FIRC0002 Stage E (Agent 10): §7 discussion in descriptive register with verdicts table (H3b fails, H4 not supported, H5a/b) and itemised limitations naming every unexecuted item with reason and bias direction (X1–X5, AR-3, IWV 2010–18, THEO-06, ECON-08 B/M); §8 conclusion rewritten; appendices — variables (A1 on one page), anchors with stripped/core sets and probe, corpus counts as macros (DATA-10), new SPID extractor appendix (ADVE-16, AI_C-13), data sources/reproducibility/data-availability statement (cluster L)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
3 changed files +347 −171
modified
paper/appendix/appendix.tex
+128 −55
@@ -1,5 +1,6 @@ | ||
| 1 | 1 | % ============================================================================ |
| 2 | −% Appendix | |
| 2 | +% Appendix — Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, | |
| 3 | +% ../results/n1/corpus_macros.tex, sections/macros_stageE.tex. | |
| 3 | 4 | % ============================================================================ |
| 4 | 5 | \newpage |
| 5 | 6 | \begin{appendices} |
@@ -12,32 +13,37 @@ | ||
| 12 | 13 | \section{Variable Definitions} |
| 13 | 14 | \label{app:variables} |
| 14 | 15 | |
| 15 | −\begin{table}[htbp] | |
| 16 | +\Cref{tab:app_variables} defines every variable used in the text, in the order in which it is | |
| 17 | +introduced. | |
| 18 | + | |
| 19 | +\begin{table}[H] | |
| 16 | 20 | \centering |
| 17 | − \caption{Variable Definitions} | |
| 21 | + \caption{Variable definitions} | |
| 18 | 22 | \label{tab:app_variables} |
| 19 | 23 | \small |
| 20 | 24 | \begin{tabular}{@{}p{5.1cm}p{9.6cm}@{}} |
| 21 | 25 | \toprule |
| 22 | 26 | Variable & Definition \\ |
| 23 | 27 | \midrule |
| 24 | − AI\,Adopt (\texttt{ai\_adoption\_wmean}) & Average over the scored filings of firm $i$ dated in calendar year $Y$ of the section-weighted mean cosine similarity of the filing's text chunks to the AI-internal-adoption anchor, equations~\eqref{eq:cos}--\eqref{eq:annual}. \\[4pt] | |
| 25 | − AI\,Infra (\texttt{ai\_infrastructure\_wmean}) & Same construction for the AI-compute-infrastructure anchor. \\[4pt] | |
| 26 | − AI\,Risk (\texttt{ai\_disruption\_risk\_wmean}) & Same construction for the AI-product-disruption-risk anchor. \\[4pt] | |
| 28 | + AI\,Adopt\newline(\texttt{ai\_adoption\_wmean}) & Average over the scored filings of firm $i$ dated in calendar year $Y$ of the section-weighted mean cosine similarity of the filing's text chunks to the AI-internal-adoption anchor, \cref{eq:cos,eq:wmean,eq:annual}. The encoder reads the first \mMaxSeq{} WordPiece tokens of each chunk (\cref{subsec:corpus_data}). \\[4pt] | |
| 29 | + AI\,Infra\newline(\texttt{ai\_infrastructure\_wmean}) & Same construction for the AI-compute-infrastructure anchor. \\[4pt] | |
| 30 | + AI\,Risk\newline(\texttt{ai\_disruption\_risk\_wmean}) & Same construction for the AI-product-disruption-risk anchor. \\[4pt] | |
| 27 | 31 | Simple mean, top-5\% tail & Filing-level unweighted mean of chunk scores, and mean of the chunks at or above the filing's 95th percentile, averaged over filings as above. \\[4pt] |
| 28 | − AI\,Compos & $\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$, with $z$ the within-formation-year cross-sectional standardization of equation~\eqref{eq:z}. \\[4pt] | |
| 29 | − Form-specific composites & AI\,Compos rebuilt from the filing-level scores of Form~10-K, Form~10-Q or Form~8-K filings only (\texttt{data/raw\_russell3000/filing\_scores}). \\[4pt] | |
| 30 | − SPID technology projects & $\log(1+n)$, where $n$ is the number of project mentions of types AI initiative, data center, cloud migration, automation, digital transformation, ERP implementation or technology deployment extracted by the SPID database from the firm's filings of year $Y$; standardized within year. Any-AI-project is an indicator for at least one AI-initiative mention or AI-tagged technology in year $Y$. \\[4pt] | |
| 31 | − Pre-determined exposure & Firm average of AI\,Compos over filing years 2019--2021, standardized across firms; binary treatment equals one above the median. \\[4pt] | |
| 32 | − Forward return & Buy-and-hold simple return from July of $Y{+}1$ to June of $Y{+}2$, equation~\eqref{eq:fwd}, from adjusted closes (dividends reinvested), requiring at least six monthly returns. \\[4pt] | |
| 32 | + AI\,Compos & $\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$, with $z$ the within-formation-year cross-sectional standardization of \cref{eq:z}. PC1 is the first principal component of the three standardized facets. \\[4pt] | |
| 33 | + Form-specific and equal-form composites & AI\,Compos rebuilt from the filing-level scores of Form~10-K, Form~10-Q or Form~8-K filings only; the equal-form composite is the mean of the three form-specific $z$-scores. \\[4pt] | |
| 34 | + Keyword baseline & Core AI terms (artificial intelligence, machine learning, neural network, large language model, \dots) per thousand words of the firm's filing sections of year $Y$, levels and $\log(1+x)$; standardized within year (\cref{subsec:construct}). \\[4pt] | |
| 35 | + Stripped and core anchors & Composites from anchor paragraphs without the cybersecurity and automation vocabulary, and from a core-AI adoption anchor, scored against the stored chunk embeddings (\cref{app:anchors}). \\[4pt] | |
| 36 | + SPID technology projects & $\log(1+n)$, where $n$ is the number of project mentions of types AI initiative, data center, cloud migration, automation, digital transformation, ERP implementation or technology deployment extracted by the SPID database from the firm's filings of year $Y$; standardized within year. Any AI project is an indicator for at least one AI-initiative mention or AI-tagged technology in year $Y$; any technology project an indicator for $n>0$. \\[4pt] | |
| 37 | + Pre-determined exposure & Firm average of AI\,Compos over filing years 2019--2021, standardized across firms; binary treatment equals one above the median. Trailing exposure: mean composite over the three filing years before the outcome year. \\[4pt] | |
| 38 | + Forward return & Buy-and-hold simple return from July of $Y{+}1$ to June of $Y{+}2$, \cref{eq:fwd}, from adjusted closes (dividends reinvested) verified against a second daily source, requiring at least six monthly returns and listing at formation; winsorized at the 1st and 99th percentiles within formation year where stated. \\[4pt] | |
| 33 | 39 | Market-adjusted return & Forward return minus the contemporaneous SPY return (IWV-adjusted version in robustness). \\[4pt] |
| 34 | − Size & Natural logarithm of market capitalization at formation: price times the shares outstanding most recently reported on the cover of a 10-K or 10-Q filed on or before the formation date (SEC XBRL \texttt{dei:EntityCommonStockSharesOutstanding}). \\[4pt] | |
| 35 | − Size group & Largest 1,000 firms by market capitalization at formation (Russell~1000 proxy) versus the remaining firms (Russell~2000 proxy), re-assigned each year. \\[4pt] | |
| 36 | − Momentum & Own return over the 12 months ending at formation, skipping the last month. \\[4pt] | |
| 37 | − Volatility & Annualized standard deviation of monthly returns over the trailing 12 months. \\[4pt] | |
| 38 | − Post & Indicator equal to one for formation (filing) years $\ge 2022$. \\[4pt] | |
| 40 | + Size & Natural logarithm of market capitalization at formation: unadjusted month-end close (second price source; Yahoo-implied raw close as fallback) times the point-in-time shares outstanding reported on the cover of the most recent 10-K or 10-Q filed on or before the formation date (SEC XBRL \texttt{dei:EntityCommonStockSharesOutstanding}, summed across classes, unit-error screen), on the same split basis; screened and winsorized within year (\cref{subsec:price}). \\[4pt] | |
| 41 | + Size group & Largest 1,000 firms by market capitalization at formation (Russell~1000 proxy) versus the remaining firms (Russell~2000 proxy), re-assigned each year; alternatively the largest 42\% of each cohort. \\[4pt] | |
| 42 | + Momentum & Own return over the 12 months ending at formation, skipping the last month; winsorized within year. \\[4pt] | |
| 43 | + Volatility & Annualized standard deviation of monthly returns over the trailing 12 months; winsorized within year. \\[4pt] | |
| 44 | + Post & Indicator equal to one for formation (filing) years $\ge 2022$ (alternative split: $\ge 2021$). \\[4pt] | |
| 39 | 45 | TECH & Equal-weighted monthly return of the Information Technology constituents of the panel minus the equal-weighted return of all constituents. \\[4pt] |
| 40 | − CAR & Cumulative market-model abnormal return, equations~\eqref{eq:mm}--\eqref{eq:car}; market model estimated 3~January--31~October 2022 on the iShares Russell~3000 ETF (IWV), event window 1~November 2022--30~June 2024. \\ | |
| 46 | + CAR & Cumulative market-model abnormal return, \cref{eq:mm,eq:car}; market model estimated 3~January--31~October 2022 on the iShares Russell~3000 ETF (IWV); day~0 is 1~December 2022; pre-launch block $[-20,-1]$ reported separately; alternative benchmarks in \cref{subsec:m_event}. \\ | |
| 41 | 47 | \bottomrule |
| 42 | 48 | \end{tabular} |
| 43 | 49 | \end{table} |
@@ -48,8 +54,16 @@ | ||
| 48 | 54 | Each concept is defined by the three paragraphs below (\texttt{src/factors/anchors.py} of the |
| 49 | 55 | EDGAR Research Database). The three paragraphs are embedded separately with |
| 50 | 56 | \texttt{all-MiniLM-L6-v2}, averaged, and normalized to a unit vector; chunk scores are cosine |
| 51 | −similarities to that vector. The anchors were written by the authors in June 2026 and have not been | |
| 52 | −revised since; no alternative anchor set has been tested. | |
| 57 | +similarities to that vector. The anchors were written by the authors in June 2026 and have not | |
| 58 | +been revised since. For the revision, two alternative anchor sets were embedded and scored | |
| 59 | +against the stored chunk embeddings (no corpus re-embedding): the original anchors with every | |
| 60 | +cybersecurity and automation phrase removed, and a core-AI adoption anchor restricted to | |
| 61 | +artificial intelligence, machine learning and generative-AI terms; the stripped composite has a | |
| 62 | +rank correlation of \mStripRho{} with the baseline at the firm-year level and \mStripQfiveStays\% | |
| 63 | +of top-quintile firm-years stay in the top quintile (\cref{subsec:construct}). A probe of written | |
| 64 | +sentences (affirmative AI, negated AI, paraphrases without keywords, automation without AI, | |
| 65 | +cybersecurity without AI, utility data-center load, unrelated text) was scored with the same | |
| 66 | +anchors; its group means and corpus percentiles are in \texttt{results/n1/}. | |
| 53 | 67 | |
| 54 | 68 | \paragraph{AI adoption (operational weighting).} (1) ``The company internally uses artificial |
| 55 | 69 | intelligence, machine learning, generative AI, large language models, automation, predictive |
@@ -98,47 +112,106 @@ reskilling, human--machine collaboration, and ``the future of work transformatio | ||
| 98 | 112 | robotics.'' The macro-risk anchor describes inflation, interest and exchange rates, recession, |
| 99 | 113 | geopolitical tensions, tariffs, commodity prices, and monetary and fiscal policy. The full texts |
| 100 | 114 | are in the source file cited above. Two of the placebo anchors (cybersecurity and automation/labour) |
| 101 | −mention AI explicitly, which bears on the interpretation of Table~\ref{tab:placebo}. | |
| 115 | +mention AI explicitly, and the adoption anchor lists cybersecurity and automation among the uses of | |
| 116 | +AI, which bears on the interpretation of \cref{tab:placebo} and motivates the stripped anchors. | |
| 102 | 117 | |
| 103 | 118 | \section{Corpus Composition} |
| 104 | 119 | \label{app:corpus} |
| 105 | 120 | |
| 106 | −The EDGAR Research Database was extended on 27--28 October 2026 from the 500 S\&P~500 constituents to | |
| 107 | −an issuer universe that merges the S\&P~500 snapshot of May 2026 with the holdings of the iShares | |
| 108 | −Russell~3000 ETF as of 23~October 2026 (2,564 distinct SEC registrants). For every issuer it | |
| 109 | −downloaded the primary document of every Form~10-K, 10-Q and 8-K filed since January 2010 | |
| 110 | −(\num{527548} filings), parsed all of them into sections and 512-word chunks (\num{1342283} sections, | |
| 111 | −\num{10720890} chunks), embedded every chunk and scored it against the eight anchors. Unlike the June | |
| 112 | −2026 release, in which only the 8-K filings and the 10-Q filings of 84 firms had been scored, the | |
| 113 | −present panel is built from all three forms for every firm; Table~\ref{tab:filingtype} in the text | |
| 114 | −reports the composition of the firm-year signal by form. The year variable of the panel is the | |
| 115 | −calendar year of the filing date. The construction of the universe, its reconciliation and the | |
| 116 | −extraction of the SPID project signals are documented in the companion methodological paper (UQO | |
| 117 | −Working Paper No.~0001). | |
| 118 | − | |
| 119 | −\section{Data Sources and Reproducibility} | |
| 121 | +The EDGAR Research Database was extended on 27--28 October 2026 from the 500 S\&P~500 constituents | |
| 122 | +to an issuer universe that merges the S\&P~500 snapshot of May 2026 with the holdings of the | |
| 123 | +iShares Russell~3000 ETF as of 23~October 2026 (2,564 distinct SEC registrants). For every issuer | |
| 124 | +it downloaded the primary document of every Form~10-K, 10-Q and 8-K filed since January 2010 | |
| 125 | +(\mnFilingsCorpusN{} filings), parsed all of them into sections and chunks (\mnSectionsCorpus{} | |
| 126 | +sections, \mnChunksCorpus{} chunks, of which \mnChunksTenK{} from annual, \mnChunksTenQ{} from | |
| 127 | +quarterly and \mnChunksEightK{} from current reports), embedded every chunk and scored it against the | |
| 128 | +eight anchors; the counts are exported from the database by a read-only audit (\texttt{code/n1}) | |
| 129 | +and enter the text as macros. Unlike the June 2026 release, in which only the 8-K filings and the | |
| 130 | +10-Q filings of 84 firms had been scored, the present panel is built from all three forms for | |
| 131 | +every firm; \cref{tab:filingtype} reports the composition of the firm-year signal by form. The | |
| 132 | +year variable of the panel is the calendar year of the filing date. The construction of the | |
| 133 | +universe, its reconciliation and the extraction of the SPID project signals are documented in | |
| 134 | +the companion methodological paper \citep{uqowp0001}. | |
| 135 | + | |
| 136 | +\section{The SPID Extractor: Model, Prompt, Settings and Validation Frame} | |
| 137 | +\label{app:spid} | |
| 138 | + | |
| 139 | +The project counts of \cref{subsec:spid} come from the SPID project database \citep{uqowp0001}. | |
| 140 | +The extractor, as recorded in the code snapshot of the FIRC-0001 pipeline archived with this | |
| 141 | +review, sends the eligible sections of each filing to the \mSpidModel{} chat-completion model | |
| 142 | +with temperature zero, a JSON-object response format and a 2,000-token output limit; long sections | |
| 143 | +are reduced to windows around project-signal keyword hits before submission. The system prompt | |
| 144 | +instructs the model to act as ``a financial analyst extracting CORPORATE PROJECTS from SEC filing | |
| 145 | +text'', defines a project as ``any concrete initiative, investment, program, expansion, | |
| 146 | +construction, product/R\&D effort, acquisition/integration, technology deployment, partnership, | |
| 147 | +or strategic program the company is pursuing'', requires a JSON object with, for each project, a | |
| 148 | +name, exactly one type from the allowed list, a one-sentence description ``grounded in the | |
| 149 | +text'', an objective, a status (planned, in progress, completed, mentioned), an amount in U.S.\ | |
| 150 | +dollars if stated, locations, technologies, partners, benefits, risks, years and a confidence | |
| 151 | +between zero and one, and closes with the rules ``Extract ONLY real initiatives, not generic | |
| 152 | +risk-factor boilerplate or hypotheticals'', ``Do not invent facts'' and ``Prefer fewer | |
| 153 | +high-quality projects over many vague ones''. The extraction was run on 27--28 October 2026 | |
| 154 | +(database file dated \mSpidDbDate); its cost is recorded in the FIRC-0001 run registry. Every | |
| 155 | +mention of the panel's registrants---\mSpidUniverseMentions{} in all---carries the model's | |
| 156 | +confidence, whose median is \mSpidConfMedian. | |
| 157 | + | |
| 158 | +No human or machine validation of the extraction exists. For the revision, a stratified sampling | |
| 159 | +frame of \mSpidSampleN{} mentions (\mSpidStrata{} strata of project type $\times$ filing form, | |
| 160 | +fixed seed) was drawn with, for every mention, the source-section excerpt in which the project | |
| 161 | +name was located (found for \mSpidEvidenceShare\% of mentions; the others show the start of the | |
| 162 | +section) and empty label columns for ``project present'' and ``AI-related'' | |
| 163 | +(\texttt{results/r3k\_v2/validation/spid\_sample.csv}). The labels are to be filled by two human | |
| 164 | +readers; a language-model second rater was proposed and not authorized. Until the labels exist, | |
| 165 | +the precision and recall of the measure are unknown, the errors-in-variables attenuation of its | |
| 166 | +slope cannot be corrected, and the project results of \cref{subsec:spidres} bound the premium on | |
| 167 | +the measure rather than on projects. A parallel frame for the embedding score (a stratified | |
| 168 | +sample of chunks by score decile and cohort, the top chunks of the top-30 and thirty median firms | |
| 169 | +of \cref{tab:topfirms}, and a codebook) is in \texttt{results/n1/} and is likewise uncoded. | |
| 170 | + | |
| 171 | +\section{Data Sources, Reproducibility and Data Availability} | |
| 120 | 172 | \label{app:repro} |
| 121 | 173 | |
| 122 | −Semantic scores and project mentions are from the EDGAR Research Database and the SPID project | |
| 123 | −database (UQO), which are not redistributed; the firm-year factor panel, the filing-level scores | |
| 124 | −by form, the SPID firm-year counts, the monthly and daily prices, the share-count history, the | |
| 125 | −factor series and the long--short series needed to reproduce every table and figure are kept in | |
| 126 | −\texttt{data/raw\_russell3000/} and \texttt{data/processed\_r3k/}. Equity prices are from Yahoo | |
| 127 | −Finance (downloaded 28~October 2026); shares outstanding are from the SEC XBRL company-facts API; | |
| 128 | −the Fama--French five factors, the momentum factor and the risk-free rate are from the Kenneth R.\ | |
| 129 | −French data library (through September 2026). The pipeline in \texttt{scripts/r3k\_*.py} runs in | |
| 130 | −numerical order: \texttt{00} downloads the factors, \texttt{01} the prices and \texttt{01c} the | |
| 131 | −share counts, \texttt{02} builds the look-ahead-free panel, \texttt{03} runs the sorts, factor | |
| 132 | −regressions, panel and Fama--MacBeth estimators, the break interaction and the size-group analyses, | |
| 133 | −\texttt{04} the placebo, sector-neutral and double sorts, exclusions, alternative signals, | |
| 134 | −horizons and benchmarks, principal components, difference-in-differences and event-time profile, | |
| 135 | −cohort and pooled tests, entrants, sector controls, technology-factor spanning, filing-type | |
| 136 | −composition, placebo break dates and the SPID interactions, \texttt{05} the event study, | |
| 137 | −\texttt{06} the figures and \texttt{07} the LaTeX tables and the number macros used in the text. | |
| 138 | −All numerical results are serialized to \texttt{results/r3k/*.json}; every number quoted in the | |
| 139 | −text is a macro generated by \texttt{07} or a table cell, so that the manuscript can be | |
| 140 | −regenerated from the data without manual editing. The June 2026 and October 2026 (v1.1) versions | |
| 141 | −of the paper, which covered the S\&P~500 only, are archived in | |
| 142 | −\texttt{docs/paper\_v1.1\_sp500\_20261026/}. | |
| 174 | +\paragraph{Sources.} Semantic scores and project mentions are from the EDGAR Research Database and | |
| 175 | +the SPID project database (UQO), which are not redistributed; the filing text and the chunk | |
| 176 | +embeddings stay in the database. Equity prices are from Yahoo Finance (monthly adjusted closes of | |
| 177 | +the 28~October 2026 vintage used by the pre-review version; a second pass of 5~November 2026 with | |
| 178 | +split histories, daily prices and benchmark ETFs); a second, licensed daily price source (FinData | |
| 179 | +OS, FirstRate data) was used read-only to verify extreme returns and to obtain unadjusted | |
| 180 | +month-end closes; shares outstanding are from the SEC XBRL company-facts API (public domain); | |
| 181 | +the Fama--French five factors, the momentum factor and the risk-free rate, monthly and daily, are | |
| 182 | +from the Kenneth R.\ French data library; point-in-time IWV holdings are the \mnIWVsnapshots{} | |
| 183 | +dated holdings files captured by the Internet Archive, listed with checksums in | |
| 184 | +\texttt{results/r3k\_v2/iwv\_holdings\_manifest.json}. | |
| 185 | + | |
| 186 | +\paragraph{Pipeline.} The pre-review exhibits (version 2.0) were produced by | |
| 187 | +\texttt{scripts/r3k\_00}--\texttt{r3k\_07} at the commit recorded in the repository README; a | |
| 188 | +re-run of that chain reproduces its results except two blocks that were stale at commit time | |
| 189 | +(the SPID regressions and the double sorts). The revised exhibits (version 2.1) are produced by | |
| 190 | +\texttt{scripts/r3k\_10}--\texttt{r3k\_23} in the order of \texttt{scripts/run\_all.sh}: second | |
| 191 | +price pass and daily factors, SEC share re-pull, second-source verification of extreme returns and | |
| 192 | +raw closes, the rebuilt panel, the core, extended, event-study and survivorship analyses, the | |
| 193 | +figures, the tables and the number macros, and the construct-validity rows from the node audit; | |
| 194 | +\texttt{tests/test\_regenerate.py} regenerates the four results files in a scratch copy and | |
| 195 | +compares them with the committed ones at a relative tolerance of $10^{-6}$ (all seeds fixed), and | |
| 196 | +a pinned environment (\texttt{requirements.lock}) is provided. Every number quoted in the text is | |
| 197 | +a macro generated from a results file or a table cell, so that the manuscript regenerates from | |
| 198 | +the data without manual editing. | |
| 199 | + | |
| 200 | +\paragraph{Data availability.} The firm-year analysis panel, the filing-level scores by form, the | |
| 201 | +SPID firm-year counts, the price and share files, the factor series, the long--short series and | |
| 202 | +the event-study abnormal returns needed to regenerate every exhibit (about 420~MB) are documented | |
| 203 | +by a SHA-256 manifest (\texttt{manifest.sha256}) and a data README that states the provenance and | |
| 204 | +licence of each file. They are not yet deposited in a public archive; the authors will release | |
| 205 | +them with the working paper (as a repository release or as an archive referenced by hash), with | |
| 206 | +the exception of the FinData/FirstRate-derived files, which are access-restricted under the | |
| 207 | +provider's licence and are needed only for the market-capitalization and verification steps (the | |
| 208 | +Yahoo-implied fallback reproduces the price months within two percent in almost every case). | |
| 209 | +Until the deposit is made, the exhibits are reproducible by the authors from documented inputs | |
| 210 | +and not by a third party. The June 2026 and October 2026 (version 1.1) versions of the paper, | |
| 211 | +which covered the S\&P~500 only, are archived in \texttt{docs/paper\_v1.1\_sp500\_20261026/}. | |
| 212 | + | |
| 213 | + | |
| 214 | +% Stage E note: the appendix-bound exhibits are printed by tables/appendix_exhibits.tex (Agent 9), which | |
| 215 | +% main.tex inputs after this appendix; every exhibit there is behind an \ExhibitOnce guard. | |
| 143 | 216 | |
| 144 | 217 | \end{appendices} |
modified
paper/sections/conclusion.tex
+42 −31
@@ -1,39 +1,50 @@ | ||
| 1 | 1 | % ============================================================================ |
| 2 | +% Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, sections/macros_stageE.tex. | |
| 2 | 3 | \section{Conclusion} |
| 3 | 4 | \label{sec:conclusion} |
| 4 | 5 | % ============================================================================ |
| 5 | 6 | |
| 6 | 7 | Using a semantic, embedding-based measure of artificial-intelligence language in the annual, |
| 7 | −quarterly and current SEC reports of some 2,500 U.S.\ listed firms over fifteen formation cohorts, | |
| 8 | −linked to look-ahead-free forward returns, factor data and an independent project-based measure of | |
| 9 | −AI engagement, we have re-examined whether and when AI language was priced in the cross-section of | |
| 10 | −returns. The answer differs from the one we gave for the S\&P~500 alone, and the difference is | |
| 11 | −instructive. In the broad market, AI language carried a modest premium throughout 2010--2024: about | |
| 12 | −\mspreadFull{} percentage points per year between the top and bottom quintiles, \mrkFMcompos{} points | |
| 13 | −per standard deviation in Fama--MacBeth regressions, with a within-sector component that survives | |
| 14 | −sector-by-year fixed effects over the full sample and a concentration among the smaller firms that | |
| 15 | −the index excludes. The sharp post-ChatGPT increase that the index displays---and that we replicate | |
| 16 | −within the new panel---is a large-cap phenomenon: among the largest thousand firms the spread rises | |
| 17 | −from \mlargePre{} to \mlargePost{} points and the event-study re-rating reaches \mevLarge{} points | |
| 18 | −at one year, whereas among the remaining firms the spread was already \msmallPre{} points before | |
| 19 | −the launch and did not rise, the 2022 cohort was negative, a placebo over ten candidate break | |
| 20 | −years ranks 2022 \mplaceboRank{}th, and the monthly long--short earned no post-launch alpha. | |
| 21 | −Disclosed AI projects carry no premium at all; the language of annual reports carries the strongest. | |
| 8 | +quarterly and current SEC reports of \mnFirms{} U.S.\ listed firms over fifteen formation cohorts, | |
| 9 | +linked to forward returns whose timing is free of look-ahead, to factor data and to an | |
| 10 | +independent project-based measure of AI engagement, we have re-examined whether and when AI | |
| 11 | +language was associated with the cross-section of returns. The answer differs from the one we | |
| 12 | +gave for the S\&P~500 alone, and the revision that followed the review of this version differs | |
| 13 | +again from our first reading of the broad universe. In the broad market, AI language carried a | |
| 14 | +modest premium throughout 2010--2024 in regression form---\mfmComposW{} points per standard | |
| 15 | +deviation on winsorized returns, with an interval that excludes zero, and \mfmFullSectorW{} within | |
| 16 | +sectors---while the sort-based expression of the same premium, \mspreadFullW{} points between the | |
| 17 | +top and bottom quintiles, has an interval that covers zero and a median near zero; every level | |
| 18 | +is conditional on the firm surviving to October 2026. The post-ChatGPT increase that the index | |
| 19 | +displays, and that we replicate within the new panel, is confined to the largest thousand firms | |
| 20 | +and to the 2024 cohort: the formal large-versus-small contrast of \mcontrastLarge{} points does | |
| 21 | +not reject equality, and the event-study gradient that supported it under the single-index | |
| 22 | +market model is absent under beta-only, market-adjusted, matched and six-factor benchmarks. In | |
| 23 | +the market as a whole the post-launch spread is not distinguishable from the pre-launch one, the | |
| 24 | +2022--24 block is not an unusual three-cohort block, and the monthly long--short earned no | |
| 25 | +post-launch alpha. Disclosed AI projects show no detectable premium on a sparse, unvalidated | |
| 26 | +measure whose minimum detectable effects exceed the language premium; the annual report carries | |
| 27 | +at least as much of the signal as the other forms. | |
| 22 | 28 | |
| 23 | −Two messages emerge. The economic message is that the pricing of AI exposure in U.S.\ equities is | |
| 24 | −better described as a persistent, partly within-industry characteristic premium concentrated in | |
| 25 | −smaller firms, overlaid after 2022 by a sector-driven re-rating of the largest AI-exposed firms, | |
| 26 | −than as a market-wide regime change triggered by a consumer event. The methodological message is | |
| 27 | −that an index sample can manufacture a regime: the S\&P~500 is the one part of the market where | |
| 28 | −the premium was absent before 2022 and large afterward, and a study confined to it reads a | |
| 29 | −large-cap episode as a change in how the market prices disclosure. Reading the full universe, | |
| 30 | −and reading all three forms of filing, reverses that reading. | |
| 29 | +Two messages follow. The economic message is that the association between AI exposure and | |
| 30 | +U.S.\ equity returns is better described as a modest, regression-robust characteristic premium | |
| 31 | +in the broad market, overlaid after 2022 by an increase that is visible only among the largest | |
| 32 | +AI-exposed firms and in one cohort, than as a market-wide regime change triggered by a consumer | |
| 33 | +event; and that the measure behind it is proximity to a technology-and-infrastructure register | |
| 34 | +rather than validated AI activity. The methodological message is that an index sample, a | |
| 35 | +current-membership universe, a bear-market event-study benchmark and independence-based inference | |
| 36 | +each flatter a post-event result in the same direction, and that a regime claim built on three | |
| 37 | +post-event cohorts should be accompanied by the tests that this revision added: cohort-level | |
| 38 | +intervals, two-way clustering and wild-cluster bootstraps, permutation and sup-Wald break tests, | |
| 39 | +tested contrasts with minimum detectable effects, multiple-testing adjustments by family, placebo | |
| 40 | +launches and several event-study benchmarks. | |
| 31 | 41 | |
| 32 | −Several next steps follow. The most important is to add post-2024 cohorts as they accrue, which is | |
| 33 | −the only way to learn whether the large-cap re-rating of 2024 was the start of a regime or an | |
| 34 | −episode. The second is a point-in-time Russell membership, which would remove the survivorship | |
| 35 | −that is more severe among small firms. The third is to validate the SPID project extraction | |
| 36 | −against human labels so that the null result on disclosed projects can be attributed to the market | |
| 37 | −rather than to measurement. Beyond these, earnings-call transcripts and patent text would sharpen | |
| 38 | −the distinction between narrative positioning and real AI activity that the contrast between | |
| 39 | −language and projects has opened. | |
| 42 | +Several next steps follow, in the order of their likely effect on the conclusions. The first is a | |
| 43 | +point-in-time universe with scored filings of delisted firms, which would turn the survivorship | |
| 44 | +bounds of this paper into an estimate. The second is to add post-2024 cohorts as they accrue, | |
| 45 | +which is the only way to learn whether the large-cap increase of 2024 was the start of a regime | |
| 46 | +or an episode. The third is to code the validation samples that have been drawn, so that the | |
| 47 | +embedding score and the project extraction acquire measured precision, and to re-embed the | |
| 48 | +corpus with chunks that fit the encoder's window. Beyond these, earnings-call transcripts and | |
| 49 | +patent text would sharpen the distinction between narrative positioning and real AI activity that | |
| 50 | +the contrast between language and projects has opened. | |
modified
paper/sections/discussion.tex
+177 −85
@@ -1,100 +1,192 @@ | ||
| 1 | 1 | % ============================================================================ |
| 2 | −\section{Discussion: Mechanism and Interpretation} | |
| 2 | +% Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, sections/macros_stageE.tex. | |
| 3 | +\section{Discussion: Interpretation, Hypotheses and Limitations} | |
| 3 | 4 | \label{sec:discussion} |
| 4 | 5 | % ============================================================================ |
| 5 | 6 | |
| 6 | −What economic mechanism fits this constellation of facts? The first version of this paper, on the | |
| 7 | −index alone, settled on a sector-level re-rating of persistent AI exposure after 2022. The broad | |
| 8 | −universe keeps that element and adds two that the index could not reveal. | |
| 7 | +The designs of this paper establish associations and their timing, not causal effects of | |
| 8 | +disclosure on prices, and the paragraphs below are written in that register. The first version | |
| 9 | +of this paper, on the index alone, settled on a sector-level re-rating of persistent AI exposure | |
| 10 | +after 2022. The broad universe keeps part of that element and changes the rest. | |
| 9 | 11 | |
| 10 | −\paragraph{A persistent, partly within-industry premium in the broad market.} In the Russell~3000 | |
| 11 | −the relation between AI language and subsequent returns is not a post-2022 phenomenon. The | |
| 12 | −composite spread is positive in three quarters of the pre-launch cohorts, the full-sample | |
| 13 | −Fama--MacBeth slope is significant with and without sector dummies, the sector-by-year fixed-effects | |
| 14 | −slope is significant over the full sample, and the premium is largest among the smaller firms that | |
| 15 | −the index excludes. The 2018--2019 cohorts, held through the first wave of enterprise machine | |
| 16 | −learning and cloud adoption, earned spreads as large as the post-launch ones. A reader of the | |
| 17 | −S\&P~500 study would have concluded that AI language was ``dormant for a decade''; the broad market | |
| 18 | −says that it was priced, modestly and intermittently, throughout the decade, among the firms whose | |
| 19 | −AI exposure was least visible to the market at large. This is consistent with the disclosure-pricing | |
| 20 | −literature's finding that qualitative filing content is processed incompletely and with delay | |
| 21 | −\citep{cohen2020lazy,hobergphillips2016}, and with the attention literature's prediction that such | |
| 22 | −processing is slowest for small, lightly covered firms \citep{hirshleifer2009driven,dellavigna2009}. | |
| 12 | +\paragraph{A modest association in the broad market, robust in regression form and fragile in | |
| 13 | +sort form.} In the Russell~3000 the relation between AI language and subsequent returns is not a | |
| 14 | +post-2022 phenomenon. The composite spread is positive in three quarters of the pre-launch | |
| 15 | +cohorts, the full-sample Fama--MacBeth slope is \mfmComposW{} points per standard deviation with | |
| 16 | +an interval that excludes zero and \mfmFullSectorW{} with sector dummies, and the sector-by-year | |
| 17 | +pooled slope is \msyFull{}. The sort evidence is weaker: the winsorized quintile spread's interval | |
| 18 | +covers zero, its median is near zero, the sector-neutral sort does not reject zero, and no | |
| 19 | +cohort-level statistic survives the multiple-testing adjustment within its family. A reader of | |
| 20 | +the S\&P~500 study would have concluded that AI language was ``dormant for a decade''; the broad | |
| 21 | +market says that it was associated with returns, modestly and intermittently, throughout the | |
| 22 | +decade in the regression sense. Every level behind that statement is conditional on survival to | |
| 23 | +October 2026, and the selection is strongest among the small firms where the first version | |
| 24 | +located the premium. The association is consistent with the disclosure-pricing literature's | |
| 25 | +finding that qualitative filing content is processed incompletely and with delay | |
| 26 | +\citep{cohen2020lazy,hobergphillips2016}, and with slower diffusion among small, lightly covered | |
| 27 | +firms \citep{honglimstein2000,hirshleifer2009driven}, but the paper has no attention proxy and | |
| 28 | +cannot attribute it to either. | |
| 23 | 29 | |
| 24 | −\paragraph{A large-cap re-rating after 2022, not a market-wide regime change.} The post-launch | |
| 25 | −increase that the index sample displays is real within the largest thousand firms and absent | |
| 26 | −outside them. Among large firms the spread triples after the launch, the one-year event-study | |
| 27 | −difference is \mevLarge{} points and highly significant, and the 2024 cohort---held through the | |
| 28 | −memory, storage and data-center rally of 2025--2026---dominates. Among smaller firms the spread does | |
| 29 | −not rise, the event-study difference is half as large and insignificant, and the long--short Sharpe | |
| 30 | −ratio falls after the launch. The generative-AI boom was, to first order, a sectoral and | |
| 31 | −large-cap reallocation of expected cash flows: the firms that were re-rated were the suppliers of | |
| 32 | −the AI build-out and the largest users, which are disproportionately large, and the re-rating | |
| 33 | −operated through sectors, as the sector-by-year fixed effects show. An attention-based account of | |
| 34 | −the post-launch period would predict the opposite size gradient---the largest effect where | |
| 35 | −attention frictions are largest---and is not supported. | |
| 30 | +\paragraph{A post-launch increase confined to the largest firms and to one cohort.} The increase | |
| 31 | +that the index sample displays is visible in the point estimates for the largest thousand firms | |
| 32 | +and absent outside them, and the test of the difference does not reject equality. Within the | |
| 33 | +largest firms the winsorized spread moves from \mlargePreW{} to \mlargePostW{} points, but it is | |
| 34 | +\mlargePostExTwentyFourW{} without the 2024 cohort---held through the memory, storage and | |
| 35 | +data-center rally of 2025--2026---and the high-AI $\times$ post $\times$ large contrast has a | |
| 36 | +two-way $t$ of \mcontrastLargeTtw{} against a minimum detectable effect of \mcontrastLargeMDE{} | |
| 37 | +points. The event study agrees under the single-index market model and disagrees under four | |
| 38 | +other benchmarks. What the evidence supports is a statement about where the post-launch increase | |
| 39 | +is \emph{visible}: among the largest, most-followed AI-exposed firms, in one cohort, and under | |
| 40 | +the benchmark that rewards the 2022 intercept. The first version's reading of a ``sectoral and | |
| 41 | +large-cap reallocation of expected cash flows'' and its identification of the re-rated firms as | |
| 42 | +``the suppliers of the AI build-out'' were mechanism claims without a test and are withdrawn. | |
| 43 | +An attention-based account of the post-launch period would predict the opposite size | |
| 44 | +gradient---the largest effect where attention frictions are largest---and is not supported by the | |
| 45 | +point estimates; nor can it be rejected by a contrast with this little power. | |
| 36 | 46 | |
| 37 | −\paragraph{Language, not projects.} The sharpest new fact is that the premium attaches to the | |
| 38 | −\emph{language} of AI exposure and not to disclosed AI \emph{projects}. Firms that announce AI | |
| 39 | −initiatives, data centers or cloud migrations earn no premium on that account, before or after the | |
| 40 | −launch, while firms whose filings resemble the AI anchors do. One interpretation is that the | |
| 41 | −embedding measures a firm's position in the AI value chain---its vocabulary of compute, | |
| 42 | −infrastructure and disruption---which is what the market priced, whereas project announcements are | |
| 43 | −a mixture of leaders and laggards catching up. Another is that project counts are too sparse to | |
| 44 | −carry a signal. Both interpretations warn against reading the semantic premium as a premium on AI | |
| 45 | −investment: it is a premium on being, or sounding like, an AI-exposed firm. | |
| 47 | +\paragraph{Language rather than projects, at this precision.} The premium attaches to the | |
| 48 | +\emph{language} of AI exposure, and no premium is detectable on disclosed AI \emph{projects}: the | |
| 49 | +binary contrasts of \cref{subsec:spidres} have intervals that comfortably include the language | |
| 50 | +premium, the project measure is sparse and unvalidated, and the 24-month reversal test shows | |
| 51 | +neither a reversal for high-language firms without projects nor a premium for those with them. | |
| 52 | +One interpretation is that the embedding measures a firm's position in the AI value chain---its | |
| 53 | +vocabulary of compute, infrastructure and disruption---and that this is what the market priced, | |
| 54 | +whereas project announcements mix leaders and laggards; another is that project counts are too | |
| 55 | +sparse to carry a signal; a third, from the AI-washing literature | |
| 56 | +\citep{erdogan2026aiwashing,wang2026aiwashing}, is that the market priced the framing of a firm | |
| 57 | +as AI-exposed. The present evidence does not separate the three. What it does say is that the | |
| 58 | +semantic premium should not be read as a premium on AI investment. | |
| 46 | 59 | |
| 47 | −\paragraph{Where the signal lives.} The annual report carries the strongest form of the signal and | |
| 48 | −the current report the weakest, by a wide margin. The first version of this paper, which read | |
| 49 | −current reports only, therefore understated the pre-launch premium and overstated the contrast with | |
| 50 | −the post-launch period. Embedding-based measures inherit the composition of the corpus they are | |
| 51 | −computed on; documenting that composition, and re-estimating on each form, is part of the method. | |
| 60 | +\paragraph{What the measure is.} The construct-validity diagnostics of \cref{subsec:construct} | |
| 61 | +qualify every use of the phrase ``AI language'' in this paper. The encoder reads the first | |
| 62 | +\mTruncWindowWords{} or so words of each \mChunkWordsMedian-word chunk; the score correlates | |
| 63 | +\mKwRho{} with a keyword count; two electric utilities rank in the top dozen because | |
| 64 | +data-center-load and generation-capacity vocabulary scores highly on the infrastructure anchor; | |
| 65 | +and negated AI sentences score only moderately below affirmative ones. The results are robust to | |
| 66 | +anchors stripped of the cybersecurity and automation vocabulary and to the exclusion of Utilities | |
| 67 | +and Energy, so they are not an artefact of the shared vocabulary, but the measure is best | |
| 68 | +described as proximity to a technology-and-infrastructure register in the opening of each filing | |
| 69 | +passage. In the terms of \citet{eisfeldt2026generative} and \citet{song2026productexposure}, the | |
| 70 | +adoption anchor is closest to an operating exposure, the infrastructure anchor to a supplier or | |
| 71 | +user position in the AI build-out that neither labour nor product measures target, and the | |
| 72 | +disruption-risk anchor to product exposure. The disruption-risk facet carries the largest spread | |
| 73 | +and the only post-launch six-factor alpha above conventional thresholds in the broad universe, | |
| 74 | +with the opposite sign to the revenue declines that Song documents for product-exposed firms; | |
| 75 | +whether this reflects risk compensation for product exposure or the fact that firms which discuss | |
| 76 | +AI as a threat to their products are also the firms that build with it cannot be decided here. | |
| 52 | 77 | |
| 53 | −\paragraph{Relation to the AI-asset-pricing literature.} The findings are complementary to | |
| 54 | −\citet{eisfeldt2023generative}, who document a sharp repricing of an occupation-based | |
| 55 | −generative-AI exposure measure in the weeks around the launch, and to \citet{babina2024aiinvestment}, | |
| 56 | −who measure AI investment from labour data. Our disclosure-language measure shows no repricing | |
| 57 | −within sixty days, a gradual one over a year among large firms, and a persistent premium among | |
| 58 | −small firms that predates the launch by years. The three measures capture different objects---labour | |
| 59 | −exposure, investment, and narrative positioning---and the differences in their timing and in their | |
| 60 | −size gradients are themselves informative about how the market priced the AI theme. | |
| 78 | +\paragraph{Where the signal lives.} The annual report carries at least as much of the priced AI | |
| 79 | +language as the quarterly and current reports, and the current report the least; the first | |
| 80 | +version of this paper, which read current reports only, therefore measured the weakest component | |
| 81 | +of the signal. Whether the forms differ statistically depends on the specification. Embedding | |
| 82 | +measures inherit the composition of the corpus they are computed on; documenting that | |
| 83 | +composition, and re-estimating on each form, is part of the method. | |
| 61 | 84 | |
| 62 | 85 | \subsection{What the evidence says about the hypotheses} |
| 63 | −H1 (unconditional pricing) is weakly supported in the broad universe: the full-sample spread and | |
| 64 | −slope are positive and significant at conventional levels, but the risk-adjusted alpha is marginal | |
| 65 | −and partly a technology tilt. H2 (regime dependence) is rejected for the market and supported only | |
| 66 | −within the largest firms, where it rests on one cohort. H3 (AI specificity) holds against ESG, | |
| 67 | −climate and macro language but not against cybersecurity and automation language. H4 | |
| 68 | −(cross-sectional, not within-firm) is supported: the premium is absorbed by firm fixed effects and | |
| 69 | −the pre-determined-exposure design shows no post-launch break. H5 (size and attention) is | |
| 70 | −rejected in its attention form---the post-launch re-rating is larger among large firms---but the | |
| 71 | −broad universe does reveal a persistent small-firm premium that the index concealed, and the | |
| 72 | −project-based measure carries no premium. | |
| 86 | +\label{subsec:verdicts} | |
| 87 | +\Cref{tab:verdicts} records the verdict on each hypothesis, clause by clause. H1 is supported in | |
| 88 | +regression form and not in sort form, and only conditionally on survival. H2 is not supported | |
| 89 | +for the market and is visible only within the largest firms, where it rests on one cohort. H3a | |
| 90 | +holds against ESG, climate and macro language and fails against cybersecurity; H3b fails, because | |
| 91 | +the post-launch long--short is spanned by the technology-sector factor. H4 is not supported: the | |
| 92 | +association is cross-sectional and absorbed by firm fixed effects, but the corrected event-time | |
| 93 | +profile shows no positive post-launch shift and the pre-launch years cannot serve as a | |
| 94 | +parallel-trends test. H5a is not supported by the point estimates, which show the larger | |
| 95 | +post-launch increase among large firms, and the formal contrast cannot reject equality in either | |
| 96 | +direction; H5b is undecidable on the project measure at its current precision. The first version | |
| 97 | +had recorded H4 as supported and H5 as rejected in its attention form; both verdicts are revised. | |
| 98 | + | |
| 99 | +\begin{table}[htbp]\centering\small | |
| 100 | +\caption{\textbf{Hypotheses and verdicts on the Russell~3000 evidence.} Clause-by-clause reading | |
| 101 | +of \cref{subsec:hypotheses}; the evidence column names the subsection where the statistic is | |
| 102 | +reported.}\label{tab:verdicts} | |
| 103 | +\begin{tabular}{@{}p{2.4cm}p{6.1cm}p{2.6cm}p{4.3cm}@{}} | |
| 104 | +\toprule | |
| 105 | +Hypothesis & Clause & Verdict & Evidence \\ | |
| 106 | +\midrule | |
| 107 | +H1 & Positive full-sample spread in sort form that survives winsorization & Not established & \cref{subsec:sorts}: winsorized interval covers zero; median near zero \\ | |
| 108 | + & Positive full-sample slope in regression form & Supported, conditional on survival & \cref{subsec:fm}: slope \mfmComposW{} (interval \mfmComposWCI) \\ | |
| 109 | +\addlinespace | |
| 110 | +H2 & Spread small before, large after the launch, in the market & Not supported & \cref{subsec:break}: pooled post$-$pre \mpooledDiffW{} ($t=\mpooledDiffWTtw$ two-way) \\ | |
| 111 | + & Increase visible in each post-launch cohort & Not supported & 2022 cohort negative; post mean rests on 2024 \\ | |
| 112 | + & 2022 stands out among break dates & Not supported & permutation $p=\mpermPW$; sup-Wald $p=\msupWaldWP$ \\ | |
| 113 | +\addlinespace | |
| 114 | +H3 & (a) Post-launch spread for AI factors, not for ESG, climate, macro & Supported for ESG, climate and macro; fails for cybersecurity & \cref{subsec:placebo}: only cybersecurity survives two-way clustering \\ | |
| 115 | + & (b) Not spanned by a technology-sector factor & Not supported & \cref{subsec:sector}: $\beta_{\text{TECH}}$ \mtechBetaPost{} post-launch; alpha falls to \mffsixTechPost\% \\ | |
| 116 | +\addlinespace | |
| 117 | +H4 & Association cross-sectional, attenuates under firm fixed effects & Supported & \cref{subsec:fm}: within-firm slope \mfeCompos{} \\ | |
| 118 | + & No differential pre-trend, positive post-launch shift of pre-determined exposure & Not supported & \cref{subsec:did}: post years below 2021; pre-years are look-back, not a trend test \\ | |
| 119 | +\addlinespace | |
| 120 | +H5 & (a) Post-launch increase larger among smaller firms & Not supported by point estimates; contrast cannot reject equality & \cref{subsec:size}: contrast \mcontrastLarge{} ($t=\mcontrastLargeTtw$, MDE \mcontrastLargeMDE) \\ | |
| 121 | + & (b) Premium visible in disclosed AI projects & Undecidable on this measure & \cref{subsec:spidres}: MDE \mspidAnyFirmFullMDE{} points; measure unvalidated \\ | |
| 122 | +\bottomrule | |
| 123 | +\end{tabular} | |
| 124 | +\end{table} | |
| 73 | 125 | |
| 74 | 126 | \subsection{Implications} |
| 75 | −For investors, the AI-language signal is a modest, persistent characteristic in the broad market, | |
| 76 | −strongest among small firms and inside industries, and a large-cap, sector-driven bet after 2022; | |
| 77 | −it is not a stable source of alpha, and the high post-launch Sharpe ratios of the index sample are | |
| 78 | −a large-cap episode. For corporate managers, the asymmetry between AI and sustainability language | |
| 79 | −illustrates how the prevailing narrative conditions the valuation consequences of what firms | |
| 80 | −emphasize, while the absence of any within-firm effect and of any premium on announced projects | |
| 81 | −means that neither emphasis nor announcement is rewarded as such. For researchers, the study shows | |
| 82 | −that a conclusion drawn from an index sample---here, a dormant-then-priced regime story---can be | |
| 83 | −an artefact of the index's size composition, and that measurement choices about which filings to | |
| 84 | −read can change the sign of a comparison. | |
| 127 | +For investors, the AI-language signal is a modest, regression-robust characteristic in the broad | |
| 128 | +market whose sort-based expression is fragile; it is not a stable source of alpha, the high | |
| 129 | +post-launch Sharpe ratios of the index sample are a large-cap, one-cohort episode, and the only | |
| 130 | +long--short with a post-launch six-factor alpha above conventional thresholds is the | |
| 131 | +disruption-risk facet, whose deflated Sharpe ratio is modest. For corporate managers, the | |
| 132 | +asymmetry between AI and sustainability language illustrates how the prevailing narrative | |
| 133 | +conditions the valuation consequences of what firms emphasize, while the absence of any | |
| 134 | +within-firm association and of any detectable premium on announced projects means that neither | |
| 135 | +emphasis nor announcement is rewarded as such at the precision of these tests. For researchers, | |
| 136 | +the study shows that a conclusion drawn from an index sample---here, a dormant-then-priced regime | |
| 137 | +story---need not hold in the market the index is drawn from; that measurement choices about | |
| 138 | +which filings to read and how much of each passage the encoder sees change the size of a | |
| 139 | +comparison; and that a current-membership universe, a bear-market event-study benchmark and | |
| 140 | +independence-based inference each flatter a post-event result in the same direction. | |
| 85 | 141 | |
| 86 | −\subsection{Limitations} | |
| 142 | +\subsection{Limitations and what was not done} | |
| 87 | 143 | \label{subsec:limits} |
| 88 | −\emph{First}, the post-launch period contains three cohorts and the last ends in June 2026; the | |
| 89 | −large-cap result rests mainly on the 2024 cohort. \emph{Second}, the universe is a current | |
| 90 | −holdings snapshot, which omits delisted firms and biases the level of small-firm returns upward; | |
| 91 | −restrictions on listing history leave the results intact, but a point-in-time Russell membership | |
| 92 | −has not been reconstructed. \emph{Third}, AI language is correlated with sector, and after 2022 | |
| 93 | −every sector control leaves nothing within sectors; the within-sector premium is a full-sample | |
| 94 | −fact, not a post-launch one. \emph{Fourth}, the semantic anchors embed the authors' definitional | |
| 95 | −choices and no alternative anchor set has been tried. \emph{Fifth}, the SPID project counts are | |
| 96 | −extracted by a language model whose precision and recall have not yet been validated against human | |
| 97 | −labels (UQO Working Paper No.~0001), so the null result on projects is a null result on a noisy | |
| 98 | −measure. \emph{Sixth}, the event-study horizons are long for a market-model design and the | |
| 99 | −parallel-trends assumption is not met. \emph{Finally}, the designs establish associations and | |
| 100 | −their timing, not causality. | |
| 144 | +The limitations below are specific; each names what was not done, why, and the direction in | |
| 145 | +which the omission biases the stated results. | |
| 146 | +\begin{enumerate}[leftmargin=2em,itemsep=2pt] | |
| 147 | +\item \emph{Survivorship.} The universe is the October-2026 IWV holdings. The selection was | |
| 148 | +measured (\cref{subsec:survivorship}) but not removed: a point-in-time Russell membership with | |
| 149 | +delisting returns and scored filings of delisted firms requires a new embedding pass of the | |
| 150 | +delisted firms' filings, which the authors did not authorize for this revision. Every level is | |
| 151 | +biased upward; the sign of the bias on the Q5$-$Q1 spread is unknown and plausibly positive, | |
| 152 | +because top-quintile firms have higher forward volatility. Attrition by AI quintile cannot be | |
| 153 | +computed on this corpus for the same reason. | |
| 154 | +\item \emph{Encoder truncation.} The encoder read about \mTruncWindowWords{} of the | |
| 155 | +\mChunkWordsMedian{} words of a typical chunk. Re-chunking the corpus to the encoder's window, or | |
| 156 | +a second encoder, requires a new embedding pass and was not executed. The score is a measure of | |
| 157 | +the opening of each passage; the direction of the bias on the return results is unknown, and the | |
| 158 | +keyword baseline and alternative anchors of \cref{subsec:construct} bound only the vocabulary | |
| 159 | +dependence, not the truncation. | |
| 160 | +\item \emph{No human-coded validation.} A stratified sample of chunks (by score decile and cohort, | |
| 161 | +plus the top chunks of the top and median firms of \cref{tab:topfirms}) with a codebook was | |
| 162 | +prepared, and a stratified sample of \mSpidSampleN{} SPID project mentions with their source text | |
| 163 | +was drawn (\cref{app:spid}); neither has been coded by human readers, and a language-model second | |
| 164 | +rater was not authorized. No label of either kind exists, so no precision, recall or inter-rater | |
| 165 | +statistic is reported, and the attenuation of the project-count slope cannot be corrected. The | |
| 166 | +project null therefore bounds the premium on the measure, not on projects. | |
| 167 | +\item \emph{Three post-launch cohorts.} The post-period is three annual observations, the last | |
| 168 | +ending in June 2026; the large-cap increase rests on one of them, the formal contrasts have | |
| 169 | +minimum detectable effects above ten points, and no post-period statistic survives a | |
| 170 | +multiple-testing adjustment. Adding cohorts as they accrue is the only remedy. | |
| 171 | +\item \emph{Benchmark dependence of the event study.} The size gradient of the event study is | |
| 172 | +present under the single-index model estimated in 2022 and absent under four other benchmarks; a | |
| 173 | +book-to-market dimension could not be added to the matched benchmark because book equity is not | |
| 174 | +in the panel. The reported gradient is an upper bound. | |
| 175 | +\item \emph{Sector and construct.} AI language is correlated with sector and with cybersecurity | |
| 176 | +and automation vocabulary; the within-sector association survives in regression form only, and | |
| 177 | +the anchors were written in 2026 and applied to 2010 filings. Two alternative anchor sets were | |
| 178 | +tested against the stored embeddings; no alternative chunking or encoder was. | |
| 179 | +\item \emph{Size partition.} Size groups use a top-1,000 cut and a fixed within-cohort percentile; | |
| 180 | +a partition by inflation-adjusted Russell~1000 dollar cut-offs was not implemented because the | |
| 181 | +reconstitution breakpoints were not verified from a public source. The two implemented partitions | |
| 182 | +agree. | |
| 183 | +\item \emph{Ambiguous 2021 cohort.} The pre-launch premium depends on where the 2021 cohort, whose | |
| 184 | +holding period straddles the launch, is placed (\cref{subsec:breakdef}). | |
| 185 | +\item \emph{Data access.} Returns were verified and capitalizations rebuilt with a licensed daily | |
| 186 | +price source that cannot be redistributed; the Yahoo-implied fallback reproduces the price months | |
| 187 | +within two percent in almost every case, and the derived panel is documented by checksum but not | |
| 188 | +yet deposited (\cref{app:repro}). | |
| 189 | +\item \emph{Register.} The designs establish associations and their timing. ``Re-rating'', | |
| 190 | +``repricing'' and ``rewarded'' are used descriptively; no design identifies a causal effect of | |
| 191 | +disclosure on returns, and the title's verb should be read in the same way. | |
| 192 | +\end{enumerate} | |
| 101 | 193 | |