FIRC Research Papers · FIRC0002 · v2.0
Does the Language of Artificial Intelligence Move Markets? A Semantic Analysis of Russell 3000 SEC Filings and the Cross-Section of Stock Returns, 2010–2026
Manel Kammoun, Djerrycharli Tandjambianda and Simon-Pierre Boucher
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
The post-ChatGPT AI premium is a large-cap episode, not a market-wide regime change
In the Russell 3000, AI language earns a modest, persistent +5.5 pp per year quintile spread (t = 2.41); the jump replicated in the S&P 500 (+5.0 to +15.0 pp) is confined to the largest thousand firms (+2.9 to +12.8 pp), and disclosed AI projects carry no premium.
Abstract
This paper measures artificial-intelligence (AI) engagement in the SEC filings of some 2,500 U.S. listed firms (the current S&P 500 constituents and the holdings of the iShares Russell 3000 ETF) by embedding every passage of their Form 10-K, 10-Q and 8-K filings of 2010–2026 with a sentence transformer and scoring it against curated AI anchor concepts (internal adoption, infrastructure exposure, product-disruption risk). The signal of filing year Y is linked to the forward twelve-month return from July of Y+1, so that it predates the holding period by at least six months. The panel of 27,468 firm-years (2,379 firms) extends the authors' earlier S&P 500 study, which rested on current reports only, to annual and quarterly reports and to the mid- and small-capitalization firms that make up most of the Russell 3000, and adds a second, independent measure of AI engagement: the AI, data-center, cloud and automation projects that a large-language-model reader extracts from the same filings (the SPID database).
Three results stand out. First, in the broad universe AI language carries a modest but persistent cross-sectional premium over the whole sample: the equal-weighted top-minus-bottom quintile spread is +5.5 percentage points per year (t = 2.41) and the Fama–MacBeth slope +3.07 points per standard deviation (t = 2.80), with a significant within-sector component. Second, the sharp post-ChatGPT increase documented for the S&P 500, replicated within the new panel (+5.0 to +15.0 points), does not generalize: in the Russell 3000 the post-launch spread is +7.9 points against +4.9 before, the 2022 cohort is negative, a placebo over ten candidate break years ranks 2022 eighth, and the monthly long–short earns no post-launch alpha. The increase is confined to the largest thousand firms (+2.9 to +12.8 points), whereas smaller firms earned about +9.6 points before the launch and +8.4 after; the ChatGPT event study shows the same size gradient (+19.5 versus +9.9 points of one-year abnormal return).
Third, disclosed AI projects carry no return premium, while the language of annual reports carries the strongest one. The evidence is consistent with a persistent, partly within-industry premium on AI exposure in the broad market and with a large-cap re-rating after 2022, not with a market-wide regime change; an index sample alone would have read a large-cap episode as a change in how the market prices disclosure.
Key findings
27,468
firm-years, 2,379 firms, formation years 2010–2024
+5.5 pp/yr
equal-weighted Q5–Q1 spread, full sample (t = 2.41)
+3.07 pp/SD
Fama–MacBeth slope (t = 2.80)
+2.9 → +12.8 pp
post-ChatGPT spread among the largest 1,000 firms
8th of 10
rank of 2022 among placebo break years
−0.22
Fama–MacBeth slope of disclosed AI projects (t = −0.24)
Scientific contributions
- A broad-market test of AI-language pricing
- The laboratory is enlarged from 500 index constituents to about 2,500 firms across the Russell 3000 capitalization spectrum, from current reports alone to the full text of 10-K, 10-Q and 8-K filings, and to 27,468 look-ahead-free firm-years.
- A modest, persistent premium in the broad market
- The equal-weighted Q5–Q1 spread is +5.5 pp per year (t = 2.41), the value-weighted spread +13.9 (t = 2.59), and the Fama–MacBeth slope +3.07 pp per SD (t = 2.80), with a within-sector slope of +1.46 (t = 4.38).
- The post-ChatGPT jump does not generalize
- Replicated in the S&P 500 members (+5.0 to +15.0 pp, t = 2.21), the increase is +4.9 to +7.9 pp in the Russell 3000 (post t = 1.34), with a negative 2022 cohort, a placebo rank of 8th out of 10 break years and a post-launch FF6 alpha of −0.3% per year.
- A size gradient, not a regime
- The largest 1,000 firms rise from +2.9 to +12.8 pp (t = 2.09) while the remaining firms move from +9.6 to +8.4; the ChatGPT event study gives +19.5 pp one-year abnormal return for the largest 1,000 versus +9.9 for the rest.
- Words versus projects
- Disclosed AI projects from the SPID database carry no premium (tech-project quintile spread −1.9 pre and −1.3 post; FM −0.22, t = −0.24), whereas 10-K language carries the strongest signal (+6.9 pre to +16.7 pp post, t = 2.61).
- A methodological warning about index samples
- The S&P 500 is the one part of the market where the premium was absent before 2022 and large afterward; a study confined to it reads a large-cap episode as a change in how the market prices disclosure.
Methodology
Methodology (5)
Semantic scoring of filings
Each 512-word chunk is embedded with a sentence transformer and scored against written anchor passages for AI adoption, infrastructure exposure and product-disruption risk; chunk scores are aggregated to a standardized firm-year composite.
Look-ahead-free portfolio sorts
The signal of filing year Y is linked to the twelve-month return from July Y+1 to June Y+2; equal- and value-weighted quintile spreads are computed for fifteen formation cohorts, with S&P 500 members and the full universe reported side by side.
Fama–MacBeth and panel regressions
Slopes per standard deviation with and without sector dummies and sector-by-year fixed effects, firm-clustered pooled tests, winsorized returns and size-neutral sorts.
Break-date placebo and factor spanning
Pre/post ChatGPT comparisons are tested against ten candidate break years; the monthly long–short is spanned on FF6 plus a TECH factor to check for post-launch alpha.
ChatGPT event study
Abnormal returns against the IWV market index over 60 days to one year (N = 2,226), with Welch and Brown–Warner crude-dependence-adjustment t-statistics, split by firm size.
Research data
- UQO EDGAR Research Database, Russell 3000 extension
- 527,548 Form 10-K, 10-Q and 8-K filings of 2,499 issuers filed since January 2010, parsed into sections and 512-word chunks; 10,720,890 chunks embedded with all-MiniLM-L6-v2 (384 dimensions) and scored by cosine similarity to anchor passages for eight concepts. Research funded by the FIRC (Fonds institutionnel de recherche et de création, UQO).
- Return and factor panel
- Monthly adjusted-close prices for 2,499 tickers from Yahoo Finance (June 2009–October 2026), SPY and IWV benchmarks, SEC share counts, Fama–French five factors plus momentum and a TECH factor; 27,468 firm-years with valid forward returns (6,793 for S&P 500 members).
- SPID AI-project counts
- Firm-year counts of AI, data-center, cloud and automation projects extracted by a large-language-model reader from the same filings (UQO Working Paper No. 0001), used as an independent, content-based measure of AI engagement.
Reproducibility
- Eight numbered scripts (scripts/r3k_00–07) from factor download and Yahoo prices to panel, core and extended analyses, event study, 14 figures and tables
- Every number in the text is a LaTeX macro generated in tables/macros.tex; results stored in results/r3k/*.json
- Previous S&P 500 version (v1.1) archived with its review report and revision changelog in docs/
- Full LaTeX source of the 48-page paper plus a LaTeX-to-Word conversion pipeline; research funded by the FIRC (Fonds institutionnel de recherche et de création, UQO)
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v2.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 48 p. |