jennyota/filing-boards

POPS4 Filing Boards

Nine tables built from United States SEC filings, refreshed every business day, each row linked to the original document on EDGAR.

Live boards: https://www.pops4.com/boards Source repository: https://github.com/HuangGoodmanAgency/filing-boards Dataset home: https://huggingface.co/datasets/jennyota/filing-boards

What makes this dataset unusual

Nothing in it is written. Every value either comes from a document that the row links to, or is a number computed from one. There is no summarisation step, no model-generated text, and no interpretation layer. The only claim a row makes is this company filed this document on this date.

Where a row carries a quote, it is a single verbatim sentence taken from the press release exhibit attached to that filing — not a paraphrase.

This makes the dataset suitable as a grounding or evaluation source, because every row is independently checkable against a public primary document.

Configs

Config	Filing	Records
cmo-moves	8-K Item 5.02	Departures and appointments of directors and officers
agreements	8-K Item 1.01	Entry into material definitive agreements
annual-meetings	Form DEF 14A	Definitive proxy statements
charter-amendments	8-K Item 5.03	Articles, bylaws, fiscal year; renames flagged
mergers	8-K Items 2.01, 5.01	Completed acquisitions and changes in control
ipo-watch	Form S-1	Companies registering to go public
restructuring	8-K Item 2.05	Costs associated with exit or disposal activities
breaches	8-K Item 1.05	Material cybersecurity incidents
auditor-watch	8-K Items 4.01, 4.02	Auditor changes, and non-reliance on earlier financial statements

Edgar & Edgarette

The same boards are live at https://www.pops4.com/boards, read by two voices: Edgar, who reports what was filed, and Edgarette, an actuarial reading of any SEC filer (exposure, experience, credibility, an 80% range, peers). Her range held 76% of past company-years against an 80% target, and every miss is published at https://www.pops4.com/boards/scorecard. Any AI can reach both through the open MCP endpoint https://www.pops4.com/boards/mcp. Numbers, never buy or sell views.

The method's own data

The dataset also carries what Edgarette's actuarial reading is built on:

• data/industry-yardstick.json — peer medians by SIC group (growth, operating margin, equity/assets, return on equity, payout), filers with $50M+ revenue, minimum 8 per group.
• data/range-calibration.json — the bands behind the published 80% revenue range.

The range is a baseline (last revenue × one plus the mean of the one- and three-year pace) widened by the 10th–90th percentile of real outcomes by size band, blended toward the whole pool by credibility. Fitted on 11,128 company-years (2018–22), tuned on 2023, tested on 2024: 83% of outcomes landed inside an 80% band. Every published range is logged in advance and scored when the next annual report arrives.

Yield, P/E and P/B are absent by design: they need a share price, which this dataset does not carry.

Fields

id (SEC accession number, the filing's own primary key) · company · ticker · exchange · cik · items (the filer's own 8-K item tags) · sector (the SEC's own industry wording) · filerCategory · revenue, revenueFY (as reported by the company in XBRL) · city, hq · formerName · event (period of report) · filed · deadline (breaches only) · quote (verbatim, where present) · source · profile.

Limitations, stated plainly

• About half of 8-K filings carry no press release exhibit, so most rows have no quote.
• Item 1.05 filings are genuinely rare — roughly one a fortnight. A near-empty breaches config is correct.
• Item 5.03 covers articles or bylaws or fiscal year. Most are not renames; genuine renames are flagged per row from the filer's own former name.
• deadline uses the shortest common statutory notification window. Real obligations follow where affected residents live, so one incident can run several clocks. It is an indicator, not legal advice.
• Item 4.01 (auditor change) is often routine; the auditor-watch config records the filing and implies nothing more.
• Coverage deliberately undercounts rather than guessing.

Licence and attribution

forecast-ledger — pre-registered ranges, resolved by law

The other configs are filings. This one is different, and it is the reason the dataset exists.

Each row is an 80% prediction interval on a company's next-year revenue, published before the outcome was known, fingerprinted at the moment of publication, and left untouched until that company files its next annual report and settles it. status is open until then.

Why that is hard to fake, and hard to copy:

• Resolution is statutory. The label arrives when the company files its 10-K, on a deadline set by law — not when a curator decides. Delayed, but scheduled.
• The grader is independent. The outcome is the company's own audited figure. A prediction hashed before an independently published outcome cannot be reward-hacked; you cannot back-date a 10-K.
• Pre-registration cannot be back-filled. Anyone with public SEC data can reconstruct historical predictions. Nobody can reconstruct having said it in public first. Each row carries its own fingerprint, the whole set carries a root hash, and the chain of daily roots sits in data/roots.log, committed to a git history we do not timestamp ourselves.
• The method is published, including the misses. Live method v3: fitted on 11,128 company-years (2018–22), tuned on 2023, tested on 2024 — 83.2% coverage, median width ±19.2%. Across all history, 69% of 7,172 company-years against an 80% target. Retired methods keep their record and their logged ranges are still scored.
• Calibration is reported by size band, not as one number — including 48% coverage for filers under $100M, which is the honest weak spot and is printed rather than smoothed.

Useful for evaluating interval calibration (coverage, width, interval score) and for outcome-supervised training in the future-as-label sense, where the passage of time supplies the label. Stated plainly: this is an evaluation-sized corpus, not a training-sized one, and it grows by roughly one range per company per fiscal year.

A range is arithmetic, not a forecast. Nothing here is investment advice.

The house behind it

POPS4 is one house with two faces. The record — these boards, the company cards and this ledger — is free and public, because an open record that anyone can check is what makes a house worth believing. The supply pays for it: corporate identity programs since 1997, 70,000+ products from 200+ authorized brands — corporate gifts, branded apparel, custom packaging and secure print — under one roof, one quote, no platform fee.

Procurement is mostly noise: chasing quotes, proofs and stock across four vendors who each own one piece. The house takes that away. The filing boards are the same instinct pointed at information — the answer is already in a public document, so stop making people hunt for it.

Both faces answer machines directly, free and without a key: https://www.pops4.com/boards/mcp for filings and cards, https://mcp.pops4.com/mcp for the catalogue.

CC BY 4.0. Free to use with credit to POPS4.

Underlying filings are published by the U.S. Securities and Exchange Commission and are public domain. This dataset adds structure and links, not content.

Citation

@misc{pops4_filing_boards,
  title  = {POPS4 Filing Boards: SEC 8-K signals with primary-source links},
  author = {Huang Goodman, Jenny},
  year   = {2026},
  url    = {https://www.pops4.com/boards},
  note   = {Hako Shikin LLC. CC BY 4.0.}
}

Maintained by Jenny Huang Goodman MPA MSc MHSA, Principal — Hako Shikin LLC.