SEC 13F Scraper — Fund Holdings & Quarterly Changes

Every manager running $100m+ files a quarterly 13F listing its US equity positions. Dumping that table is easy, and it is what every 13F scraper does. It is also wrong three ways.

The three traps

1. The values changed units in 2023 — silently, by a factor of 1000.

SEC amended Form 13F for filings made from 3 January 2023: values are whole dollars. Before that they were thousands. Berkshire Hathaway's own filings:

Filed	tableValueTotal	Actually
2022-11-14	296,096,640	$296.1 billion
2026-05-15	263,095,703,570	$263.1 billion

Read both literally and the 2022 portfolio comes out as $296 million. This Actor normalises everything to dollars and labels which convention each filing used.

2. One holding is filed as many rows.

The same CUSIP repeats once per internal manager group. Berkshire's Q1 2026 filing lists ALLY FINL INC three separate times. Row-dumping reports three positions where there is one — and every weight, rank and count derived from it is wrong.

Verified on the same filing: 90 rows collapse to 29 real positions, 18 of which were split. This Actor aggregates by CUSIP and tells you how many rows were merged, so the arithmetic is auditable.

3. Amendments replace a quarter; they do not add one.

A 13F-HR/A restates a period already filed. Treat it as another quarter and you double-count; ignore it and you are reading data the filer has withdrawn. This Actor keeps the latest filing per period and names the accession it supersedes.

What nobody sells: the diff

A list of positions is a snapshot. What moved is the information. Every quarter is compared against the one before it and each position is labelled:

Action	Meaning
NEW	Position opened this quarter
EXITED	Position closed entirely
ADDED	Share count up by more than the threshold
TRIMMED	Share count down by more than the threshold
HELD	No meaningful change

Decided on share count, never dollar value. If a stock halves and the manager does nothing, the dollar value halves — that is the market moving, not a trade. Value change is still reported, just never used to classify.

Real output, Berkshire Q1 2026 vs Q4 2025:

Action	Holding	Shares before → after	Δ
NEW	Delta Air Lines	0 → 39,809,456	+$2,646.5M
ADDED	Alphabet Class C	17,846,142 → 54,249,798	+204%
TRIMMED	Constellation Brands	13,000,000 → 632,890	−95.1%
TRIMMED	Nucor Corp	6,407,749 → 3,907,075	−39%

Positions are matched on CUSIP, not issuer name — filers type names by hand and they drift. Share classes are tracked separately, which is why Alphabet can legitimately appear as NEW in one class and ADDED in another.

Input

{
  "ciks": ["1067983", "1350694"],
  "quarters": 4,
  "changesOnly": true,
  "minPositionValue": 10000000
}

These are the manager's CIK, not the companies they hold. Berkshire Hathaway is 1067983, Bridgewater 1350694, Renaissance Technologies 1037389.

Field	Default	Notes
ciks	["1067983"]	Manager CIKs.
quarters	4	Newest first. One extra is read as the comparison baseline.
startPeriod / endPeriod	—	YYYY-MM-DD, quarter ends.
changesOnly	false	Skip the holdings table, return only what moved.
minPositionValue	—	Drop the long tail of tiny positions.
trimThresholdPercent	5	Share-count move before ADDED / TRIMMED fires.
includeAmendments	true	Prefer a 13F-HR/A over the original.
contactEmail	—	Goes in the User-Agent, nowhere else.

Output

portfolio — one row per quarter:

{
  "type": "portfolio",
  "managerName": "Berkshire Hathaway Inc",
  "periodOfReport": "2026-03-31",
  "positions": 29,
  "rowsFiled": 90,
  "positionsWithSplitRows": 18,
  "totalValue": 263095703570,
  "reportedTotalValue": 263095703570,
  "totalsAgree": true,
  "largestHolding": "APPLE INC",
  "largestHoldingPercent": 22.0,
  "concentrationHHI": 1185
}

totalsAgree cross-checks the sum of the rows against the total the manager reported on its own cover page. On the run above it reconciles exactly, which is the strongest available proof that the units and aggregation are right.

change — one row per position that moved. holding — one row per position per quarter, with weight, rank and implied price. quarter_changes — the per-quarter headline.

Reliability

• Official SEC endpoints only: data.sec.gov and www.sec.gov/Archives. No login, no session, no rendered-page scraping.
• The information table has an arbitrary numeric filename (53405.xml on one filing, 18337.xml on another). It is discovered from the filing's own index.json, never guessed.
• Runs under SEC's published 10 requests/second ceiling with a single serial pacer, and sends a declared User-Agent with a contact address — SEC returns 403 to anything without one.
• One unreadable quarter never fails the run; it lands in RUN_SUMMARY.

Limits

• 13F covers US-listed equities, ADRs, certain convertibles and options. It does not cover cash, bonds, foreign listings or short positions. A 13F is not the whole portfolio and never was.
• Filings are due 45 days after quarter end, so the data is six weeks stale on arrival. That is the form, not the Actor.
• Filings before roughly 2013 are plain text with no XML table. Those are skipped with a recorded reason.
• PRN rows are bond principal amounts, not share counts. They are kept separate and no per-share price is derived for them.
• This is public regulatory data returned as filed. It is not investment advice and no forecast is implied by any field.