spb/uqo-research-db

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 (i)~extracts10the canonical sections (Risk Factors, MD\&A, legal proceedings, financial statements, and so on);11(ii)~splits each section into overlapping text chunks; (iii)~embeds every chunk into a12384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence transformer; and (iv)~scores each13chunk by cosine similarity to a set of \emph{anchor} passages defining eight economic concepts. The14anchors are short, deliberately written paragraphs that articulate the concept in natural language;15for example, the AI-adoption anchor describes a firm ``internally using artificial intelligence,16machine learning, generative AI, large language models, automation, predictive analytics \dots\ to17improve productivity, operations, efficiency, decision-making.'' Cosine similarity to such an anchor18measures how closely a passage's \emph{meaning} matches the concept, regardless of the exact words19used---capturing paraphrase and context in a way keyword counts cannot.2021Chunk-level scores are aggregated to the filing level using section weights that emphasize the parts22of the document where the concept is most likely to be discussed substantively (e.g.\ MD\&A and Risk23Factors), and then to the firm-fiscal-year level. For each concept the database reports the mean,24maximum, standard deviation, top-5\% tail, and section-weighted mean (\texttt{wmean}) of chunk25similarities, along with first and second differences, percentage changes, and three-year rolling26statistics. The eight concepts comprise three AI facets---\emph{AI adoption}, \emph{AI27infrastructure}, \emph{AI disruption risk}---and five others used as placebos: \emph{ESG commitment},28\emph{climate risk}, \emph{cybersecurity}, \emph{automation/labor}, and \emph{macro risk}.2930The resulting panel covers all 500 current S\&P~500 members across fiscal years 2010--2026, a total31of 7{,}869 firm-years. We take the section-weighted mean (\texttt{wmean}) as our headline signal32because it up-weights the forward-looking, strategy-rich sections; we verify robustness to the33simple mean and the top-5\% tail in Section~\ref{sec:robustness}. Our three primary signals are34therefore \texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}), \texttt{ai\_infrastructure\_wmean}35(\textbf{AI\,Infra}), and \texttt{ai\_disruption\_risk\_wmean} (\textbf{AI\,Risk}); we also form a36standardized opportunity composite, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$.3738\subsection{Price and factor data}39\label{subsec:price}40We download monthly adjusted-close prices for all 500 tickers from Yahoo Finance over 2009--202641(handling share-class tickers such as BRK.B and BF.B), the SPDR S\&P~500 ETF (SPY) as the market42benchmark, and the Fama--French five factors \citep{famafrench2015}, the momentum factor, and the43risk-free rate from the Kenneth R.\ French data library. Monthly simple returns are computed from44adjusted prices, which incorporate dividends and splits. A market-capitalization proxy is formed as45shares outstanding (from Yahoo Finance) times the historical price; we use it for value-weighting and46as a size control, acknowledging the shares-held-constant approximation.4748\subsection{Timing and the construction of tradeable returns}49\label{subsec:timing}50The single most important design choice in a disclosure--return study is timing. A 10-K covering51fiscal year $Y$ becomes public only when filed, typically in February or March of year $Y{+}1$. To52guarantee that our signal is in the public information set before any return is measured, we follow53the Fama--French portfolio convention: signals from fiscal year $Y$ are used to form portfolios at54the end of June of $Y{+}1$, and we measure the forward twelve-month return from July of $Y{+}1$55through June of $Y{+}2$. The roughly three-month gap between the typical filing date and June~3056ensures no look-ahead bias even for late filers. After requiring a valid forward return with adequate57monthly coverage, we obtain 6{,}852 firm-years spanning formation years 2010--2024, with return58realizations extending through mid-2026.5960\subsection{Controls and standardization}61\label{subsec:controls}62For each formation year we compute cross-sectional $z$-scores of every signal, so that all63coefficients are expressed per cross-sectional standard deviation and the powerful secular rise in AI64language (documented below) is differenced out---in any given year we compare firms to their65contemporaneous peers. We control for the three workhorse return determinants: \emph{size} (log66market-cap proxy), \emph{momentum} (the firm's own return over the twelve months ending at formation,67skipping the last month), and \emph{idiosyncratic volatility} (the annualized standard deviation of68monthly returns over the trailing year). In regressions, returns and continuous controls are69winsorized at the 1st and 99th percentiles to limit the influence of outliers; in portfolio sorts we70use raw returns. Appendix~\ref{app:variables} gives precise definitions.7172\subsection{Summary statistics}73\label{subsec:summary}74Table~\ref{tab:summ} reports summary statistics for the estimation sample. The average forward75twelve-month return is 18.6\% (reflecting the strong 2011--2026 bull market for U.S.\ large caps),76with a wide cross-sectional dispersion (standard deviation 41.5\%). The AI signals are small in77level---cosine similarities of disclosure to narrow anchors---but exhibit meaningful cross-sectional78variation. Table~\ref{tab:corr} reports correlations. Two features stand out. First, the three AI79signals are highly collinear (pairwise $\rho$ between 0.75 and 0.86): firms that write more about80adopting AI also write more about AI infrastructure and AI risk. They are best understood as three81lenses on a single latent construct---\emph{AI salience}---a point we formalize with82principal-components analysis in Section~\ref{sec:robustness}. Second, the raw (unconditional)83correlation between each AI signal and the forward return is positive but small ($\approx 0.04$),84foreshadowing that the unconditional relation is weak and that the action lies in its time variation.8586\input{tables/tab_summary}87\input{tables/tab_corr}8889\subsection{The secular rise of AI language}90\label{subsec:trend}91Before turning to returns, Figure~\ref{fig:trend} documents the raw phenomenon that motivates our92within-year design. The cross-sectional mean similarity to all three AI concepts trended steadily93upward through the 2010s and inflected sharply after 2022; by 2025--2026 AI language saturates94S\&P~500 disclosure. This is precisely why our tests rely on within-year cross-sectional ranking: the95interesting question is not whether the index as a whole talks more about AI over time (it96emphatically does), but whether, \emph{in a given year}, firms that talk about AI more than their97peers subsequently earn different returns. Figure~\ref{fig:sectorbox} shows that AI disclosure is98also unevenly distributed across sectors, concentrated in Information Technology, Communication99Services, and Industrials---a fact that will matter for interpreting the sector channel.100101\begin{figure}[htbp]\centering102\includegraphics[width=0.9\textwidth]{fig4_ai_language_trend.png}103\caption{Cross-sectional mean of section-weighted cosine similarity to each AI anchor, by fiscal104year. All three measures trend upward and inflect after the 2022 launch of ChatGPT (dashed line).}105\label{fig:trend}106\end{figure}107108\begin{figure}[htbp]\centering109\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}110\caption{Distribution of the AI-adoption score by GICS sector (box plots, outliers suppressed; red111line $=$ median). AI disclosure is concentrated in technology-intensive sectors.}112\label{fig:sectorbox}113\end{figure}114115\subsection{Which firms disclose the most AI? Descriptive evidence}116\label{subsec:topfirms}117Table~\ref{tab:topfirms} lists the ten most AI-intensive S\&P~500 firms by average AI-adoption score118over the full sample. The list is instructive in two ways. First, it has strong face validity: it119features enterprise-technology and software firms (Hewlett Packard Enterprise, Intel, IBM, Tyler120Technologies, Autodesk), an algorithm-centric platform (Uber), and data-intensive financials (BNY121Mellon, Berkshire Hathaway), exactly the kinds of firms one would expect to write substantively about122deploying AI. Second, the list is \emph{not} confined to the technology sector: industrials (Cummins)123and energy (EOG Resources) appear, confirming that intensive AI disclosure cuts across the GICS124taxonomy and that our continuous, text-based measure captures variation a sector dummy would miss.125126A further descriptive fact is central to our interpretation: the AI signal is highly127\emph{persistent}. The year-over-year cross-sectional rank autocorrelation of the AI composite is1280.88, and even at a five-year horizon it remains 0.73. Firms that are AI-intensive disclosers stay129AI-intensive; the measure captures a slow-moving structural characteristic of the firm rather than130transient wording. This persistence has two implications we exploit later: it explains why firm fixed131effects absorb most of the signal (there is little within-firm variation to identify), and it132underpins the difference-in-differences design, in which we fix each firm's exposure at its pre-event133level precisely because that level is stable.134135\input{tables/tab_topfirms}136