spb/wp17_uqo

1% ============================================================================2\section{Data and Variable Construction}3\label{sec:data}4% ============================================================================56\subsection{The semantic factor panel}7\label{subsec:semantic}8The backbone of our study is the EDGAR Research Database, an infrastructure that downloads, parses,9and semantically scores SEC filings for S\&P~500 firms. For each filing the pipeline proceeds in four10steps. It (i)~extracts the canonical sections of the 10-K---Risk Factors, Management's Discussion and11Analysis (MD\&A), legal proceedings, the financial statements, and so on---so that the location as12well as the content of each passage is known; (ii)~splits each section into overlapping text chunks of13a few sentences, the overlap ensuring that a thought spanning a chunk boundary is not lost;14(iii)~embeds every chunk into a 384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence15transformer, a compact and widely validated model that maps semantically similar passages to nearby16points in the embedding space; and (iv)~scores each chunk by its cosine similarity to a set of17\emph{anchor} passages that define eight economic concepts. The anchors are short, deliberately18written paragraphs that articulate each concept in natural language. The AI-adoption anchor, for19instance, describes a firm ``internally using artificial intelligence, machine learning, generative20AI, large language models, automation, predictive analytics \dots\ to improve productivity,21operations, efficiency, decision-making.'' Cosine similarity to such an anchor measures how closely a22passage's \emph{meaning} matches the concept, on a scale from $-1$ to $1$, regardless of the exact23words used---so that a sentence on ``deploying neural networks to forecast demand'' scores highly on24AI adoption even though it shares no keyword with the anchor. This is precisely the paraphrase- and25context-sensitivity that keyword counts lack.2627Chunk-level scores are aggregated to the filing level using section weights that emphasize the parts28of the document where the concept is most likely to be discussed substantively (e.g.\ MD\&A and Risk29Factors), and then to the firm-fiscal-year level. For each concept the database reports the mean,30maximum, standard deviation, top-5\% tail, and section-weighted mean (\texttt{wmean}) of chunk31similarities, along with first and second differences, percentage changes, and three-year rolling32statistics. The eight concepts comprise three AI facets---\emph{AI adoption}, \emph{AI33infrastructure}, \emph{AI disruption risk}---and five others used as placebos: \emph{ESG commitment},34\emph{climate risk}, \emph{cybersecurity}, \emph{automation/labor}, and \emph{macro risk}.3536The resulting panel covers all 500 current S\&P~500 members across fiscal years 2010--2026, a total37of 7{,}869 firm-years. We take the section-weighted mean (\texttt{wmean}) as our headline signal38because it up-weights the forward-looking, strategy-rich sections; we verify robustness to the39simple mean and the top-5\% tail in Section~\ref{sec:robustness}. Our three primary signals are40therefore \texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}), \texttt{ai\_infrastructure\_wmean}41(\textbf{AI\,Infra}), and \texttt{ai\_disruption\_risk\_wmean} (\textbf{AI\,Risk}); we also form a42standardized opportunity composite, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$.4344\subsection{Price and factor data}45\label{subsec:price}46We download monthly adjusted-close prices for all 500 tickers from Yahoo Finance over 2009--202647(handling share-class tickers such as BRK.B and BF.B), the SPDR S\&P~500 ETF (SPY) as the market48benchmark, and the Fama--French five factors \citep{famafrench2015}, the momentum factor, and the49risk-free rate from the Kenneth R.\ French data library. Monthly simple returns are computed from50adjusted prices, which incorporate dividends and splits. A market-capitalization proxy is formed as51shares outstanding (from Yahoo Finance) times the historical price; we use it for value-weighting and52as a size control, acknowledging the shares-held-constant approximation.5354\subsection{Timing and the construction of tradeable returns}55\label{subsec:timing}56The single most important design choice in a disclosure--return study is timing, because a signal57that inadvertently peeks at information unavailable in real time can manufacture predictability that58no investor could have traded on. A 10-K covering fiscal year $Y$ becomes public only when filed,59typically in February or March of year $Y{+}1$. To guarantee that our signal is in the public60information set before any return is measured, we follow the Fama--French portfolio convention:61signals from fiscal year $Y$ are used to form portfolios at the end of June of $Y{+}1$, and we measure62the forward twelve-month return from July of $Y{+}1$ through June of $Y{+}2$. The roughly three-month63gap between the typical filing date and June~30 builds in a margin that rules out look-ahead bias even64for late filers, while the July-to-June holding window is the standard one in the cross-sectional65literature and so keeps our estimates comparable to it. After requiring a valid forward return with66adequate monthly coverage, we obtain 6{,}852 firm-years spanning formation years 2010--2024, with67return realizations extending through mid-2026.6869\subsection{Controls and standardization}70\label{subsec:controls}71For each formation year we compute cross-sectional $z$-scores of every signal, so that all72coefficients are expressed per cross-sectional standard deviation and the powerful secular rise in AI73language (documented below) is differenced out---in any given year we compare firms to their74contemporaneous peers. We control for the three workhorse return determinants: \emph{size} (log75market-cap proxy), \emph{momentum} (the firm's own return over the twelve months ending at formation,76skipping the last month), and \emph{idiosyncratic volatility} (the annualized standard deviation of77monthly returns over the trailing year). In regressions, returns and continuous controls are78winsorized at the 1st and 99th percentiles to limit the influence of outliers; in portfolio sorts we79use raw returns. Appendix~\ref{app:variables} gives precise definitions.8081\subsection{Summary statistics}82\label{subsec:summary}83Table~\ref{tab:summ} reports summary statistics for the estimation sample. The average forward84twelve-month return is 18.6\% (reflecting the strong 2011--2026 bull market for U.S.\ large caps),85with a wide cross-sectional dispersion (standard deviation 41.5\%). The AI signals are small in86level---cosine similarities of disclosure to narrow anchors---but exhibit meaningful cross-sectional87variation. Table~\ref{tab:corr} reports correlations. Two features stand out. First, the three AI88signals are highly collinear (pairwise $\rho$ between 0.75 and 0.86): firms that write more about89adopting AI also write more about AI infrastructure and AI risk. They are best understood as three90lenses on a single latent construct---\emph{AI salience}---a point we formalize with91principal-components analysis in Section~\ref{sec:robustness}. Second, the raw (unconditional)92correlation between each AI signal and the forward return is positive but small ($\approx 0.04$),93foreshadowing that the unconditional relation is weak and that the action lies in its time variation.9495\input{tables/tab_summary}96\input{tables/tab_corr}9798\subsection{The secular rise of AI language}99\label{subsec:trend}100Before turning to returns, Figure~\ref{fig:trend} documents the raw phenomenon that motivates our101within-year design. The cross-sectional mean similarity to all three AI concepts trended steadily102upward through the 2010s and inflected sharply after 2022; by 2025--2026 AI language saturates103S\&P~500 disclosure. This is precisely why our tests rely on within-year cross-sectional ranking: the104interesting question is not whether the index as a whole talks more about AI over time (it105emphatically does), but whether, \emph{in a given year}, firms that talk about AI more than their106peers subsequently earn different returns. Figure~\ref{fig:sectorbox} shows that AI disclosure is107also unevenly distributed across sectors, concentrated in Information Technology, Communication108Services, and Industrials---a fact that will matter for interpreting the sector channel.109110\begin{figure}[htbp]\centering111\includegraphics[width=0.9\textwidth]{fig4_ai_language_trend.png}112\caption{\textbf{The secular rise of AI language in S\&P~500 disclosure.} The figure plots the113cross-sectional mean of the section-weighted cosine similarity of 10-K text to each of the three AI114anchors (adoption, infrastructure, and disruption risk), by fiscal year over 2010--2026. All three115measures trend steadily upward through the 2010s and inflect sharply after the November 2022 launch of116ChatGPT (dashed vertical line), by which point AI language saturates index disclosure. This secular117trend motivates the within-year cross-sectional design used throughout the paper.}118\label{fig:trend}119\end{figure}120121\begin{figure}[htbp]\centering122\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}123\caption{\textbf{AI disclosure across GICS sectors.} The figure shows box plots of the firm-level124AI-adoption score within each GICS sector (boxes span the interquartile range, the red line marks the125sector median, and outliers are suppressed for legibility). AI disclosure is concentrated in126technology-intensive sectors---Information Technology, Communication Services, and Industrials---a127concentration that motivates the sector-channel analysis and the sector-neutral robustness checks128reported later.}129\label{fig:sectorbox}130\end{figure}131132\subsection{Which firms disclose the most AI? Descriptive evidence}133\label{subsec:topfirms}134Table~\ref{tab:topfirms} lists the ten most AI-intensive S\&P~500 firms by average AI-adoption score135over the full sample. The list is instructive in two ways. First, it has strong face validity: it136features enterprise-technology and software firms (Hewlett Packard Enterprise, Intel, IBM, Tyler137Technologies, Autodesk), an algorithm-centric platform (Uber), and data-intensive financials (BNY138Mellon, Berkshire Hathaway), exactly the kinds of firms one would expect to write substantively about139deploying AI. Second, the list is \emph{not} confined to the technology sector: industrials (Cummins)140and energy (EOG Resources) appear, confirming that intensive AI disclosure cuts across the GICS141taxonomy and that our continuous, text-based measure captures variation a sector dummy would miss.142143A further descriptive fact is central to our interpretation: the AI signal is highly144\emph{persistent}. The year-over-year cross-sectional rank autocorrelation of the AI composite is1450.88, and even at a five-year horizon it remains 0.73. Firms that are AI-intensive disclosers stay146AI-intensive; the measure captures a slow-moving structural characteristic of the firm rather than147transient wording. This persistence has two implications we exploit later: it explains why firm fixed148effects absorb most of the signal (there is little within-firm variation to identify), and it149underpins the difference-in-differences design, in which we fix each firm's exposure at its pre-event150level precisely because that level is stable.151152\input{tables/tab_topfirms}153