FIRC Research Papers · FIRC0001 · v1.0
From Corporate Disclosures to a Russell 3000 Strategic Project Database: Identification, Measurement, Validation, and Research Applications
Manel Kammoun, Djerrycharli Tandjambianda and Simon-Pierre Boucher
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
121,099 disclosed strategic projects, each traceable to an SEC filing passage
SPID reads 527,548 filings of 2,564 issuers since 2010 and extracts 308,065 project mentions into a 21-category taxonomy; the paper documents the instrument and its audit, not yet a validated measurement.
Abstract
This paper documents the construction of SPID, a project-level research database that reconstructs the strategic initiatives disclosed by U.S.-listed companies in their SEC filings. Starting from an issuer universe that merges a current S&P 500 snapshot with the holdings of the iShares Russell 3000 ETF (2,564 distinct SEC registrants), the pipeline collects every 10-K, 10-Q and 8-K filed since January 2010 (527,548 documents), segments them into 1.34 million sections, and submits 1.12 million eligible sections to a schema-constrained large-language-model extractor.
The result is 308,065 evidence-linked project mentions, resolved into 121,099 candidate projects for 2,492 issuers, each tied to its SEC accession number, section and source passage, together with a controlled 21-category taxonomy, status, cited years, locations, technologies, partners and disclosed amounts. The paper describes the universe definition and its reconciliation, the distributed acquisition and parsing architecture, the extraction contract and its confidence semantics, the mention-to-project resolution rule and its known failure modes, and the point-in-time structure of the timelines.
A validation framework (stratified human annotation, precision and recall on an independent gold standard, linkage audits and coverage gates) is set out, stating explicitly which components are implemented and which remain to be executed. Descriptive patterns are reported for the full corpus (M&A and integration account for 34.5% of projects; a USD amount is attached to 46.7% of projects, median 128.5 M USD), with the caveat that a disclosed project is neither a verified investment nor a realized expenditure. SPID should be read as a research instrument under validation rather than as an exhaustive census of Russell 3000 projects.
Key findings
527,548
SEC filings (10-K, 10-Q, 8-K) of 2,564 issuers, 2010–2026
1.12 million
filing sections submitted to the LLM extractor
308,065
evidence-linked project mentions
121,099
resolved candidate projects for 2,492 issuers
34.5%
of projects are M&A and post-merger integration
128.5 M USD
median largest disclosed amount per project
Scientific contributions
- A precise measurement target
- A project is defined as a distinct, identifiable, issuer-associated strategic initiative, separated from a project mention, a generic strategy, a historical or third-party reference and an uncertain candidate, and tied to the existing 21-category SPID taxonomy.
- A scalable, reproducible architecture
- A frozen issuer universe, distributed acquisition across ten Apple-silicon cluster nodes and one external server under a shared SEC rate budget, section-level parsing of 527,548 filings and a resumable, idempotent extraction step costing roughly 775 USD.
- An extraction contract exposed to criticism
- The prompt, the JSON schema, the confidence threshold (mentions below 0.45 dropped) and the deterministic resolution rule are documented in enough detail to be audited; 35.2% of mentions carry confidence 1.0 and 32.7% carry 0.9, a calibration question left for the annotation study.
- Descriptive patterns with explicit caveats
- M&A and integration dominate (34.5% of projects), followed by capex programs (8.9%) and partnerships (8.2%); technology categories are small in count (264–449 projects each) but carry high amounts (median 128–166 M USD for data centers and AI initiatives).
- A validation framework with stated status
- Stratified human annotation by two independent coders, precision/recall/F1 on a gold standard, confidence calibration, linkage audits and coverage gates are specified; the paper states which gates are met and which are not, and asks that SPID be cited as a database under validation.
Methodology
Methodology (5)
Universe definition and reconciliation
S&P 500 snapshot and IWV holdings are matched to SEC CIKs (2,564 registrants, 2,499 filing the three forms); the fund-holdings proxy at one date is acknowledged as a substitute for licensed Russell membership.
Distributed acquisition and parsing
Eleven machines (ten MacLustr Apple-silicon nodes and one OVHcloud server) download under SEC fair-access limits with an interleaved CIK allocation, parse filings into sections and export tables to Parquet.
Schema-constrained LLM extraction
1.12 million eligible sections (focus windows around lexical cues) are submitted to an extractor returning structured mentions with category, status, amounts, years, locations, technologies, partners and a self-reported confidence.
Mention-to-project resolution and timelines
A deterministic key resolves mentions into projects with point-in-time timelines; its known failure modes (merging distinct phases, splitting renamed initiatives) are documented, with explainable linkage and human adjudication proposed.
Validation protocol
Stratified gold-standard annotation, inter-annotator agreement (Cohen's κ), mention-level precision/recall/F1, category macro/micro-F1, field accuracy, abstention rate and confidence calibration with bootstrap intervals, to be executed before derived data release.
Research data
- SPID (SEC Project Intelligence Database), Russell 3000 release
- 527,548 Form 10-K, 10-Q and 8-K filings of 2,564 issuers (current S&P 500 snapshot merged with iShares IWV holdings) filed January 2010–October 2026, parsed into 1.34 million sections; 308,065 mentions and 121,099 projects stored in DuckDB (spid.duckdb, 3.2 GB) on top of the 68 GB UQO EDGAR research corpus. Research funded by the FIRC (Fonds institutionnel de recherche et de création, UQO).
- Project attributes and graph
- Each project carries category, status, cited years, locations, technologies, partners and disclosed amounts (USD amount on 120,240 mentions and 56,500 projects; median 128.5 M USD, P99 12.5 B USD) and is projected into a graph of 183,472 nodes; health care (23,679 projects) and industrials (17,890) contribute the most.
Reproducibility
- Corpus and project layers reproducible from the uqo-research-db repository (revision 2324b58), the frozen universe artifact, download logs and the extraction ledger
- Every number in the paper regenerated by code/export_stats.py (DuckDB) and code/make_tables.py (CSV to LaTeX tables and figures)
- Claim register (Appendix D) separating implemented, documented-but-unverified, proposed and empirically verified claims; raw filings not redistributed under SEC fair-access policy
- Full LaTeX source of the 39-page paper; research funded by the FIRC (Fonds institutionnel de recherche et de création, UQO)
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 39 p. |