spb/wp17_uqo

1% ============================================================================2\section{Robustness and Identification}3\label{sec:robustness}4% ============================================================================56The post-2022 change is large on average and, as Section~\ref{subsec:break} showed, concentrated.7This section confronts it with the alternative explanations a sceptical reader would raise: that it8is mechanical, that it is a mega-cap rally in disguise, that it is sensitive to how the signal is9built or to which filings were scored, that it reflects the composition of a current-membership10universe, that it merely tracks the technology sector, and that the inference behind the event11study and the difference-in-differences estimates is overstated. Some of these explanations are12rejected and some are not.1314\subsection{Placebo: is the change specific to AI? (H3)}15If the post-2022 spread were a mechanical artefact of the pipeline, or a generic ``high-disclosure16firms did well'' phenomenon, it should appear for any semantic factor. Table~\ref{tab:placebo} and17Figure~\ref{fig:placebo} re-run the sub-period and interaction design on all eight factors. The18three AI factors show post-2022 spreads of $+13.6$ to $+14.3$ points. Macro-risk language shows no19spread ($-0.5$). ESG-commitment and climate-risk language earn negative post-2022 spreads ($-8.8$20and $-5.0$ points), in line with the post-2021 reversal in the sustainable-investing literature21\citep{pastor2022dissecting}. Two non-AI factors, however, also rise: cybersecurity ($+13.5$),22which is bundled with technology exposure and whose anchor itself mentions AI-powered threat23detection, and automation/labour ($+7.2$), whose anchor refers to AI and robotics. The placebo24therefore rules out a generic disclosure effect and a pipeline artefact, and it shows a rotation25away from sustainability language, but it does not isolate AI from the broader technology theme26with which AI, cybersecurity, and automation language travel. In the interaction regressions none27of the eight coefficients other than AI infrastructure is significant, and the three post-period28annual spreads do not support finer distinctions.2930\input{tables/tab_placebo}3132\begin{figure}[htbp]\centering33\includegraphics[width=0.92\textwidth]{fig7_placebo.png}34\caption{\textbf{Placebo test across semantic factors.} Post-ChatGPT high-minus-low (Q5$-$Q1)35return spread for each of the eight semantic factors. The AI-related factors (in red) and36cybersecurity cluster at the top; automation/labour language also earns a positive spread;37ESG-commitment and climate-risk language earn negative spreads, and macro-risk language none. The38contrast rules out a generic disclosure effect but not a broader technology theme.}\label{fig:placebo}39\end{figure}4041\subsection{Event study around ChatGPT}42The portfolio results use annual formation and forward returns; the event study provides a daily,43cleanly dated complement. Table~\ref{tab:event} and Figure~\ref{fig:event} report market-model44cumulative abnormal returns by pre-event AI tercile. The ordering is monotone at every horizon and45widens over time. The High-minus-Low CAR difference is $+2.9$ points at 60 trading days, $+10.1$46points at one year, and $+14.6$ points at eighteen months; the Welch $t$-statistics of the first47version, 1.77, 1.97, and 1.89, treat the 160 High and 162 Low firms as independent. Because48abnormal returns are purged of market beta, the spread is not a difference in beta exposure. A49cross-sectional regression of one-year CAR on pre-event exposure yields $+3.1$ points per standard50deviation \tstat{1.39}; adding sector fixed effects reduces it to $+2.1$ points \tstat{0.99},51again locating much of the effect in a sector channel. Figure~\ref{fig:eventbin} shows that the52relation is smooth across the exposure distribution.5354\input{tables/tab_event}5556Table~\ref{tab:eventcda} re-examines the inference. The average pairwise correlation of57estimation-window residuals across the 322 High and Low firms is 0.039; for a one-sample test on58$n$ firms, \citet{kolari2010} show that the independence-based $t$-statistic overstates the true59one by the factor $\sqrt{(1+(n-1)\bar\rho)/(1-\bar\rho)}$, which here is almost four. The portfolio60statistic of equation~\eqref{eq:cda} avoids the problem by treating the daily High-minus-Low61abnormal return as one series. With the estimation-window standard deviation it gives $t=1.35$ at6260 days, $1.45$ at six months, $2.28$ at one year, and $2.63$ at eighteen months; with the63event-window standard deviation, $1.54$, $1.65$, $2.60$, and $2.99$. The one-year and64eighteen-month differences are therefore significant under tests that allow for dependence,65whereas the sixty-day difference---the horizon at which an attention shock should be most66visible---is not. The pattern of a difference that accumulates over a year and a half, rather than67appearing within weeks of the launch as in \citet{eisfeldt2023generative}, is more naturally read68as the gradual sector-level repricing that the annual results also display than as a discrete69attention shock, and at such horizons the market-model benchmark is coarse \citep{fama1998}.7071\input{tables/tab_event_cda}7273\begin{figure}[htbp]\centering74\includegraphics[width=\textwidth]{fig8_event_car.png}75\caption{\textbf{Cumulative abnormal returns around ChatGPT.} The left panel plots the mean76market-model cumulative abnormal return (CAR) by pre-event AI tercile over the event window; the77right panel plots the High$-$Low CAR spread. The tercile ordering is monotone and the spread78widens gradually, reaching about $+10$ percentage points one year after the launch. All terciles79drift negative because the market model was estimated in the 2022 bear market; only differences80between terciles are interpretable.}\label{fig:event}81\end{figure}8283\begin{figure}[htbp]\centering84\includegraphics[width=0.66\textwidth]{fig13_event_binscatter.png}85\caption{\textbf{One-year abnormal return versus pre-event AI exposure.} Binned scatter, with86fitted line, of each firm's one-year cumulative abnormal return after ChatGPT against its pre-event87AI exposure (filings of calendar 2021). The relation is positive and smooth across the exposure88distribution.}\label{fig:eventbin}89\end{figure}9091\subsection{Difference-in-differences and the cross-sectional nature of the effect (H4)}92Table~\ref{tab:did} reports the difference-in-differences estimates of equation~\eqref{eq:did}, in93which treatment is the firm's pre-determined (2019--2021) exposure interacted with the post-202294indicator, within firm and year fixed effects. The continuous-exposure coefficient is $+2.7$95percentage points per standard deviation \tstat{1.51}; the binary version is $+1.7$ points96\tstat{0.78}. Both are positive and neither is significant at conventional levels. The97disappearance of the effect under firm fixed effects in Table~\ref{tab:panel} and the weakness of98the DiD estimates say the same thing: once the permanent component of AI exposure is absorbed,99little remains to be priced.100101\input{tables/tab_did}102103The event-time profile of equation~\eqref{eq:evt}, in Figure~\ref{fig:eventtime}, is less104favourable to the regime interpretation than the first version of this paper reported. With105standard errors clustered by firm, the premium on fixed pre-launch exposure is $+3.9$ points per106standard deviation in 2015 \tstat{2.8}, $+3.0$ in 2016 \tstat{2.3}, and $+2.3$ in 2018107\tstat{1.5}, falls to zero or below in 2017, 2019, and 2020, and is $+3.4$ in 2022 \tstat{2.2},108$+0.5$ in 2023 \tstat{0.3}, and $+6.8$ in 2024 \tstat{3.0}. The 2015 premium is as large as the1092022 premium; only 2024 stands apart. The parallel-trends assumption of the DiD design is thus not110satisfied in the strict sense, and the figure is better read as showing that firms with persistently111AI-intensive filings earned a premium in several years scattered across the sample, with an112unusually large one in the last. The first version's confidence bands, computed with year-clustered113errors, were spuriously narrow (standard errors of 0.1--0.4 points instead of 1.0--2.3), and the114statement that the profile was ``flat and near zero through 2021'' is withdrawn.115116\begin{figure}[htbp]\centering117\includegraphics[width=0.82\textwidth]{fig11_eventtime.png}118\caption{\textbf{Event-time profile of the premium on pre-determined AI exposure.} Estimated return119premium on firms' pre-determined (2019--2021) AI exposure by formation year, relative to 2021,120with 95\% confidence intervals from standard errors clustered by firm (equation~\eqref{eq:evt}). The121premium is significantly positive in 2015, 2016, 2022, and 2024, and largest in 2024; the pre-launch122profile is not flat.}123\label{fig:eventtime}124\end{figure}125126\subsection{Construction, horizon, and universe robustness}127Table~\ref{tab:robust} collects the construction checks. The post-2022 spread is essentially128unchanged whether the signal is the section-weighted mean (base, $+14.6$), the simple mean129($+13.4$), or the top-5\% tail ($+14.9$); for 8-K filings the first two coincide, so this check130speaks mainly to the 10-Q filings of the 84 firms. The spread survives replacing the three AI131signals with their first principal component, which explains 86.8\% of their common variance132(post-2022 spread $+13.4$). It is $+14.1$ points when the Magnificent~7 are excluded and $+14.5$133when the ten largest firms by market capitalization are dropped (Figure~\ref{fig:mag7}): the134post-period spread is not the footprint of a handful of trillion-dollar winners, which is135unsurprising given that its largest contributors are mid-sized memory and storage firms. Two checks136are less favourable. The six-month forward-return spread in the post-period is $+1.1$ points,137computed on the 2022 and 2023 cohorts only, so the post-2022 relation is not visible at a138six-month horizon in the cohorts that can be evaluated. And the sector-neutral composite, which139compares firms only to others within their GICS sector, produces a post-2022 spread of $+3.8$140points \tstat{0.84}, a quarter of the baseline: most of the spread operates between sectors.141142\input{tables/tab_robust}143144\begin{figure}[htbp]\centering145\includegraphics[width=0.78\textwidth]{fig10_mag7.png}146\caption{\textbf{The post-2022 spread is not a mega-cap artefact.} Post-2022 composite Q5$-$Q1147return spread for the full index, after excluding the Magnificent~7, and after excluding the ten148largest firms by market capitalization, together with the pre-period spread excluding the149Magnificent~7. The spread is essentially unchanged by the exclusions.}\label{fig:mag7}150\end{figure}151152\subsection{Composition of the corpus: the 8-K-only signal}153\label{subsec:corpus}154Because the firm-year score of equation~\eqref{eq:annual} weights filings equally, and because the15510-Q filings that were scored belong to 84 alphabetically contiguous firms and score about 0.07156higher on the AI-adoption anchor than 8-K filings, the baseline signal partly sorts firms on whether157their quarterly reports happened to be parsed. The share of firm-years containing a scored 10-Q158rises from 3\% in the bottom AI quintile to 30\% in the top quintile, and within a formation year a159scored 10-Q raises the composite $z$-score by 0.67 standard deviations \tstat{6.7}. This is a160measurement artefact, not an economic one. Panel~A of Table~\ref{tab:reviewrobust} rebuilds the161composite from 8-K filings only, using the filing-level scores exported from the database162(\texttt{data/processed/filing\_type\_scores.parquet}). The 8-K-only composite has a correlation of1630.95 with the baseline, a scored 10-Q no longer predicts it ($-0.10$ standard deviations,164$t=-1.1$), and its spreads are $+1.6$ points before 2022 and $+16.3$ after \tstat{2.10}, with165cohort values of $+6.3$, $+1.3$, and $+41.3$; its monthly long--short earns a post-period Sharpe166ratio of 1.42 \tstat{2.25}. Dropping the 84 firms altogether gives $+0.6$ before and $+15.0$167after. The post-2022 change is therefore not an artefact of the partially parsed 10-Q sample,168and the paper's results should be read as results about the language of current reports. That169the 2023 cohort turns from $-4.0$ to $+1.3$ when the 10-Q firms are removed from the signal shows,170however, that cohort-level spreads are sensitive to measurement details of this kind.171172\input{tables/tab_review_robust}173174\subsection{Entrants and survivorship}175\label{subsec:entrants}176Panel~B of Table~\ref{tab:reviewrobust} addresses the current-membership universe. Firms that177joined the index after 2010 and have no price history before July 2009 earn higher average returns178than seasoned firms (they were selected into the index partly because of those returns), but they179are spread evenly across the AI quintiles (11--12\% of each), so they do not generate the180cross-sectional pattern. Among the 416 seasoned firms the composite spread is $+1.0$ point before1812022 and $+17.3$ after \tstat{2.36}, with cohort values of $+9.1$, $+1.4$, and $+41.3$; among182firm-years with at least 36 months of price history at formation it is $+2.8$ and $+14.9$. The183check does not address the firms that left the index before 2026, for which we have no data; the184level of returns is biased upward by their absence, and a firm that disclosed heavily on AI and185subsequently failed is missing from the bottom of the return distribution.186187\subsection{Definition of the break}188Panel~C of Table~\ref{tab:reviewrobust} moves the 2021 cohort, whose holding period straddles the189launch. Ending the pre-period in 2020 gives a pre-period spread of $+2.0$ points \tstat{1.17};190starting the post-period in 2021 gives $+12.7$ points over four cohorts, of which the first three191average $+4.1$. The conclusion that the post-period average depends on the 2024 cohort is192unaffected by where the line is drawn.193194\subsection{Sector controls and the technology factor}195\label{subsec:sector}196Panel~D of Table~\ref{tab:reviewrobust} and Panel~C of Table~\ref{tab:postinf} address the concern197that AI language proxies for the technology sector. In annual cross-sectional regressions with198size, momentum, and volatility controls, the post-period slope on the composite is $+5.5$ points199per standard deviation \tstat{2.00}; with GICS-sector dummies it is $+2.9$ \tstat{1.36}, and the2002024 slope falls from $+14.3$ to $+10.2$. In pooled regressions of the winsorized forward return201with firm-clustered errors, year fixed effects give $+2.9$ \tstat{2.76} and sector-by-year fixed202effects $+1.3$ \tstat{1.22}. In the monthly series the long--short loads 0.47 on the203technology-sector factor after 2022 and its FF6 alpha falls from $+5.2\%$ to $+1.9\%$ per year204\tstat{0.86} when the factor is added; the pre-period FF6 alpha with the sector factor is205$-0.4\%$. A within-sector premium on AI language after 2022 therefore exists in point estimate but206is not statistically distinguishable from zero in any of these specifications; the statistically207detectable part of the post-2022 change is the between-sector part.208209\subsection{An alternative performance outcome: risk}210Performance is not only about mean returns. High-AI firms exhibit modestly higher forward realized211volatility, and the gap widens after 2022 (a Q5$-$Q1 difference of about 2.9 annualized volatility212points, $t=3.1$, against 2.0 points before). Part of the post-2022 return differential is213compensation for the greater volatility of AI-exposed firms during a period of rapid technological214repricing.215