spb/wp17_uqo

1% ============================================================================2\section{Data and Variable Construction}3\label{sec:data}4% ============================================================================56\subsection{The scored filing corpus}7\label{subsec:corpus_data}8The backbone of the study is the EDGAR Research Database, an infrastructure built by the authors9that downloads, parses, and semantically scores SEC filings for the 500 current S\&P~50010constituents. The acquisition layer retrieved every Form~10-K, 10-Q, and 8-K filed by these firms11from January 2010 to June 2026 (\num{134124} filings). The parsing layer splits each document into12its canonical sections where a section template exists (the quarterly report is parsed into its13financial statements, management's discussion and analysis (MD\&A), quantitative disclosures,14controls and procedures, legal proceedings, risk factors, and unregistered sales; the current15report is kept as a single body) and then into overlapping chunks of about 500 words, the overlap16ensuring that a thought spanning a chunk boundary is not lost. The embedding layer maps every17chunk to a 384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence transformer18\citep{reimers2019sbert}, a compact and widely validated model that places semantically similar19passages close together.2021Table~\ref{tab:corpus} describes the corpus that had actually been parsed and scored when the22panel used in this paper was built. All \num{103926} Form~8-K filings of the 500 firms were23processed, as were \num{3710} of the \num{22542} Form~10-Q filings, belonging to 84 firms, almost all24of them with tickers between SMCI and ZTS (the parser was interrupted during an alphabetical pass;25Appendix~\ref{app:corpus} lists them); none of26the \num{7655} Form~10-K filings had been processed. The first version of this paper described the27corpus as the full text of annual reports, which was incorrect. The scored text is therefore28dominated by current reports: a typical 8-K contributes two chunks and a typical firm-year contains29about thirteen filings and 28 chunks (interquartile range 17--57). Current reports are short and30event-driven---earnings releases, material agreements, executive changes, results of shareholder31votes---so the AI measure captures the extent to which a firm's \emph{announcements} of the year32are couched in AI-related language, rather than the extent of its annual strategic narrative. The3310-Q filings that were processed carry roughly thirty times more chunks each and, because they34contain MD\&A and risk-factor sections, score about 0.07 higher on the AI-adoption anchor on35average than an 8-K. Section~\ref{subsec:corpus} shows what this implies for the firm-year scores and36re-estimates the main results on 8-K filings alone.3738\input{tables/tab_corpus}3940\subsection{Anchor concepts and semantic scores}41\label{subsec:semantic}42Each of eight economic concepts is defined by three short paragraphs written in the register of a43filing; the full texts are reproduced in Appendix~\ref{app:anchors}. The AI-adoption anchor, for44instance, describes a firm that ``internally uses artificial intelligence, machine learning,45generative AI, large language models, automation, predictive analytics \dots\ to improve46productivity, operations, efficiency, decision-making.'' The three paragraphs of a concept are47embedded, averaged, and normalized, so that each concept $a$ is represented by a unit vector48$\bar{\mathbf a}_a$. The score of chunk $c$ on concept $a$ is the cosine similarity49\begin{equation}50  s_{c,a} \;=\; \frac{\mathbf e_c^{\prime}\,\bar{\mathbf a}_a}{\lVert \mathbf e_c\rVert\,\lVert \bar{\mathbf a}_a\rVert}\in[-1,1],51  \label{eq:cos}52\end{equation}53where $\mathbf e_c$ is the embedding of the chunk. Because similarity is computed in the embedding54space, a passage on ``deploying neural networks to forecast demand'' scores highly on AI adoption55even though it shares no keyword with the anchor; this paraphrase sensitivity is what keyword56counts lack. The eight concepts comprise three AI facets---\emph{AI adoption}, \emph{AI57infrastructure}, and \emph{AI disruption risk}---and five others used as placebos: \emph{ESG58commitment}, \emph{climate risk}, \emph{cybersecurity}, \emph{automation/labour}, and \emph{macro59risk}.6061Chunk scores are aggregated to the filing level in three ways. The simple mean averages all chunks62of filing $f$; the top-5\% tail averages the chunks at or above the 95th percentile of the filing's63scores; and the section-weighted mean, our headline measure, weights chunks by the section in which64they appear,65\begin{equation}66  \bar s^{\,w}_{f,a} \;=\; \frac{\sum_{c\in f} w_{\sigma(c),a}\, s_{c,a}}{\sum_{c\in f} w_{\sigma(c),a}},67  \label{eq:wmean}68\end{equation}69where $\sigma(c)$ is the section of chunk $c$ and the weights $w$ emphasize MD\&A and risk-factor70text for the relevant concept (for AI adoption and infrastructure, MD\&A receives weight 2.0 and71risk factors 1.2; for AI disruption risk, risk factors 2.0 and MD\&A 1.5; financial statements72receive 0.5; the undivided body of an 8-K receives 1.0). Since 8-K filings have a single section,73the weighted and simple means coincide for them, and the weights only affect the 10-Q filings of74the 84 firms. Filing-level scores are then averaged, without weights, over all scored filings dated75in calendar year $Y$:76\begin{equation}77  \text{AI}_{i,Y,a} \;=\; \frac{1}{|\mathcal F_{i,Y}|}\sum_{f\in\mathcal F_{i,Y}} \bar s^{\,w}_{f,a},78  \label{eq:annual}79\end{equation}80where $\mathcal F_{i,Y}$ is the set of scored filings of firm $i$ in year $Y$. An important81consequence of equation~\eqref{eq:annual} is that a 10-Q with sixty chunks and an 8-K with two82chunks carry the same weight, so that the presence of a scored 10-Q in a firm-year raises the83annual score mechanically. The panel covers the 500 current constituents over filing years842010--2026, a total of \num{7869} firm-years. Our three primary signals are85\texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}), \texttt{ai\_infrastructure\_wmean}86(\textbf{AI\,Infra}), and \texttt{ai\_disruption\_risk\_wmean} (\textbf{AI\,Risk}); we also form a87standardized opportunity composite, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$,88where $z(\cdot)$ denotes the within-year cross-sectional standardization defined below.8990\subsection{Price and factor data}91\label{subsec:price}92We download monthly adjusted-close prices for all 500 tickers from Yahoo Finance over June 2009 to93June 2026 (handling share-class tickers such as BRK.B and BF.B), the SPDR S\&P~500 ETF (SPY) as the94market benchmark, and the Fama--French five factors \citep{famafrench2015}, the momentum factor95\citep{carhart1997}, and the risk-free rate from the Kenneth R.\ French data library (available96through April 2026). Monthly simple returns are computed from adjusted prices, which incorporate97dividends and splits. A market-capitalization proxy is formed as the latest shares outstanding98(from Yahoo Finance, June 2026) times the historical price; we use it for value-weighting and as a99size control, acknowledging that it holds shares constant. Daily prices from September 2021 to100July 2024 are used for the event study.101102\subsection{Universe and survivorship}103\label{subsec:universe}104The universe is the S\&P~500 membership as of June 2026, not the historical membership. Firms that105left the index before 2026 are absent, and firms that joined it after 2010---83 of the 500 have106Yahoo Finance prices beginning after July 2009, among them Uber, Palantir, and Airbnb---enter the107panel with their pre-inclusion history. This is a survivorship and look-ahead bias at the level of108the universe \citep{browngoetzmannross1995}: it raises the average level of returns (entrants earn10910.6 percentage points more per year than seasoned firms in our panel) and could bias the110cross-sectional AI--return relation if recent entrants were concentrated in high-AI quintiles. We111show in Section~\ref{subsec:entrants} that entrants are spread evenly across the AI quintiles and112that the main results hold among firms listed since 2009, but the reader should bear in mind that113the panel is not a point-in-time index.114115\subsection{Timing and the construction of tradeable returns}116\label{subsec:timing}117The most important design choice in a disclosure--return study is timing, because a signal that118inadvertently uses information unavailable in real time can manufacture predictability that no119investor could have traded on. Every filing contributing to the signal of year $Y$ is dated between1201~January and 31~December of $Y$, and it is public on its filing date. We follow the121portfolio-formation convention of \citet{famafrench2015}: signals of year $Y$ are used to form122portfolios at the end of June of $Y{+}1$, and the forward return is the buy-and-hold return from123July of $Y{+}1$ through June of $Y{+}2$,124\begin{equation}125  r_{i,Y+1} \;=\; \prod_{m=\text{Jul}\,Y+1}^{\text{Jun}\,Y+2}\bigl(1+r_{i,m}\bigr)-1 ,126  \label{eq:fwd}127\end{equation}128computed when at least six monthly returns are available in the window. The gap between the last129filing of year $Y$ and portfolio formation is therefore at least six months and on average about130twelve, which rules out look-ahead bias but also means that the signal is stale relative to the131quarterly cadence of current reports; this works against finding predictability. Two consequences132of the calendar-year convention deserve notice. The signal of the 2022 cohort includes 8-K filings133dated in December 2022, after the ChatGPT launch, so a small part of that signal is itself a134reaction to the event; the difference-in-differences design of Section~\ref{sec:robustness}, which135fixes exposure at its 2019--2021 level, is immune to this. And the holding period of the 2021136cohort (July 2022--June 2023) straddles the launch; we treat this cohort as pre-launch in the137baseline and show in Section~\ref{sec:robustness} that moving it does not change the conclusions.138After requiring a valid forward return we obtain \num{6852} firm-years (497 firms) spanning139formation years 2010--2024, with return realizations extending to June 2026; the forward return of140the 2024 cohort ends with the partial June 2026 observation, prices having been downloaded on14117~June 2026. Filing years 2025 and 2026 contribute signals but no forward return.142143\subsection{Controls and standardization}144\label{subsec:controls}145For each formation year we compute cross-sectional $z$-scores of every signal,146\begin{equation}147  z(x_{i,Y}) \;=\; \frac{x_{i,Y}-\bar x_{Y}}{\mathrm{sd}_Y(x)} ,148  \label{eq:z}149\end{equation}150so that all coefficients are expressed per cross-sectional standard deviation and the secular rise151in AI language documented below is removed: in any given year we compare firms to their152contemporaneous peers. We control for three return determinants measured at formation:153\emph{size} (log market-cap proxy), \emph{momentum} (the firm's own return over the twelve months154ending at formation, skipping the last month), and \emph{volatility} (the annualized standard155deviation of monthly returns over the trailing year). In regressions, returns and continuous156controls are winsorized at the 1st and 99th percentiles; in portfolio sorts we use raw returns and157report winsorized spreads as a robustness check. Appendix~\ref{app:variables} gives precise158definitions.159160\subsection{Summary statistics}161\label{subsec:summary}162Table~\ref{tab:summ} reports summary statistics for the estimation sample. The average forward163twelve-month return is 18.6\%, reflecting the strong 2011--2026 market for U.S.\ large caps and the164survivorship of the universe, with a wide cross-sectional dispersion (standard deviation 41.5\% and165a maximum of 969\%). The AI signals are small in level---cosine similarities of short passages to166narrow anchors---but exhibit meaningful cross-sectional variation. Table~\ref{tab:corr} reports167correlations. The three AI signals are highly collinear (pairwise $\rho$ between 0.74 and 0.86):168firms whose filings resemble the adoption anchor also resemble the infrastructure and risk anchors.169They are best understood as three lenses on a single latent construct, AI salience, a point we170formalize with principal components in Section~\ref{sec:robustness}. The raw correlation between171each AI signal and the forward return is positive but small ($\approx 0.04$).172173\input{tables/tab_summary}174\input{tables/tab_corr}175176\subsection{The secular rise of AI language}177\label{subsec:trend}178Figure~\ref{fig:trend} documents the phenomenon that motivates the within-year design. The179cross-sectional mean similarity to all three AI concepts trended upward through the 2010s and180inflected sharply after 2022; by 2025--2026 AI language is pervasive in the filings of the index.181The interesting question is therefore not whether the index as a whole talks more about AI over182time, but whether, in a given year, firms that talk about AI more than their peers subsequently earn183different returns. Figure~\ref{fig:sectorbox} shows that AI language is unevenly distributed across184sectors, concentrated in Information Technology, Communication Services, and Industrials, a fact185that matters for the sector analysis.186187\begin{figure}[htbp]\centering188\includegraphics[width=0.9\textwidth]{fig4_ai_language_trend.png}189\caption{\textbf{The secular rise of AI language in S\&P~500 filings.} The figure plots the190cross-sectional mean of the section-weighted cosine similarity of the scored filings (Form~8-K and,191for 84 firms, Form~10-Q) to each of the three AI anchors, by filing year over 2010--2026. All three192measures trend upward through the 2010s and inflect after the November 2022 launch of ChatGPT193(dashed vertical line). The trend motivates the within-year cross-sectional design.}194\label{fig:trend}195\end{figure}196197\begin{figure}[htbp]\centering198\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}199\caption{\textbf{AI language across GICS sectors.} Box plots of the firm-year AI-adoption score200within each GICS sector (boxes span the interquartile range, the red line marks the sector median,201outliers are suppressed). AI language is concentrated in Information Technology, Communication202Services, and Industrials, which motivates the sector-neutral and sector-factor analyses of203Section~\ref{sec:robustness}.}204\label{fig:sectorbox}205\end{figure}206207\subsection{Which firms score highest?}208\label{subsec:topfirms}209Table~\ref{tab:topfirms} lists the ten firms with the highest average AI-adoption score. The list210contains enterprise-technology and software firms (Hewlett Packard Enterprise, Intel, IBM, Tyler211Technologies, Autodesk), a platform company (Uber), data-intensive financials (BNY Mellon,212Berkshire Hathaway), an industrial (Cummins), and an energy producer (EOG Resources). Four of the213ten belong to the group of 84 firms whose 10-Q filings were scored, which mechanically raises their214scores; the ranking should therefore be read as descriptive. A further fact is central to the215interpretation: the signal is persistent. The average year-over-year cross-sectional rank216correlation of the composite is 0.88, and at a five-year lag it is still 0.73. Firms whose filings217score high on AI stay high; the measure captures a slow-moving characteristic rather than transient218wording. This explains why firm fixed effects absorb most of the signal, and it underpins the219difference-in-differences design that fixes each firm's exposure at its pre-launch level.220221\input{tables/tab_topfirms}222