How this is made

The newest 10-K defines the rows. A line the company reported for years and has since dropped doesn’t appear at all — not as a blank row, not anywhere. Conversely, a line introduced recently shows a short history.

Remedy: build the row set from the union of all filings rather than one, and carry a "last reported in" marker per row. This is substantially harder because line order and labels conflict across filings — you’d need a merge strategy for rows that exist in some years and not others.

The alias rule requires at least two periods of exact agreement, zero disagreements, matching sign convention and time basis, and that the alternate concept isn’t itself a line on the statement. That’s strong evidence, but it’s still inference. Two genuinely different concepts that happen to coincide for one company can slip through — an issuance line and an employee-stock-plan line that always matched, say.

Remedy: a reviewed exception list, so a human can veto a specific pair. Raising the agreement threshold also helps, at the cost of missing real renames that had short overlap windows.

Each concept independently takes its most recently filed value. If a later filing restated one line but didn’t carry another, one column ends up mixing vintages, and the internal articulation breaks. Observed in Berkshire Hathaway (two years) and Walmart (one year), with discrepancies around 0.05%.

Remedy: treat a statement as a coherent artifact — for each period, pick the single most authoritative filing and take the whole column from it, falling back per-concept only where that filing is silent.

The model anchors on a 10-K. Foreign private issuers file 20-F (or 40-F for Canadian issuers) and are skipped entirely — ASML, HSBC, Alibaba and ARM among them. Many also report under IFRS rather than US GAAP.

Remedy: extend the form filter and handle IFRS taxonomies. Non-trivial: 20-F statement structures differ, and IFRS concept names are a separate vocabulary.

Companies don’t tag a Q4 duration — the fourth quarter only exists inside the annual figure. Flow items (revenue, expenses, cash flows) show Q1, Q2, Q3 and FY. Balance-sheet lines are unaffected, since every quarter end has its own snapshot.

Remedy: derive Q4 as FY minus the first three quarters. Correct for additive items, wrong for per-share figures, averages and ratios — so it needs a per-concept rule about what may be subtracted.

A 10-Q reports both the quarter and the cumulative year to date. Keeping both would put two different numbers under one column heading, so only the single quarter is kept.

Remedy: add the duration to the column identity and show YTD as its own set of columns. Roughly doubles the width of the quarterly output.

Where a figure was later restated, the newest version wins. You cannot see what was originally reported — which is precisely what a restatement study would want.

Remedy: keep all vintages with the filing that reported each, and let the consumer choose "as originally filed" or "as most recently restated". The data to do this is already collected; only the collapse discards it.

Only reports the SEC classifies as financial statements are included. The notes, where segment detail, tax reconciliations, debt schedules and lease maturities live, are not.

Remedy: include the note reports too. They’re available and structured the same way; the cost is a much larger and messier output, since note tables are irregular.

A product or geography split attaches to a line only if that statement declares the matching dimension. This prevents a concept from dragging every breakdown it has anywhere into every statement — but a filer who reports a breakdown without declaring the axis on that statement will have it omitted.

Remedy: fall back to attaching breakdowns by concept where no axis is declared, flagged as lower confidence.

Header rows carry no arithmetic, so their nesting level is guessed from the rows beneath them. It’s usually right and occasionally puts a heading one level off.

Remedy: nothing available in this dataset — the indentation genuinely isn’t published. Parsing the rendered filing itself would settle it.

If no arithmetic network matches a report well enough, the whole statement renders without indentation. Equity statements are the usual casualty.

Remedy: fall back to a presentation-order heuristic, or hand-build hierarchies for the common statement types.

The concept-change log tracks a line by its label, since the concept is the thing that moved. A company that reworded a label and retagged in the same filing appears as one line ending and another starting, rather than as a change.

Remedy: cross-reference the value-agreement evidence, which catches renames the label-based match misses. Both signals exist; they’re just not currently merged.

Everything keys on CIK. Mergers, redomiciliations and re-listings create new CIKs, so history simply starts there — Broadcom begins in 2015 because the current entity was created by the 2018 restructuring, and its predecessor sits under a different CIK.

Remedy: a CIK-succession map, hand-curated or derived from EDGAR’s former-name records.

Companies occasionally report a figure at a non-reporting date — shares authorised on the day of a stock split, for instance. These are real disclosures and get their own column, labelled with the month to distinguish them from quarters. They interrupt the rhythm of an otherwise regular time series.

Remedy: filter them out, at the cost of dropping genuine data.

A fiscal year is labelled for the calendar year it ends in. That matches how Apple, Microsoft and NVIDIA name their own fiscal years — but some retailers with January and February year ends call it the previous year. The label is internally consistent; it may not match a given company’s marketing.

Remedy: a per-company override, or simply show the period end date alongside.

Dollars, share counts and per-share amounts sit in the same column, distinguished only by row. Summing a column is meaningless.

Remedy: the unit is carried per row and could drive separate sections or sheets.

Everything comes from the SEC’s quarterly Financial Statement Data Sets, which lag filings by roughly a couple of months, and from whichever archive files are present on disk. A missing quarterly file leaves a hole with no warning.

Remedy: validate archive continuity on load and warn on gaps.

Companies can invent their own concepts when the standard taxonomy doesn’t fit. These work fine within a company and are meaningless across companies. Roughly a fifth of the concept changes observed involve one.

Remedy: nothing to fix, really — but they’re flagged in the change log so you can see where a company left the standard vocabulary.