Methodology

What this site is

Insider Atlas is a descriptive research site built on public SEC filings. Most insider-data sites show you transactions and let you draw your own conclusions; Insider Atlas shows you the transactions, what the academic literature has found about patterns like them, and where the two diverge. It never turns those records into a recommendation or a security-level outcome — it says “3 insiders bought $X of stock in the same week,” cites the papers that studied that pattern, and stops there.

What this site is NOT

This is the part you should read twice.

• Not investment advice. Nothing here is a recommendation to buy, sell, or hold any security. No page on this site tells you what to do with money.
• Not a prediction engine. There are no forecasts, no backtests, and no security scores. The literature pages rank empirical support in published research, not future performance.
• Not a trading tool. The data is at least a day old by construction (see the update lag below), and the academic evidence says most of what insiders file is uninformative anyway (Cohen, Malloy & Pomorski 2012).
• Not a substitute for the filings. Every row links to its EDGAR source; when in doubt, the filing is the truth and this site is a lossy summary of it.

If a cluster event on this site looks exciting, the correct next step is to read the underlying filings and the cited papers — not to trade.

Data flow and provenance

Everything on this site comes from a daily batch pipeline against SEC EDGAR — no real-time feed, no vendor data, no scraping of other aggregators. The provenance chain for any number you see is:

SEC accession number → EDGAR archive URL → normalized JSON in the repository → this page.

1. 6pm ET, weekdays — a scheduled job fetches the day’s Form 4 index from EDGAR (master.YYYYMMDD.idx under the daily index).
2. For each Form 4 / 4A accession, the pipeline fetches the primary ownership XML document from the EDGAR archives.
3. The XML is parsed (Table I non-derivative + Table II derivative rows, footnotes, relationship flags) and normalized into a versioned JSON schema. Every record keeps its accession number and a resolvable sec.gov archive URL.
4. Normalized filings are committed to the repository as data/form4/YYYY-MM-DD.json — an append-only public audit trail.
5. Company metadata (ingest/company_metadata.py) enriches each issuer with its SIC code, SIC description, sector mapping, and exchange from data.sec.gov/submissions/ (cached 7 days).
6. Cluster detection (ingest/cluster_detector.py) runs over the ingested window, writing one file per ISO week to data/clusters/.
7. Aggregation (ingest/aggregator.py) produces the per-company, per-insider, per-sector, and site-wide JSON the pages are built from.
8. The static site is rebuilt from the committed JSON. What you see is exactly what is in the repository.

The pipeline respects SEC fair-access rules: at most 10 requests per second (we run at 9 with backoff on 429/503) and a User-Agent that identifies the project and a contact address.

If you find a row you can’t trace to a filing, that’s a bug — please report it.

Update lag

Timeliness is bounded in two places, and we disclose both rather than pretend otherwise:

• The SEC’s side: Section 16 requires a Form 4 to be filed within 2 business days of the transaction. A trade executed Monday can legally surface Wednesday evening.
• Our side: we re-ingest once per day (~6pm ET, weekdays) and rebuild the site from the committed data. Every page carries a “Data as of” stamp; there is no real-time feed and no LIVE indicator anywhere, because neither would be honest.

In the worst case a transaction appears here roughly three business days after it happened. Amended filings (Form 4/A) can also restate transactions from much earlier dates — cluster files for past weeks may therefore change when amendments arrive.

Cluster definition (v0.2)

A cluster event is: 3 or more distinct insiders at the same issuer making same-direction trades (all acquisitions or all dispositions) within the same ISO calendar week, with total reported value of at least $100,000.

The detection algorithm, in the order it runs:

1. Every transaction is bucketed by (ISO week of trade date, issuer CIK).
2. Routine transactions are filtered out first — codes A, D, F, I, M (grants, dispositions to the issuer, tax withholding, discretionary plan moves, option exercises under Rule 16b-3). Detection runs on the opportunistic remainder.
3. Buckets with mixed direction after the routine filter (some insiders acquiring, some disposing) are rejected outright.
4. Buckets with 3+ distinct insider CIKs and ≥ $100k total reported value become cluster events.
5. A bucket whose activity is entirely routine is still evaluated over the routine transactions, and — if it qualifies — surfaces as a routine-only window (routine = true). These are excluded from every headline cluster count and shown only behind a drill-down, because the literature finds them uninformative.

Notes on the construction:

• The 3-insider / same-week construction follows Lakonishok & Lee (2001), whose “strong buy” definition required agreement across multiple insiders.
• The routine-code exclusion is a lightweight proxy for the Cohen, Malloy & Pomorski (2012) finding that more than half of Form 4 activity is routine and carries essentially no documented return association. (Their original classifier uses 5-year per-insider trading histories; the code-based proxy is the locked v1 compromise.)
• Concentration tiers on cluster cards (WATCH 3–4 insiders, ELEVATED 5–7, HIGH 8+) describe the insider count only. They are labels, not scores.
• Joint filings: one Form 4 can be filed jointly by several related reporting owners (a fund, its adviser, its general partner). The shared transactions are counted once in event dollar totals, but each co-filer counts as a member — so a single joint filing can technically satisfy the 3-insider threshold with related entities. The “source filings” count on each cluster page shows how many independent filings back the event; treat 1-filing clusters skeptically.
• Implausible values: occasionally a filer enters the aggregate dollar amount in both the shares and price fields, producing trillion-dollar rows. Any single transaction over $10B is flagged, shown with its provenance, and excluded from all dollar totals.
• Rule 10b5-1 plan trades are flagged, not excluded — you can filter them yourself.

There is no SEC definition of a “cluster”; other sites use windows from 7 to 90 days. Ours is deliberately the tight end, and the definition is versioned — if it changes, the version stamps on cluster pages change with it.

Routine vs opportunistic

The single most important split on this site, exposed everywhere rather than buried:

All dollar aggregates on this site (sector heatmap, KPI chips, company/insider flow columns) use opportunistic transactions only, so a wave of January option grants doesn’t masquerade as conviction buying.

Heatmap metric

Each sector cell’s fill encodes net open-market flow ÷ gross flow over the trailing 7 days, using non-routine codes only: (buys − sells) / (buys + sells), clamped to [−1, +1]. Equal cell sizes are deliberate — the grid compares direction and intensity across sectors, not market size. Color is never the only encoding: every cell carries a signed mono caption.

Dates

Form 4 has at least four dates. We keep them separate and you should too:

13F institutional holdings (Phase 3)

Quarterly Form 13F-HR reports, ingested per quarter from EDGAR (ingest/13f_quarterly.py). Every 13F-HR filed in the 45-day window after quarter-end is scanned (cover page + summary), and the top 50 managers by reported 13(f) value get their full information table ingested, plus their previous-quarter report for change classification. Positions are matched across quarters by (CUSIP, put/call flag, shares-vs-principal type) after summing rows split by investment discretion; each position is classified NEW / INCREASED / DECREASED / UNCHANGED / CLOSED. When no previous-quarter report is matched, changes show as no data — never as “no change”.

13F cluster events are the Form 4 analog with the manager as the entity and the quarter as the window: 3+ tracked managers reporting the same position as NEW or INCREASED in the same quarter, over open equity positions only (no puts/calls, no principal-amount rows) with a $1M combined added-value noise floor (Form 4 uses $100k; 13F positions run an order larger). The framing is always “N managers added X” — with a 45-day lag and quarter-end snapshots, a shared add is routinely index membership or style flow (Lewellen 2011), not coordinated conviction.

Standing 13F caveats (rendered on every holdings page, not just here): 45-day filing lag; long-only US-listed 13(f) securities — no shorts, no international, options too partial to interpret; positions are quarter-end snapshots that miss intra-quarter round trips; only managers over $100M in 13(f) securities file, and positions under 10,000 shares and $200,000 may be omitted. Sector labels on manager pages are joined from the Form 4 pipeline’s issuer metadata, so their coverage is partial by construction.

What 50 years of literature supports — the full synthesis

This is the complete signal-support synthesis the site is built from, reproduced in full rather than distilled. Every rank below describes empirical support in published research — the strength and durability of documented historical associations — never a forecast, score, or recommendation. The papers themselves are annotated on the literature pages.

Empirical support ladder

What the literature treats as noise

For return association in published research, the following are noise, and this site never headlines them:

• Isolated small-dollar Form 4 sales after large option exercises.
• Routine annual diversification sales.
• A single large-shareholder secondary sale without officer participation.
• Dollar-weighted measures dominated by one block trade.
• Equal-weighted small-cap abnormal returns quoted without liquidity and cost adjustment.
• 13F ownership-percentage changes that simply track index reconstitution.

The ten principles this site is built on

The synthesis distills into ten commitments, all of them visible in the UI:

1. Buy ≠ sell asymmetry is a first-class UI fact, not a footnote.
2. Size stratification context accompanies any return-association claim quoted from the literature.
3. Routine vs opportunistic flags appear wherever the data allow.
4. Cluster definition transparency: N insiders, window length, and included roles are stated and versioned (v0.2, above).
5. Form-type literacy: Form 4 is an executed trade, Form 144 is an intent, 13F is a quarterly snapshot — never conflated.
6. Horizon honesty: announcement-window reactions (Brochet 2010), multi-month drift, and implementable-after-costs results (Seyhun 1986; Jeng, Metrick & Zeckhauser 2003) are different findings and cited as such.
7. No survivorship shortcuts: quoted results come from delisting-aware samples, reported both equal- and value-weighted where the papers do.
8. No productized backtest hero charts: the site’s job is context for public data, not selling a strategy.
9. The ladder of papers is citable end-to-end — a reader can move from Jaffe (1974) → Seyhun (1986) → Lakonishok & Lee (2001) → Jeng, Metrick & Zeckhauser (2003) → Cohen, Malloy & Pomorski (2012) without marketing mythology.
10. The dead-cat-bounce callout sits next to every sales visualization: sales often reflect liquidity and diversification, not a forecast of decline.

One-sentence takeaway

Across five decades of published research, insiders’ open-market purchases — especially non-routine, multi-insider clusters at smaller, high-asymmetry firms — carry the most durable documented associations; average sales and raw institutional headcount changes largely do not.

What we deliberately don’t do

• No forecasts, backtests, or security scores. The literature pages rank empirical support in published research, not company-level outcomes.
• No fake liveness. Every data view carries a “Data as of” stamp; there is no green LIVE dot anywhere.
• No sell-side alarmism. Average insider sales are not a reliable negative signal (Lakonishok & Lee 2001; Jeng, Metrick & Zeckhauser 2003) — sales visualizations on this site always carry that caveat, and sell-direction cluster pages carry it too.
• No non-public data. If it isn’t in an EDGAR filing, it isn’t here.

What makes this different