\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{graphicx,booktabs,amsmath,amssymb,siunitx,caption,float}
\usepackage[hidelinks]{hyperref}
\usepackage{xcolor,sectsty,setspace}
\usepackage{titling}
\definecolor{navy}{HTML}{1b3a6b}
\sectionfont{\color{navy}}\subsectionfont{\color{navy}}
\captionsetup{font=small,labelfont={bf,color=navy}}
\graphicspath{{../figures/}}
\setlength{\parskip}{0.45em}\setlength{\parindent}{0pt}
\sisetup{group-minimum-digits=4,group-separator={,}}
\newcommand{\tstat}[1]{{\footnotesize(\textit{t}=#1)}}

\title{\color{navy}\bfseries Does the Language of Artificial Intelligence Move Markets?\\[2mm]
\large A Semantic Analysis of S\&P~500 SEC Filings and Subsequent Stock Returns, 2010--2026}
\author{EDGAR Research Database Project\\
\small Manel Kammoun \quad Djerrycharli Tandjambianda \quad Simon-Pierre Boucher\\
\small Universit\'e du Qu\'ebec en Outaouais (UQO)}
\date{June 2026}

\begin{document}
\maketitle

\begin{abstract}
\noindent We measure the \emph{semantic intensity} of artificial-intelligence (AI) language in the
annual SEC filings (10-K) of S\&P~500 constituents using sentence-embedding cosine
similarity to curated AI ``anchor'' concepts, and ask whether this disclosure
intensity is related to the cross-section of subsequent stock returns. Combining a
panel of 7{,}869 firm-years (500 firms, fiscal years 2010--2026) with monthly
prices from Yahoo Finance and the Fama--French factor data, we form strictly
tradeable forward 12-month returns and study them with portfolio sorts,
factor-model spanning regressions, two-way fixed-effects panels, Fama--MacBeth
regressions, and a structural-break test around the launch of ChatGPT
(November~2022). Our central finding is \textbf{conditional}: over the full sample
the relation between AI-disclosure intensity and future returns is positive but
economically modest and only marginally significant; however, it \emph{more than
sextuples} after 2022. Firms in the top quintile of AI-disclosure intensity
outperformed the bottom quintile by \textbf{+2.4 percentage points} per year before
ChatGPT but by \textbf{+14.6 percentage points} per year afterward. The effect is a
\emph{cross-sectional, regime-dependent re-rating}---it is absorbed by firm fixed
effects, indicating the market repriced which \emph{firms} were AI-exposed rather
than rewarding year-to-year changes in any single firm's wording.
\end{abstract}

\section{Introduction}
The diffusion of generative AI has been the defining corporate-narrative shift of
the 2020s. A natural question for investors and researchers is whether the
\emph{disclosure} of AI engagement---how intensively a firm writes about adopting,
building, or being threatened by AI in its mandatory filings---carries information
about future equity performance. This study answers that question for the S\&P~500
using a purpose-built semantic measurement of filing text rather than naive keyword
counts.

We exploit the EDGAR Research Database, which parses every 10-K/10-Q/8-K filing into
sections, embeds each text chunk with the \texttt{all-MiniLM-L6-v2} sentence
transformer, and scores it by cosine similarity to professionally curated anchor
passages describing eight economic concepts. Three of these capture distinct facets
of AI exposure:
\begin{itemize}\setlength\itemsep{0.1em}
  \item \textbf{AI~adoption} --- internal use of AI/ML to improve operations;
  \item \textbf{AI~infrastructure} --- capital exposure to AI compute (GPUs, data centers, cloud);
  \item \textbf{AI~disruption risk} --- the risk that AI commoditizes the firm's products.
\end{itemize}
We aggregate chunk scores to a section-weighted annual firm score, link them to
realized returns, and subject the relation to a battery of asset-pricing tests.

Our contribution is threefold. First, we move beyond keyword dictionaries to a
\emph{semantic} measure of AI exposure that captures meaning, not just term
frequency. Second, we provide a clean, look-ahead-free mapping from disclosure to
\emph{tradeable} forward returns. Third, and most importantly, we document that the
pricing of AI language is \emph{not} a stable factor premium but a
\textbf{regime-dependent re-rating} that switched on with the generative-AI boom.

\section{Data}
\paragraph{Semantic factors.} The firm-year panel contains 7{,}869
observations covering all 500 current S\&P~500 members across fiscal years
2010--2026. For each of the eight anchor concepts the database reports the mean,
maximum, standard deviation, top-5\% tail, and section-weighted mean
(\texttt{wmean}) of chunk-level cosine similarities, together with first/second
differences and rolling statistics. We use the section-weighted mean as our
headline signal because it up-weights the Risk-Factors and MD\&A sections where
forward-looking strategic content concentrates.

\paragraph{Prices and factors.} We download monthly adjusted close prices for all
500 tickers from Yahoo Finance (2009--2026), the SPDR S\&P~500 ETF (SPY) as the
market benchmark, and the Fama--French five factors plus momentum and the risk-free
rate from the Kenneth French data library.

\paragraph{Timing.} A 10-K covering fiscal year $Y$ is public by roughly
February--March of $Y{+}1$. To eliminate look-ahead bias we form signals at the
\textbf{end of June of $Y{+}1$} and measure the \textbf{forward 12-month return over
July~$Y{+}1$ to June~$Y{+}2$}---the Fama--French portfolio convention. The signal is
therefore always known before the holding period begins. This yields 6{,}852
firm-years with a tradeable forward return spanning formation years 2010--2024
(return realizations through mid-2026).

\paragraph{Variables and treatment.} For each formation year we compute
cross-sectional $z$-scores of every signal (so coefficients are per-standard-deviation
and the secular rise in AI language is differenced out), plus controls for size
(log market-cap proxy), 12-month momentum, and trailing volatility. Returns and
controls are winsorized at the 1st/99th percentiles in regressions. Table~\ref{tab:summ}
reports summary statistics.

\input{tab_summary}

A first caveat is visible in the correlation structure: the three AI signals are
highly collinear ($\rho = 0.75$--$0.86$). They are best read as three lenses on a
single latent construct---\emph{AI salience} in corporate disclosure---rather than as
independent factors. This collinearity matters for interpreting multivariate
specifications below.

\section{The Rise of AI Language}
Before turning to returns we document the raw phenomenon. Figure~\ref{fig:trend}
shows that the cross-sectional mean similarity to all three AI concepts rose
steadily over the decade and accelerated sharply after 2022: by 2025--2026, AI
language saturates S\&P~500 disclosure. This secular trend is precisely why our
tests rely on \emph{within-year cross-sectional} ranking---in any given year we ask
which firms talk about AI \emph{more than their peers}, not whether the index as a
whole talks about AI more than it used to.

\begin{figure}[H]\centering
\includegraphics[width=0.92\textwidth]{fig4_ai_language_trend.png}
\caption{Cross-sectional mean of section-weighted cosine similarity to each AI anchor,
by fiscal year. All three measures trend upward and inflect after the 2022 launch of
ChatGPT (dashed line).}\label{fig:trend}
\end{figure}

\section{Results}
\subsection{Portfolio sorts}
Each formation year we sort firms into quintiles on each AI signal and compute the
equal-weighted (EW) and value-weighted (VW) forward return of each quintile, plus
the high-minus-low (Q5$-$Q1) spread. Statistical significance uses Newey--West
(3-lag) $t$-statistics on the annual spread series. Table~\ref{tab:sorts} summarizes.

\input{tab_sorts}

Over the full sample the relation is monotone-ish and positive but economically
modest: the EW Q5$-$Q1 spread is +4.6~pp/yr for AI~adoption \tstat{1.83},
+3.9~pp/yr for AI~infrastructure \tstat{1.58}, and +5.3~pp/yr for AI~disruption
risk \tstat{2.27}. Value-weighting attenuates the adoption and composite spreads,
indicating the effect is stronger among smaller index members. The left panel of
Figure~\ref{fig:quint} shows the full-sample quintile pattern.

\subsection{Does AI disclosure earn risk-adjusted alpha?}
We next build an annually rebalanced, equal-weighted, self-financing long-short
portfolio (top minus bottom quintile) and regress its monthly return on the CAPM and
on the Fama--French five-factor-plus-momentum (FF6) model. Table~\ref{tab:alpha}
reports intercepts (alphas).

\input{tab_alpha}

Unconditionally, the AI-opportunity long-short earns about +2.3\%/yr, of which the
FF6 alpha is +1.9\%/yr but statistically insignificant \tstat{1.24}: most of the raw
spread is compensation for known factors and noise. The one exception is the
\emph{AI-disruption-risk} spread, whose FF6 alpha of +2.8\%/yr is marginally
significant \tstat{2.10}---firms that flag AI as a competitive threat earned a small
premium, consistent with a risk-premium interpretation. The headline message of the
full-sample tests is therefore one of \emph{weak, mostly-not-significant} pricing.

\subsection{Panel regressions and Fama--MacBeth}
Table~\ref{tab:panel} reports two-way (firm $\times$ year) fixed-effects panel
regressions of forward returns on the AI $z$-scores and controls, with standard
errors double-clustered by firm and year, alongside Fama--MacBeth estimates.

\input{tab_panel}

The two contrasting columns are instructive. In the \textbf{Fama--MacBeth}
cross-sectional regressions the AI-adoption and AI-infrastructure premia are positive
(+1.3 and +1.6~pp per standard deviation) with $t$-statistics near 1.6--1.7---weakly
positive pricing in the cross-section. In the \textbf{two-way fixed-effects} panel,
however, every AI coefficient collapses to essentially zero. Because firm fixed
effects remove all permanent cross-firm differences, this tells us the signal lives
\emph{across} firms, not \emph{within} them: a given firm writing relatively more
about AI from one year to the next does not predict its own subsequent return. The
market priced \emph{which firms are AI-exposed}, not the annual wording wiggle. (Size
carries the familiar negative premium throughout, $t\approx-8$.)

\subsection{The structural break: AI pricing switched on after ChatGPT}
The full-sample averages mask a sharp regime change. Splitting at the November~2022
launch of ChatGPT, Figure~\ref{fig:quint} (center vs.\ right panel) shows the
quintile pattern is nearly flat before 2022 and steeply increasing afterward. The
annual Q5$-$Q1 spreads in Figure~\ref{fig:spread} are small and mixed-sign through
2021, then large and positive for every post-2022 formation year.

\begin{figure}[H]\centering
\includegraphics[width=\textwidth]{fig1_quintile_returns.png}
\caption{Mean forward 12-month return by AI-disclosure composite quintile. The
high-minus-low spread is +2.4~pp before ChatGPT and +14.6~pp after.}\label{fig:quint}
\end{figure}

\begin{figure}[H]\centering
\includegraphics[width=0.92\textwidth]{fig3_annual_spread.png}
\caption{Annual high-minus-low (Q5$-$Q1) forward-return spread on the AI-disclosure
composite, by 10-K fiscal year. Post-ChatGPT formation years (red) are uniformly and
substantially positive.}\label{fig:spread}
\end{figure}

Table~\ref{tab:break} quantifies the break two ways: sub-period quintile spreads and
a pooled interaction of the AI $z$-score with a post-2022 indicator.

\input{tab_break}

The numbers are striking. The AI-adoption Q5$-$Q1 spread rises from +2.2\%/yr
\tstat{2.04} pre-2022 to \textbf{+14.3\%/yr} \tstat{2.15} post-2022; the composite
rises from +2.5\% to +14.6\%/yr. In the interaction regressions the post-2022
incremental slope is positive for all three signals and statistically significant for
AI~infrastructure \tstat{2.63}, the facet most directly tied to the firms (chipmakers,
cloud, data-center, and power names) that led the 2023--2025 rally. The cumulative
long-short wealth curve in Figure~\ref{fig:cum} makes the timing visible: the AI
portfolios drift sideways-to-up for a decade, then inflect upward after the dashed
ChatGPT line.

\begin{figure}[H]\centering
\includegraphics[width=0.92\textwidth]{fig2_cumulative_ls.png}
\caption{Cumulative value of \$1 invested in the equal-weighted, annually
rebalanced, self-financing high-minus-low AI-disclosure portfolio.}\label{fig:cum}
\end{figure}

Figure~\ref{fig:bin} confirms the pattern is not driven by outliers: the binned
relation between the AI composite and forward returns is flat pre-2022 and clearly
upward-sloping post-2022.

\begin{figure}[H]\centering
\includegraphics[width=\textwidth]{fig5_binscatter.png}
\caption{Binned scatter of forward return against the AI-disclosure composite, by
sub-period. The conditional slope steepens markedly after ChatGPT.}\label{fig:bin}
\end{figure}

\section{Discussion}
Three conclusions follow. \textbf{(1) AI disclosure is informative, but
conditionally.} It is not a stable, decade-long return factor; it is a signal that
the market began to price only once generative AI became salient. This is consistent
with an \emph{attention} or \emph{re-rating} mechanism: the textual evidence of AI
engagement existed for years, but capital flowed toward AI-exposed firms only after a
narrative catalyst. \textbf{(2) The effect is cross-sectional, not within-firm.} The
disappearance of the coefficient under firm fixed effects warns against trading
year-to-year changes in a single firm's AI verbosity; the durable signal is the
persistent \emph{level} of a firm's AI exposure relative to peers. \textbf{(3) The
three AI lenses largely coincide.} Their high collinearity means one should not
over-interpret which specific facet (adoption vs.\ infrastructure vs.\ risk)
``wins''; they jointly proxy a single AI-salience construct, and the infrastructure
lens simply loads most heavily on the firms that led the boom.

\section{Limitations}
The post-2022 window contains only three formation years, so the regime result, while
economically large and robust across signals and specifications, rests on a short
post-period and should be read as suggestive rather than definitive. The market-cap
proxy holds shares outstanding at their latest value; survivorship is present because
we use current index membership; and the semantic anchors, though carefully curated,
embed the authors' definitional choices. Finally, disclosure intensity is correlated
with sector (Information Technology firms naturally score higher), so part of the
post-2022 premium reflects a sector tilt toward technology---though the within-year,
size- and momentum-controlled tests mitigate this.

\section{Conclusion}
Using a semantic, embedding-based measure of AI language in S\&P~500 filings linked
to Yahoo Finance returns, we find that AI disclosure carried little pricing power for
most of the 2010s and then became strongly rewarded after the launch of ChatGPT. The
high-minus-low AI-disclosure return spread jumped from roughly +2~pp to roughly
+15~pp per year. The pattern is a regime-dependent, cross-sectional re-rating rather
than a persistent factor premium---evidence that markets price the \emph{language} of
AI, but only once AI commands their attention.

\vspace{1em}\hrule\vspace{0.4em}
{\footnotesize\textit{Data:} EDGAR Research Database (UQO); prices from Yahoo Finance;
factors from the Kenneth R.\ French data library. All return tests use look-ahead-free
July-formation forward returns. Code and tables: \texttt{analysis/}.}
\end{document}