-XML filings, omitted filings, failed downloads, failed/missing historical indexes and stale-cache use. Older text-only documents remain unparsed. A failed history page never silently becomes an empty history. Overlapping index pages are deduplicated by accession. Network requests have timeouts and stay paced below SEC's published request limit.
Every insider transaction, newest first — plan sells labeled, no-indication sells
flagged, pre-2023 sells marked plan status ? instead of pretending to know.
Every ticker you've rated, sorted by uncle rate, with each boat's loudest dimension.
- The classic insider-trading literature (Lakonishok & Lee 2001; Finnerty 1976) found pooled insider sales carry little predictive signal — the informativeness was in purchases — and mimicking profits in small/OTC stocks were eaten by trading costs (Rozeff & Zaman 1988; Lin & Howe 1990). uncle-watch's sell-based dimensions are not independently validated as return predictors. The uncle rate is a reading aid that ranks which filings deserve your attention; it is not, and must never be read as, a tradeable signal.
- The 10b5-1 checkbox is only mandatory on filings from April 2023, so plan status is
three-valued; pre-2023 sells are
unknownand excluded from every scored dimension rather than counted as unscheduled — each dimension prints how many it left out. The unscored exit zone still includes them and prints its basis. A micro cap whose whole story is pre-2023 will therefore score low on clusters, ladders, and discipline: that is the tool refusing to guess, not a clean bill of health. Readuncle actions. - "No 10b5-1 indication" means exactly that — the filing didn't claim the affirmative defense. It is not Cohen/Malloy/Pomorski's behavioral routine/opportunistic classification (theirs is built from multi-year calendar regularity), and their results don't validate this filter.
- Foreign-issuer coverage varies by period and applicable exemption. From March 18, 2026, directors/officers of covered FPIs became subject to Section 16(a) reporting, with SEC exemptions for qualifying cases. A 20-F/40-F/6-K flag is only a coverage prompt, not a determination of an individual's legal obligation. Earlier histories may still be absent from EDGAR; local-regime reporting such as SEDI is not yet integrated. See the current SEC reporting overview.
- A Form 4/A supersedes only the original it names (via
dateOfOriginalSubmission), so a sibling Form 4 for the same period survives; an amendment that names no original date falls back to superseding its whole (issuer, owners, period) group. Multi-owner joint filings are still attributed to each owner (slight double-count). - The dilution dimension counts all offering paperwork — a mega cap's bond 424B2s score the same as a shell's equity ATM. Read the evidence line, not just the number.
- Submissions and prices are cached for 24h. Ownership XMLs are retained and reused locally; amendments have separate accessions. A local cache is not proof of what was publicly available at every historical instant.
whoanalyzes non-derivative transactions; derivative tables and standalone holdings are not converted into sales. Issuer history can include a holdings filing without a purchase/sale event. Known omissions are not proof of absence.- Price-context ratios for sales before the available daily-price feed are now unavailable; the tool never substitutes a later quote for an earlier sale.
- The weights and the $25k ladder floor are hand-tuned on a handful of anchors, not backtested. This is a reading tool, not a trading system.
v0.4 (local development) — historical submission indexes, explicit coverage, CIK-based person/issuer continuity, and an independent personal-history report. As-of filtering precedes amendment handling; 3/A and 5/A cannot erase a Form 4. Unsupported/missing files remain visible, and historical price context never borrows a future quote. FPI coverage wording updated for the 2026 rule change. The six scoring formulas and weights remain scoring version 3; personal behavior comparisons do not enter the composite.
v0.3 (2026-09-01) — three fixes from a second code read, each with a test:
- Form 4/A dedupe dropped innocent siblings: within one (issuer, owners, period) group, any amendment used to erase every original — a grant filed a week before a corrected sale went missing. Amendments now replace only the original whose filing date they name; an undated amendment keeps the old conservative behaviour.
- v0.2's "refuse to guess" about pre-2023 plan status only reached the ratio dimension; cluster density, ladder acceleration, and exit discipline still treated every unknown-status sell as unscheduled. All six scored dimensions now use only sells that affirmatively lack a 10b5-1 indication, and print the excluded count. The exit zone (unscored) keeps them, labeled.
- Ladder acceleration had no value floor: a 50% → 100% exit of a hundred-dollar stake scored 100, same as a seven-figure one. A run must now be worth ≥ $25k to count.
On the two README anchors nothing moved (QUBT 83, GOOG 17); the fixes bite on tickers with pre-2023 histories, corrected filings, or token-sized insiders.
Also in v0.3: the SEC User-Agent lost its parenthesised URL — sec.gov started answering 403 to it, which broke every fresh clone — and a failed EDGAR/Yahoo fetch now falls back to a stale local cache with a loud warning instead of crashing.
v0.2 (2026-08-30) — scoring semantics audited against the 1974–2003 insider-trading literature by gold-digger, a sibling project that digs up old research and points it at live systems. The audit found a mislabeled citation and two semantic overclaims; this release fixes them:
- Renamed "Opportunistic ratio" → "No 10b5-1 indication" and removed the CMP 2012 citation from the filter: their routine/opportunistic split is behavioral (calendar regularity across years), not plan-checkbox status, and never validated this filter.
- Plan status became three-valued around the April 2023 checkbox rule; pre-2023 "opportunistic counts skew high" is fixed by refusing to guess.
- Removed "Net insider flow" from the composite (an all-sell/zero-buy mega cap scored 100 on it — absence of buying is not evidence of danger). Open-market buying is now an unscored counter-signal. Remaining six weights renormalized.
Deferred until the full texts are read: a Kahle (2000)–style event feature (abnormal insider selling before an equity/convertible issue) and a real behavioral routine/opportunistic classifier worthy of the CMP name.
This project was built conversationally with an AI agent (Claude) from one person's observation that the same names kept selling the same stock at the same heights. Zero dependencies, plain Node, every score traceable to a filing. Fork it, point it at your ticker.
MIT