TeX 69.3%
Python 30.6%
1\documentclass[11pt]{article}2\usepackage[margin=1in]{geometry}3\usepackage{graphicx,booktabs,amsmath,amssymb,siunitx,caption,float}4\usepackage[hidelinks]{hyperref}5\usepackage{xcolor,sectsty,setspace}6\usepackage{titling}7\definecolor{navy}{HTML}{1b3a6b}8\sectionfont{\color{navy}}\subsectionfont{\color{navy}}9\captionsetup{font=small,labelfont={bf,color=navy}}10\graphicspath{{../figures/}}11\setlength{\parskip}{0.45em}\setlength{\parindent}{0pt}12\sisetup{group-minimum-digits=4,group-separator={,}}13\newcommand{\tstat}[1]{{\footnotesize(\textit{t}=#1)}}1415\title{\color{navy}\bfseries Does the Language of Artificial Intelligence Move Markets?\\[2mm]16\large A Semantic Analysis of S\&P~500 SEC Filings and Subsequent Stock Returns, 2010--2026}17\author{EDGAR Research Database Project\\18\small Manel Kammoun \quad Djerrycharli Tandjambianda \quad Simon-Pierre Boucher\\19\small Universit\'e du Qu\'ebec en Outaouais (UQO)}20\date{June 2026}2122\begin{document}23\maketitle2425\begin{abstract}26\noindent We measure the \emph{semantic intensity} of artificial-intelligence (AI) language in the27annual SEC filings (10-K) of S\&P~500 constituents using sentence-embedding cosine28similarity to curated AI ``anchor'' concepts, and ask whether this disclosure29intensity is related to the cross-section of subsequent stock returns. Combining a30panel of 7{,}869 firm-years (500 firms, fiscal years 2010--2026) with monthly31prices from Yahoo Finance and the Fama--French factor data, we form strictly32tradeable forward 12-month returns and study them with portfolio sorts,33factor-model spanning regressions, two-way fixed-effects panels, Fama--MacBeth34regressions, and a structural-break test around the launch of ChatGPT35(November~2022). Our central finding is \textbf{conditional}: over the full sample36the relation between AI-disclosure intensity and future returns is positive but37economically modest and only marginally significant; however, it \emph{more than38sextuples} after 2022. Firms in the top quintile of AI-disclosure intensity39outperformed the bottom quintile by \textbf{+2.4 percentage points} per year before40ChatGPT but by \textbf{+14.6 percentage points} per year afterward. The effect is a41\emph{cross-sectional, regime-dependent re-rating}---it is absorbed by firm fixed42effects, indicating the market repriced which \emph{firms} were AI-exposed rather43than rewarding year-to-year changes in any single firm's wording.44\end{abstract}4546\section{Introduction}47The diffusion of generative AI has been the defining corporate-narrative shift of48the 2020s. A natural question for investors and researchers is whether the49\emph{disclosure} of AI engagement---how intensively a firm writes about adopting,50building, or being threatened by AI in its mandatory filings---carries information51about future equity performance. This study answers that question for the S\&P~50052using a purpose-built semantic measurement of filing text rather than naive keyword53counts.5455We exploit the EDGAR Research Database, which parses every 10-K/10-Q/8-K filing into56sections, embeds each text chunk with the \texttt{all-MiniLM-L6-v2} sentence57transformer, and scores it by cosine similarity to professionally curated anchor58passages describing eight economic concepts. Three of these capture distinct facets59of AI exposure:60\begin{itemize}\setlength\itemsep{0.1em}61 \item \textbf{AI~adoption} --- internal use of AI/ML to improve operations;62 \item \textbf{AI~infrastructure} --- capital exposure to AI compute (GPUs, data centers, cloud);63 \item \textbf{AI~disruption risk} --- the risk that AI commoditizes the firm's products.64\end{itemize}65We aggregate chunk scores to a section-weighted annual firm score, link them to66realized returns, and subject the relation to a battery of asset-pricing tests.6768Our contribution is threefold. First, we move beyond keyword dictionaries to a69\emph{semantic} measure of AI exposure that captures meaning, not just term70frequency. Second, we provide a clean, look-ahead-free mapping from disclosure to71\emph{tradeable} forward returns. Third, and most importantly, we document that the72pricing of AI language is \emph{not} a stable factor premium but a73\textbf{regime-dependent re-rating} that switched on with the generative-AI boom.7475\section{Data}76\paragraph{Semantic factors.} The firm-year panel contains 7{,}86977observations covering all 500 current S\&P~500 members across fiscal years782010--2026. For each of the eight anchor concepts the database reports the mean,79maximum, standard deviation, top-5\% tail, and section-weighted mean80(\texttt{wmean}) of chunk-level cosine similarities, together with first/second81differences and rolling statistics. We use the section-weighted mean as our82headline signal because it up-weights the Risk-Factors and MD\&A sections where83forward-looking strategic content concentrates.8485\paragraph{Prices and factors.} We download monthly adjusted close prices for all86500 tickers from Yahoo Finance (2009--2026), the SPDR S\&P~500 ETF (SPY) as the87market benchmark, and the Fama--French five factors plus momentum and the risk-free88rate from the Kenneth French data library.8990\paragraph{Timing.} A 10-K covering fiscal year $Y$ is public by roughly91February--March of $Y{+}1$. To eliminate look-ahead bias we form signals at the92\textbf{end of June of $Y{+}1$} and measure the \textbf{forward 12-month return over93July~$Y{+}1$ to June~$Y{+}2$}---the Fama--French portfolio convention. The signal is94therefore always known before the holding period begins. This yields 6{,}85295firm-years with a tradeable forward return spanning formation years 2010--202496(return realizations through mid-2026).9798\paragraph{Variables and treatment.} For each formation year we compute99cross-sectional $z$-scores of every signal (so coefficients are per-standard-deviation100and the secular rise in AI language is differenced out), plus controls for size101(log market-cap proxy), 12-month momentum, and trailing volatility. Returns and102controls are winsorized at the 1st/99th percentiles in regressions. Table~\ref{tab:summ}103reports summary statistics.104105\input{tab_summary}106107A first caveat is visible in the correlation structure: the three AI signals are108highly collinear ($\rho = 0.75$--$0.86$). They are best read as three lenses on a109single latent construct---\emph{AI salience} in corporate disclosure---rather than as110independent factors. This collinearity matters for interpreting multivariate111specifications below.112113\section{The Rise of AI Language}114Before turning to returns we document the raw phenomenon. Figure~\ref{fig:trend}115shows that the cross-sectional mean similarity to all three AI concepts rose116steadily over the decade and accelerated sharply after 2022: by 2025--2026, AI117language saturates S\&P~500 disclosure. This secular trend is precisely why our118tests rely on \emph{within-year cross-sectional} ranking---in any given year we ask119which firms talk about AI \emph{more than their peers}, not whether the index as a120whole talks about AI more than it used to.121122\begin{figure}[H]\centering123\includegraphics[width=0.92\textwidth]{fig4_ai_language_trend.png}124\caption{Cross-sectional mean of section-weighted cosine similarity to each AI anchor,125by fiscal year. All three measures trend upward and inflect after the 2022 launch of126ChatGPT (dashed line).}\label{fig:trend}127\end{figure}128129\section{Results}130\subsection{Portfolio sorts}131Each formation year we sort firms into quintiles on each AI signal and compute the132equal-weighted (EW) and value-weighted (VW) forward return of each quintile, plus133the high-minus-low (Q5$-$Q1) spread. Statistical significance uses Newey--West134(3-lag) $t$-statistics on the annual spread series. Table~\ref{tab:sorts} summarizes.135136\input{tab_sorts}137138Over the full sample the relation is monotone-ish and positive but economically139modest: the EW Q5$-$Q1 spread is +4.6~pp/yr for AI~adoption \tstat{1.83},140+3.9~pp/yr for AI~infrastructure \tstat{1.58}, and +5.3~pp/yr for AI~disruption141risk \tstat{2.27}. Value-weighting attenuates the adoption and composite spreads,142indicating the effect is stronger among smaller index members. The left panel of143Figure~\ref{fig:quint} shows the full-sample quintile pattern.144145\subsection{Does AI disclosure earn risk-adjusted alpha?}146We next build an annually rebalanced, equal-weighted, self-financing long-short147portfolio (top minus bottom quintile) and regress its monthly return on the CAPM and148on the Fama--French five-factor-plus-momentum (FF6) model. Table~\ref{tab:alpha}149reports intercepts (alphas).150151\input{tab_alpha}152153Unconditionally, the AI-opportunity long-short earns about +2.3\%/yr, of which the154FF6 alpha is +1.9\%/yr but statistically insignificant \tstat{1.24}: most of the raw155spread is compensation for known factors and noise. The one exception is the156\emph{AI-disruption-risk} spread, whose FF6 alpha of +2.8\%/yr is marginally157significant \tstat{2.10}---firms that flag AI as a competitive threat earned a small158premium, consistent with a risk-premium interpretation. The headline message of the159full-sample tests is therefore one of \emph{weak, mostly-not-significant} pricing.160161\subsection{Panel regressions and Fama--MacBeth}162Table~\ref{tab:panel} reports two-way (firm $\times$ year) fixed-effects panel163regressions of forward returns on the AI $z$-scores and controls, with standard164errors double-clustered by firm and year, alongside Fama--MacBeth estimates.165166\input{tab_panel}167168The two contrasting columns are instructive. In the \textbf{Fama--MacBeth}169cross-sectional regressions the AI-adoption and AI-infrastructure premia are positive170(+1.3 and +1.6~pp per standard deviation) with $t$-statistics near 1.6--1.7---weakly171positive pricing in the cross-section. In the \textbf{two-way fixed-effects} panel,172however, every AI coefficient collapses to essentially zero. Because firm fixed173effects remove all permanent cross-firm differences, this tells us the signal lives174\emph{across} firms, not \emph{within} them: a given firm writing relatively more175about AI from one year to the next does not predict its own subsequent return. The176market priced \emph{which firms are AI-exposed}, not the annual wording wiggle. (Size177carries the familiar negative premium throughout, $t\approx-8$.)178179\subsection{The structural break: AI pricing switched on after ChatGPT}180The full-sample averages mask a sharp regime change. Splitting at the November~2022181launch of ChatGPT, Figure~\ref{fig:quint} (center vs.\ right panel) shows the182quintile pattern is nearly flat before 2022 and steeply increasing afterward. The183annual Q5$-$Q1 spreads in Figure~\ref{fig:spread} are small and mixed-sign through1842021, then large and positive for every post-2022 formation year.185186\begin{figure}[H]\centering187\includegraphics[width=\textwidth]{fig1_quintile_returns.png}188\caption{Mean forward 12-month return by AI-disclosure composite quintile. The189high-minus-low spread is +2.4~pp before ChatGPT and +14.6~pp after.}\label{fig:quint}190\end{figure}191192\begin{figure}[H]\centering193\includegraphics[width=0.92\textwidth]{fig3_annual_spread.png}194\caption{Annual high-minus-low (Q5$-$Q1) forward-return spread on the AI-disclosure195composite, by 10-K fiscal year. Post-ChatGPT formation years (red) are uniformly and196substantially positive.}\label{fig:spread}197\end{figure}198199Table~\ref{tab:break} quantifies the break two ways: sub-period quintile spreads and200a pooled interaction of the AI $z$-score with a post-2022 indicator.201202\input{tab_break}203204The numbers are striking. The AI-adoption Q5$-$Q1 spread rises from +2.2\%/yr205\tstat{2.04} pre-2022 to \textbf{+14.3\%/yr} \tstat{2.15} post-2022; the composite206rises from +2.5\% to +14.6\%/yr. In the interaction regressions the post-2022207incremental slope is positive for all three signals and statistically significant for208AI~infrastructure \tstat{2.63}, the facet most directly tied to the firms (chipmakers,209cloud, data-center, and power names) that led the 2023--2025 rally. The cumulative210long-short wealth curve in Figure~\ref{fig:cum} makes the timing visible: the AI211portfolios drift sideways-to-up for a decade, then inflect upward after the dashed212ChatGPT line.213214\begin{figure}[H]\centering215\includegraphics[width=0.92\textwidth]{fig2_cumulative_ls.png}216\caption{Cumulative value of \$1 invested in the equal-weighted, annually217rebalanced, self-financing high-minus-low AI-disclosure portfolio.}\label{fig:cum}218\end{figure}219220Figure~\ref{fig:bin} confirms the pattern is not driven by outliers: the binned221relation between the AI composite and forward returns is flat pre-2022 and clearly222upward-sloping post-2022.223224\begin{figure}[H]\centering225\includegraphics[width=\textwidth]{fig5_binscatter.png}226\caption{Binned scatter of forward return against the AI-disclosure composite, by227sub-period. The conditional slope steepens markedly after ChatGPT.}\label{fig:bin}228\end{figure}229230\section{Discussion}231Three conclusions follow. \textbf{(1) AI disclosure is informative, but232conditionally.} It is not a stable, decade-long return factor; it is a signal that233the market began to price only once generative AI became salient. This is consistent234with an \emph{attention} or \emph{re-rating} mechanism: the textual evidence of AI235engagement existed for years, but capital flowed toward AI-exposed firms only after a236narrative catalyst. \textbf{(2) The effect is cross-sectional, not within-firm.} The237disappearance of the coefficient under firm fixed effects warns against trading238year-to-year changes in a single firm's AI verbosity; the durable signal is the239persistent \emph{level} of a firm's AI exposure relative to peers. \textbf{(3) The240three AI lenses largely coincide.} Their high collinearity means one should not241over-interpret which specific facet (adoption vs.\ infrastructure vs.\ risk)242``wins''; they jointly proxy a single AI-salience construct, and the infrastructure243lens simply loads most heavily on the firms that led the boom.244245\section{Limitations}246The post-2022 window contains only three formation years, so the regime result, while247economically large and robust across signals and specifications, rests on a short248post-period and should be read as suggestive rather than definitive. The market-cap249proxy holds shares outstanding at their latest value; survivorship is present because250we use current index membership; and the semantic anchors, though carefully curated,251embed the authors' definitional choices. Finally, disclosure intensity is correlated252with sector (Information Technology firms naturally score higher), so part of the253post-2022 premium reflects a sector tilt toward technology---though the within-year,254size- and momentum-controlled tests mitigate this.255256\section{Conclusion}257Using a semantic, embedding-based measure of AI language in S\&P~500 filings linked258to Yahoo Finance returns, we find that AI disclosure carried little pricing power for259most of the 2010s and then became strongly rewarded after the launch of ChatGPT. The260high-minus-low AI-disclosure return spread jumped from roughly +2~pp to roughly261+15~pp per year. The pattern is a regime-dependent, cross-sectional re-rating rather262than a persistent factor premium---evidence that markets price the \emph{language} of263AI, but only once AI commands their attention.264265\vspace{1em}\hrule\vspace{0.4em}266{\footnotesize\textit{Data:} EDGAR Research Database (UQO); prices from Yahoo Finance;267factors from the Kenneth R.\ French data library. All return tests use look-ahead-free268July-formation forward returns. Code and tables: \texttt{analysis/}.}269\end{document}270