PIPELINE.md

# How this dataset is built

Every step is in [`recipe/`](recipe/), at the same revision as the data.

## What could not be built, and why it matters

The first job was finding out what is actually reachable. The answer is
narrower than the name "corporate actions" suggests, and worth stating plainly
because everything else follows from it.

SEC's Financial Statement Data Sets carry `num.txt` — **numeric facts only**.
Dividend dates are typed as dates in XBRL, so they are in the filings and not
in the bulk data. The XBRL API is not a way around it: a request for
`DividendsPayableDateOfRecordDayMonthAndYear` returns 404, because that
endpoint serves unit-bearing facts and a date has no unit.

So there are no ex-dates, no record dates, no pay dates here. There is no free
structured source for them. What remains is still worth having: per-share
amounts by fiscal period, and split ratios.

## Dividends

Six tags carry per-share dividends, and they mean different things.
`CommonStockDividendsPerShareDeclared` is a decision; `…CashPaid` is a cash
movement; the preferred variants are a claim that ranks ahead of the common
holder. They are kept apart rather than coalesced into one number.

Facts from both the consolidated and the dimensional tables are read. The
dimensional ones matter: a dividend declared after the period closed is tagged
`SubsequentEventType=SubsequentEvent`, which makes it the one dividend fact
that is not history — the declaration is already public when the filing is.

`period_start` is derived from `period_end` and the fact's own `quarters`. A
per-share dividend dated to an instant is a filer error and is dropped rather
than assigned a day it did not cover.

## Splits: the trace, not the event

There is no free feed of US stock splits. There is, in a dataset that never
overwrites what a filing said, the trace a split leaves: it forces the company
to restate every earlier per-share figure by the ratio. Two filings covering
the same quarter, one before the split and one after, differ by exactly that
factor.

This is stronger than the signal most people reach for. A jump in shares
outstanding looks identical for a two-for-one split and for an equity raise
that doubled the count. Only a split reaches back and rewrites the past.

Three sources are combined and each is labelled:

| Method | Filers | Standing |
|---|---|---|
| `xbrl_tag` | 514 rows | the filer tagged the conversion ratio — authority |
| `eps_restatement` | 5 105 rows | inferred from restated per-share figures |
| `share_count` | — | never used alone; cannot tell a split from an issuance |

Validated against the tagged ratios: **73% of the companies with a tagged ratio
also have the same ratio inferred**, and the inferred method reaches seven
times as many companies.

### Four things this got wrong first

**The ratio list was handwritten and incomplete.** It had 2:1 through 20:1 and
a few m:n forms, and no 6:1 — so Deckers' 2024 split, the very case this
dataset exists to fix, was invisible. The evidence was all there; the ratio
simply was not in the table.

**Generating the list instead was worse.** Every simple fraction with small
numerator and denominator gives 144 ratios, thirteen of whose tolerance bands
overlap, and an ordinary 1.83× restatement then resolved to "eleven-for-six".
The list is written out again, completed, and nothing between 0.85 and 1.18 is
admitted at all: a five percent stock dividend and a five percent restatement
leave the same trace.

**One restated period is not evidence.** A single figure that happens to land
on a clean ratio produced a four-for-one Tesla split in 2020 that never
happened. Two is the floor; four or more earns `medium`; the count grades the
claim rather than gating it, because a real split restates every prior period a
filing shows — Apple's 2020 split left 36 of them.

**Extreme ratios need more.** Earnings of minus two cents restated to minus
forty dollars is a genuine one-for-a-thousand consolidation, and it is also
what a rounding change looks like on a company whose EPS never left the third
decimal. Ratios past 50× or under 0.02 now require the stronger evidence
threshold.

### The detection window

A split gets a window, never a date. `detected_after` is the acceptance of the
last filing that still used the old figures; `detected_before` the first that
used the new. Deckers lands in 2024-08-01 → 2024-10-31, and the split was
2024-09-16.

Combining windows across restated periods takes their intersection, which is
tighter. The intersection can be empty, and when it is, that is not an
arithmetic slip — it means the evidence spans two events at the same ratio, a
company that split two-for-one twice. The union is used then, a window that
certainly contains them, rather than publishing one that ends before it starts.

## The adjustment factor

The table the rest of it is for. Walking the splits backwards from today gives,
for every span, the product of every split that happened after it: Apple is
28.0 before 2014, 4.0 between, 1.0 now. Multiply an as-filed share count by it,
or divide an as-filed EPS, and the figure lines up with a split-adjusted price
series.

The gate checks that the newest span of every filer has a factor of exactly
one. Anything else means the walk started from the wrong end, and the whole
company's history would be off by a constant.

## Verification

[`quality.py`](recipe/quality.py) gates publication. Beyond the usual null and
range checks: every published ratio must be one companies actually declare
(1.83 is a restatement that slipped through, not a split), detection windows
must not end before they start, and the agreement between the tagged and
inferred methods must stay above 65% — a drop means the restatement signal has
started picking up something that is not a split.

## Schedule

A Hugging Face Job runs weekly on Monday at 08:10 UTC, half an hour after the
fundamentals rebuild it reads. Nothing else is fetched: this dataset has no
source of its own.

## What is not done

* **No dates for dividends.** See the top of this document.
* **No splits before 2009.** The restatement trace needs XBRL.
* **No separation of special from regular dividends.** The filings usually do
  not distinguish them either.
* **No stock dividends under 18%.** Indistinguishable from a restatement.
* **`share_count` is computed but not published as a method.** It corroborates;
  it cannot stand alone.