PIPELINE.md

# How this dataset is built

Every step is in [`recipe/`](recipe/), published at the same revision as the
data. This document records what the steps do and what went wrong on the way.

## The idea

A security master is usually bought. The free alternatives are snapshots —
`company_tickers.json` states today's CIK-to-ticker map and carries no past, so
a company that changed symbol is listed under the new one for its entire life,
and every backtest built on it silently reads one company's prices against
another's fundamentals.

The history exists, though, scattered across filings nobody reads for this
purpose. Four million Form 4 submissions each print the issuer's trading symbol
on a known date. Every XBRL filing header carries the company's name, its former
name and the date it changed. Settlement records tie a CUSIP to a ticker twice
a month since 2004. Collected with their dates, those are a security master.

So this dataset has almost no source of its own. It is assembled from the
sibling datasets and from three SEC files, and it publishes the evidence
alongside the conclusions.

## The ticker box

[`symbols.py`](recipe/symbols.py) exists because 2.3% of Form 4 filings do not
put a ticker in the ticker box. They put an exchange prefix (`NYSE: KRC`),
parentheses (`(SIRI)`), two share classes at once (`ISCA, ISCB`, `JWA/JWB`,
`CRDA CRDB`), a symbol spelled with spaces (`N O G`), a venue note
(`NWIN(OB)`), a CIK, or a word meaning absence (`NONE`, `N/A`). Read literally
they become 3 428 tickers that never traded, and Hubbell, Wiley and
Brown-Forman lose their real symbols.

The parser reaches 98.97% of four million filings; the remainder is genuine
absence. Two cases in it are worth keeping in mind, because both were bugs
first:

* `NYSE/TRN` splits into two chunks and the first is a venue, not a
  four-letter ticker — but `CBOE` alone is Cboe Global Markets, whose ticker is
  its own name. A venue name is discarded only when some other chunk parsed.
* `NWIN(OB)` lost its symbol because brackets were stripped before the trailing
  note was read, turning it into `NWIN(OB` and matching nothing.

## Observations, then spans

Everything collected is a dated statement: *this filer stated this identifier
on this day*. Ten million of them. The spans in `identifiers` are a groupby over
that, and nothing more — first observation, last observation, how many, from
which sources.

The restraint is deliberate. It would be easy to emit `valid_to` as the day
before the next ticker appeared, which reads better and claims something the
evidence does not support: nobody publishes the day a ticker changed hands, and
a filing gap is a filing gap, not a delisting. `observations` ships so that
anyone who wants stronger intervals can derive them from the same evidence.

## What "current" means

The first definition was "the newest value this filer stated". It made a
company that stopped filing in 2014 permanently current under its 2014 ticker,
and when another company later took that ticker over, both were current at
once — 1 218 tickers, 6% of the live universe, each pointing at two companies.

Now `is_current` also requires the filer to still be filing, and 54 remain.
Those are genuine: dual filers and reorganisations mid-transfer.
`is_latest_for_filer` keeps the old meaning, because "what did this company last
call itself" is a real question about dormant companies, and keying issuer
attributes on `is_current` would leave every dormant filer nameless.

## The CUSIP bridge

13F and N-PORT name a holding by CUSIP; everything else names a company by CIK.
The composition is: settlement records give CUSIP to ticker on a date, filings
give ticker to filer over a span, and the join is on the ticker with the dates
required to overlap. Overlap is what separates a confident match from a
plausible one, and it matters precisely for the reused tickers this dataset
exists to disentangle.

Two things were learned by measuring rather than assuming.

**The count is the wrong measure.** By raw CUSIP count the bridge looked 37%
complete, which reads like a failure. Against actual 13F holdings the same
bridge covers 83.1% of holding rows and 87.5% of position value. Two thirds
of the CUSIPs ever traded are one-off warrants, foreign lines and delisted
shells that no strategy touches. The build now gates on match rate weighted
by settlement traffic, excluding funds.

**Most funds cannot resolve, and that is correct.** `IVV`, `IWM` and `VOO`
settle like shares, but they are lines inside a registrant trust's filings and
file nothing under a ticker of their own. Counting them as misses is what made
the bridge look broken; they carry `looks_like_fund` and sit outside the gated
denominator. *Never* was too strong, though, and both documents said it: a fund
that is its own trust does file, and `SPY`, `GLD`, `DIA` and `MDY` resolve to
the trust that issues them. 749 fund rows carry a CIK.

**The first six characters are the issuer.** A CUSIP's prefix identifies the
company, the suffix the instrument, so a class B share or a preferred of a
company whose common stock already resolved belongs to the same filer. Saying
so is a fact about how CUSIPs are assigned, not a guess — but the instrument
itself was never matched, only its issuer, so it carries its own method and a
middle confidence. It lifted coverage from 37% to 69% of pairs.

**Some of that lift was wrong, and the coverage gate rewarded it.** The prefix
rule borrowed a filer from *any* match, and a fund family is one prefix: when
iShares' `ICF` collided with ICF International's ticker, all 74 iShares funds
under `464287` became ICF International — IVV, IWM and EFA among them — and 14
Vanguard funds became a Van Kampen muni trust the same way. Checked against
the filer's own names, 8.6% of the 7 653 inherited fund rows were right. The
weighted-match gate counts any CIK, right or wrong, so it read the damage as
progress. Now only a date-confirmed, non-fund match may lend its issuer, a
prefix that points at two filers lends nothing, and funds never inherit;
inherited rows that agree on a name went from 84% to 96%.

**A fund is not the company that shares its ticker.** SPDR's `MDY` matched
Midway Gold by ticker and date, PIMCO's `CORP` a Michigan bank. Names are still
never used to attach a CUSIP; they are used to refuse one — a fund whose name
shares no distinctive word with anything the filer was ever called gets
`name_conflict` and no CIK, and the next candidate is considered. MDY now
resolves to the SPDR MidCap 400 trust. `CORP` still does not: "PIMCO
INVESTMENT GRADE CORP BD" carries no word the fund pattern knows.

**One symbol, three spellings.** Settlement records write `BRKB`, filings
`BRK.B` or `BRK-B`. The join used the literal string, so multi-class companies
matched only where a filer happened to type the record's spelling. Tickers
are now joined with the class separator removed; the published `ticker` is
still the one the record stated.

**The same inputs have to give the same bridge.** When two filers have each
stated a symbol exactly once, they tie on every key the candidates are ranked
by, and the winner was whichever row the join emitted first: feeding the
published pairs back through the builder moved 36 rows to a different filer.
The CIK now breaks the last tie — arbitrary, but the same every run, which is
what lets anyone rebuild `cusip_map` from the published tables and get the
file that was published.

## Verification

[`quality.py`](recipe/quality.py) gates publication. The check that earns its
place is ambiguity: a ticker current for two filers means somebody's history
has been attached to the wrong company, and nothing else in the pipeline would
notice. It caught the `is_current` definition above.

The other floors — ticker coverage among active filers, weighted CUSIP match,
issuer count — are set below measured values so a regression fails and noise
does not. All of them were set *after* measuring; the first attempt used
guessed thresholds, and both guesses were wrong in opposite directions.

## Schedule

A Hugging Face Job runs [`jobs/run.py`](jobs/run.py) weekly on Monday at 07:40
UTC, after the sibling datasets have refreshed: restore, lint, test, rebuild,
verify, publish. Weekly rather than daily because the inputs are weekly at
best — settlement files land twice a month and the sibling datasets are the
things that move.

## What is not done

* **No history before Form 4.** Ticker evidence effectively starts in 2006.
  EDGAR does not carry what a company traded as in 1998, so neither does this.
* **No exchange history worth the name.** Filers rarely name the venue and SEC's
  current map covers only live listings, so 8 228 spans is all the evidence
  there is.
* **No ISIN, no LEI, no FIGI.** 13F carries a FIGI column and N-PORT carries
  ISIN; neither is ingested yet. The schema has room for them.
* **No fuzzy name matching.** A CUSIP could often be attached to a filer by
  comparing issuer names, and a name match that is wrong is indistinguishable
  from one that is right. Unmatched is more useful than plausibly wrong.