TeX 69.3%
Python 30.6%
1% ============================================================================2\section{Related Literature and Hypotheses}3\label{sec:literature}4% ============================================================================56Our study sits at the intersection of five strands of research: textual analysis in finance, the7pricing of corporate disclosure, technology and innovation in asset pricing, investor attention and8narrative economics, and the nascent literature on the economics of artificial intelligence. Each9contributes a building block---a measurement tool, an object of study, a candidate risk factor, a10behavioural mechanism, and a comparison case---and a sixth, methodological strand disciplines the11inference. We take the strands in turn and close with the hypotheses they jointly motivate.1213\subsection{Textual analysis and natural language processing in finance}14A now-large literature extracts economic signals from financial text. \citet{tetlock2007} showed15that the tone of a \textit{Wall Street Journal} column forecasts aggregate returns, and16\citet{tetlock2008} extended this to firm-level news. \citet{lm2011} demonstrated that17finance-specific sentiment dictionaries materially outperform general-purpose word lists in 10-K18analysis, launching a wave of dictionary-based studies; \citet{lm2016} and \citet{lm2020} survey19the field and its pitfalls, including the sensitivity of results to document selection and20pre-processing. A second generation moved from counting words to modelling them: \citet{gentzkow2019}21review text-as-data methods in economics, \citet{kekellyxiu2019} develop a supervised topic model22that predicts returns from news, \citet{kelly2019textregression} propose a sparse text-regression23framework, and \citet{bybee2024narrative} recover a ``state of the economy'' from business news.24A third generation uses distributed representations of meaning. \citet{hanleyhoberg2019} apply word25embeddings to bank 10-Ks to identify emerging risks before they are named in dictionaries;26\citet{limai2021} measure corporate culture from earnings calls with word embeddings;27\citet{huang2023finbert} show that a domain-adapted transformer outperforms dictionaries and28bag-of-words classifiers on financial text; and \citet{cong2022textfactor} build interpretable29``textual factors'' from embeddings. Our paper belongs to this embedding-based generation. The30limitations of keyword counting are acute for a theme like AI: a count cannot distinguish ``we do31not currently use artificial intelligence'' from ``we are an artificial-intelligence leader,'' it32misses the paraphrase of a firm that describes machine learning without writing ``AI,'' and it33treats boilerplate and substance identically. We instead measure the semantic similarity of each34passage to anchor concepts with a sentence transformer \citep{reimers2019sbert}, a design close in35spirit to the similarity-to-reference-text constructions of \citet{hanleyhoberg2019}. The36literature also warns, however, that text-based measures inherit the composition of the corpus they37are computed on \citep{lm2016,lm2020}; Section~\ref{sec:data} therefore documents that composition38in detail.3940\subsection{The pricing of corporate disclosure}41A parallel literature asks whether the \emph{content} and \emph{changes} of mandatory disclosure42predict returns. \citet{cohen2020lazy} show that firms that change the language of their annual43reports relative to the prior year experience predictable future returns and operating outcomes.44\citet{hoberg2010product} and \citet{hobergphillips2016} use 10-K product descriptions to build45text-based industry classifications that carry asset-pricing content. \citet{lopezlira2023risk}46extracts priced risk exposures from risk-factor sections. The general lesson is that the47qualitative portions of filings contain value-relevant information that markets process48incompletely and with delay. We contribute a concept-level measure---AI engagement---computed49mainly from current reports rather than annual reports, and we characterize how its relation to50returns varies over time, a dimension the disclosure-pricing literature has largely left implicit.5152\subsection{Technology, innovation, and the cross-section of returns}53Asset pricing has long sought to connect technological change to risk and return.54\citet{hirshleifer2013innovative} show that innovative efficiency predicts returns;55\citet{kogan2017technological} build market-based measures of innovation from patents and link them56to growth and reallocation; \citet{koganpapanikolaou2014} show that investment-specific technology57shocks shape the cross-section. A more recent literature studies AI exposure directly.58\citet{babina2024aiinvestment} measure firm-level AI investment from job postings and r\'esum\'es59and document faster growth among AI-adopting firms; \citet{acemoglu2022aijobs} and \citet{webb2020}60measure the exposure of occupations and establishments to AI; \citet{acemoglu2020wrong} study the61labour-market effects of industrial robots. Closest to us, \citet{eisfeldt2023generative} construct62a firm-level measure of \emph{generative}-AI exposure from the task content of occupations and show63that a long--short portfolio on that measure earned large returns in the weeks after ChatGPT's64release; \citet{jha2024chatgpt} use ChatGPT itself to extract expected investment and employment65policies from earnings calls; and \citet{lopezliratang2026} show that a large language model can66rank news headlines in a way that predicts next-day returns. Our contribution relative to this work67is to measure AI exposure from the firm's own disclosure language, available for the entire index68at annual frequency, and to characterize the full time profile of its pricing across fifteen69cohorts rather than an event window. Where \citet{eisfeldt2023generative} document a sharp70repricing in the weeks around the launch, we ask whether a persistent, text-based exposure measure71carried a premium before the launch and whether the premium after it is distinguishable from a72technology-sector effect.7374\subsection{Investor attention, salience, and narrative economics}75Why might value-relevant disclosure be priced only after a catalyst? The behavioural literature76offers a mechanism: limited investor attention. \citet{barberodean2008} show that individual77investors are drawn to attention-grabbing stocks; \citet{da2011} construct an attention proxy from78Google search and link it to prices; \citet{dellavigna2009} document under-reaction to information79released when attention is low; and \citet{hirshleifer2009driven} show that earnings news is80underweighted when investors are distracted. \citet{shiller2017narrative} argues that contagious81narratives drive major asset-price movements, and \citet{bakerwurgler2006} show that sentiment82differentially affects hard-to-value, speculative stocks. These mechanisms predict that a salient83consumer event such as ChatGPT could draw attention to information that had been sitting in filings84for years. They do not, however, distinguish a firm-level attention shock from a sector-level85revision of expected cash flows, and our tests are designed to separate the two.8687\subsection{ESG and climate disclosure as a comparison case}88Our placebo design connects to the literature on the pricing of ESG and climate disclosure.89\citet{pastor2021sustainable} model how sustainability preferences depress the expected returns of90green assets in equilibrium, and \citet{pastor2022dissecting} show that green stocks' strong91realized returns through 2021 reflected unexpected increases in climate concern rather than high92expected returns. \citet{bolton2021risk} document a carbon premium, and \citet{sautner2023} build a93firm-level climate-exposure measure from earnings-call language that is priced in options and94returns. Consistent with the post-2021 reversal in this literature, we find that high-ESG and95high-climate disclosure earned negative spreads after 2022, which helps to rule out a generic96``more disclosure earns more'' explanation of the AI result.9798\subsection{Short samples, data snooping, and event-study inference}99Two methodological literatures bear directly on how much weight our post-2022 estimates can carry.100The first concerns the credibility of cross-sectional return patterns. \citet{harveyliuzhu2016}101argue that the multiplicity of tested predictors requires $t$-statistics well above conventional102thresholds; \citet{houxuezhang2020} find that most published anomalies fail to replicate under103standard procedures; and \citet{mcleanpontiff2016} document post-publication decay. A spread104estimated on three annual observations, one of which is extreme, must be read in that light. The105second concerns inference in event studies when all firms share a single event date, so that106abnormal returns are cross-sectionally correlated \citep{brownwarner1985,kolari2010} and107event-induced variance may differ from estimation-window variance \citep{boehmer1991}, and in108long-horizon settings, where the choice of benchmark and the compounding of abnormal returns109matter \citep{barberlyon1997,fama1998,mitchellstafford2000,kothariwarner2007}. Our event-study110inference follows the portfolio approach recommended in that literature. Finally, the use of a111current-membership universe raises the survivorship concerns analysed by112\citet{browngoetzmannross1995}, which we address with restrictions on listing history.113114\subsection{Hypotheses}115\label{subsec:hypotheses}116Our analysis is organized around four hypotheses.117\begin{description}[leftmargin=2.4em,style=nextline]118\item[H1 (Unconditional pricing).] If AI exposure carries a stable risk premium or persistent119mispricing, then sorting on AI-disclosure intensity yields a positive and significant risk-adjusted120return spread over the full 2010--2024 formation sample.121\item[H2 (Regime dependence).] If instead AI exposure is repriced only once AI becomes salient, the122return spread is small before and large after the late-2022 generative-AI catalyst, and the123increase is visible in each post-launch cohort rather than in a single one.124\item[H3 (AI-specificity).] If the post-2022 spread reflects an AI repricing rather than a125mechanical or market-wide artefact, it appears for AI-related semantic factors but not for unrelated126factors such as ESG, climate, or macro risk, and it is not spanned by a technology-sector factor.127\item[H4 (Cross-sectional, not within-firm).] If the market repriced \emph{which} firms are128AI-exposed rather than rewarding incremental wording, the effect is concentrated in the129cross-section, attenuates under firm fixed effects, and displays no pre-trend when exposure is held130fixed at its pre-launch level.131\end{description}132\noindent The evidence rejects H1, supports the first clause of H2 but not the second, supports133H3 with respect to the placebo factors but not with respect to the sector factor, and supports H4134only in part.135