# Tradevo Data — Methodology

*Point-in-time US equity fundamentals, built honestly from SEC EDGAR.*

A small, honest fundamentals dataset built entirely from free SEC EDGAR filings.
Every value is stamped with the date it *first became public*, so a backtest can only
ever use what was actually knowable at each point in time.

This document exists because, with data, **transparency is the product.** If you can't
see how it was built and validated, you can't trust it — so here's exactly how it's built.

---

## Problem 1 — Lookahead bias (the silent backtest killer)

A company's fiscal year has a *period end* (e.g. Apple's FY2024 ended **2024-09-28**) but
the numbers aren't *public* until the 10-K is filed weeks later (Apple's was filed
**2024-11-01**). A backtest that joins fundamentals on the period-end date is using
information that didn't exist yet — it's peeking into the future.

In this sample, on rows with a reliable filing date, fundamentals became public on
average **43 days after** the period ended (max 61). That's the future-peek a naive
join silently grants you on *every data point*. It makes strategies look better in
testing than they are in life.

**Live example, straight from the data:**

| As of date | Newest AAPL annual revenue you could *honestly* know | Filed |
|---|---|---|
| 2024-10-15 | FY2023 — $383.3B | 2023-11-03 |
| 2024-11-15 | FY2024 — $391.0B | 2024-11-01 |

Same company, one month apart, a different "latest known" number. This dataset encodes
that; a naive one hands you FY2024 too early.

## Problem 2 — Dirty / inconsistent raw data

EDGAR is free but messy: companies tag the same concept under different XBRL labels and
switch them over time (e.g. revenue under `Revenues` in older years, then
`RevenueFromContractWithCustomerExcludingAssessedTax`). A naive pull locks onto whichever
tag it sees first and can return a *stale year* or the *wrong number*. We resolve tags
to the most recent reporting and validate every row (see below).

---

## How it's built

1. **Source** — SEC EDGAR `companyfacts` API (`data.sec.gov`). Public domain, free, no key.
2. **Annual figures** — only `10-K` filings; for flow items (revenue, net income) only
   ~full-year durations (345–385 days) are kept, so quarters and stub periods can't leak in.
3. **Tag resolution** — among candidate XBRL tags for each concept, we choose the one whose
   data extends to the *most recent* period (ties broken by priority). This kills the
   stale-tag bug.
4. **Point-in-time stamping** — `first_filed` = the *earliest* filing that reported a given
   period. That is the first date the number was knowable. `lag_days` = first_filed − period_end.
5. **Restatement detection** — if a later filing revised a period's value by >0.5%, the row is
   marked `restated = True`, and we keep both `original_value` (first knowable) and
   `latest_value` (most recent).

   **`restated` means "the number changed," not "the accountants were wrong."** Three different
   things trip this flag and it is worth knowing which you are looking at:

   - **Retroactive split adjustment** (20 of the 189 rows here, all `DilutedShares`). When a
     company splits, EDGAR's current values are restated back through history, so Amazon's
     FY2020 diluted share count reads **510,000,000** as originally filed and **10,198,000,000**
     today — the 2022 20-for-1 split applied backwards. Same for AAPL (4:1), NVDA (4:1, then
     10:1), TSLA (5:1, then 3:1), WMT (3:1), NFLX (7:1, then 10:1), AVGO (10:1) and NKE (2:1).
     Nothing is wrong with either number; they are denominated in different shares. This is the
     single most dangerous one for a backtest, because pairing today's share count with a
     historical price understates market cap by exactly the split factor.
   - **Accounting-standard adoption.** Salesforce's FY2017–18 net income moves from $179.6M to
     $323M and $127.5M to $360M — that is ASC 606 adopted retrospectively, not an error.
   - **Genuine revisions and amendments**, including 10-K/A refilings.

   We deliberately do **not** collapse these into one "corrected" number. The point of the
   dataset is that `original_value` is what you could actually have known on `first_filed`, and
   `latest_value` is what a modern data vendor would hand you for the same period. The gap
   between them is the lookahead you are trying to measure.
6. **QA + reliability** — every row is checked (filing-lag range, value magnitude). Rows whose
   only available filing is a much-later one (common for the oldest years, where the original
   10-K predates XBRL) are marked `filed_reliable = False` rather than shipped with a
   misleading date.

## Columns

| column | meaning |
|---|---|
| `ticker`, `cik` | company identity |
| `concept` | Revenue / NetIncome / OperatingCashFlow / EPSDiluted / DilutedShares / Assets / StockholdersEquity |
| `xbrl_tag` | the exact SEC tag the value came from (full provenance) |
| `fiscal_year`, `period_end` | the period the value covers |
| `first_filed` | date it first became public (the point-in-time stamp) |
| `lag_days` | first_filed − period_end (the lookahead gap) |
| `filed_reliable` | True if the original 10-K is in XBRL (trust the date) |
| `original_value` | value as first reported |
| `latest_value` | value as most recently reported |
| `restated` | True if later revised >0.5% |
| `qa_status` | `clean` or the specific flag raised |

## This sample's coverage

- 40 large-cap US companies, 7 concepts (revenue, net income, operating cash flow, diluted EPS, diluted shares, assets, equity), up to 12 fiscal years each
- 3,280 point-in-time rows; revenue history depth averages 11.7 years
- 3,240/3,280 rows carry a reliable filing date (mean lag 43 days, max 61); 40 oldest-year/edge rows flagged for resolution
- 189 restatements detected (same-tag revisions >0.5%, including 10-K/A amendments). 20 of
  those are retroactive split adjustments to diluted share counts, not accounting errors — see
  "Restatement detection" above for why they are kept rather than collapsed

## What the adversarial audit caught (and fixed)

Before launch we ran an adversarial audit: independent agents re-derived values straight from
the raw EDGAR filings and diffed them against this dataset. It caught two real bugs:

1. **Fiscal-year labeling off-by-one (~25,000 rows in the full dataset).** Comparative-year
   figures (prior-year numbers re-reported inside a later 10-K) were tagged with the *filing's*
   fiscal year instead of the period they actually covered — e.g. Alphabet's FY2014 revenue
   (period ending 2014-12-31) was labeled FY2015. Fix: `fiscal_year` is now derived from
   `period_end` (the year the period ends; periods ending Jan 1–7 belong to the prior fiscal
   year, which preserves 52/53-week calendars like JNJ's).
2. **Diluted share counts reported on different scales by different filers.** Some companies
   tag share counts in millions (McDonald's files `716.4`), others in thousands (National
   Beverage files `93,620`), others in full units. We previously multiplied every sub-100k
   value by 1e6, which was wrong for the thousands filers by a factor of 1,000 — so we removed
   that correction rather than keep guessing. **We do not infer scale from magnitude.** These
   rows keep the filer's own number and carry `qa_status = FLAG:implausible_value`, and the API
   returns `null` for them rather than a figure we can't stand behind. Correcting them properly
   (deriving scale from the company's own NetIncome/EPS ratio) is on the roadmap; publishing a
   guess is not.

Item 1 is fixed in the published sample. Item 2 is *flagged, not fixed* — see above. Neither
bug affected `period_end`, and the 3,240 rows marked `filed_reliable = True` retain their
verified point-in-time stamps. The 40 oldest-year/edge dates that cannot be established from
the original XBRL filing remain explicitly unreliable rather than being presented as exact.

We publish this because "our data is audited" only means something if you also publish what the
audit found. A pipeline this size with zero bugs found would just mean nobody looked. If you
find something else, open an issue — corrections get published, not buried.

## Known limitations (we mark them, we don't hide them)

- **Oldest-year filing dates**: 40 rows where only a later XBRL filing exists; flagged, not faked.
  (Resolvable by cross-referencing the EDGAR submissions index — on the roadmap.)
- **Annual only** for now; quarterly (10-Q) point-in-time is the next build.
- **Banks/insurers**: "revenue" is an approximate concept for financials; treat JPM-type names with care.
- **40-company sample**: this repository is intentionally limited to 40 companies. The paid
  API serves the live full universe; exact totals are published at https://tradevodata.com/status.

The entire pitch is the line above each of these: a clean dataset *tells you what it doesn't know.*