Schema

Three tables, joined on accession_number — the SEC's unique identifier for a single filing. This page is the same documentation that ships with the data, published in full so you can evaluate it before paying for anything.

filings (1 row per SEC form)
   │
   ├── accession_number ──▶ owners  (1+ rows: who filed it)
   │
   └── accession_number ──▶ trans   (0+ rows: what they transacted)

One filing can name several owners — co-filers and affiliated funds — and report several transactions together. That one-to-many shape is the source of the most common analytical mistake, covered at the bottom of this page.

filings

One row per form submitted to EDGAR. 1,533,585 rows.

owners

One row per person or entity named on a filing. 1,669,356 rows, 132,229 distinct insiders. Several rows per filing is normal, not a duplication bug.

trans

One row per transaction line, and the largest table. 2,317,601 rows.

Form types

Transaction codes

The distribution here is the single most important thing to understand before your first query.

The two mistakes

Group filings inflate dollar totals

When a private-equity or venture firm transacts, the fund, its general partner, its management company and named partners may each file the same transaction. Summing value across owners rows multiplies one purchase several times over. Deduplicate on (issuer_cik, trans_date, value) before aggregating.

This matters more than it sounds: the inflation is concentrated in exactly the companies where institutional holders are active, so it distorts comparisons between companies rather than adding uniform noise you could ignore.

Automatic transactions outnumber decisions five to one

Codes A, F and M together outnumber P by roughly five to one, and they are almost entirely mechanical — vesting schedules, tax withholding, option mechanics. Filtering to trans_code = 'P' is what isolates the transactions an insider actually chose to make.

Three more things worth knowing

• Identify entities by CIK, never by name. owner_name and issuer_name are free text and vary between filings for the same party; the CIK fields are stable.
• Filing date is when the market learned. trans_date is when the trade happened; filing_date is when it became public. Any analysis of what was knowable at a point in time keys on filing_date.
• Source errors are real and preserved. A handful of filings carry mispunched share counts or prices, producing transaction values in the billions. We parse faithfully rather than silently correcting, so filter implausible outliers before computing averages.