SPB Git
spboucher.ai

spb/wp17_uqo

Public

FIRC0002 Stage E (Agent 10): §3 data — encoder window (AI_C-17, THEO-14), SPID provenance (ADVE-16, AI_C-13), universe conditional on survival (ADVE-01, EDIT-04, DATA-03), market-cap rebuild and return hygiene BEFORE→AFTER (ECON-01, EDIT-01, DATA-02, DATA-04), winsorization as applied (AI_C-01, ECON-05); §4 methods — inference rules, break tests, event-time with main effect (THEO-02, ECON-09), contrasts and MDEs (EDIT-03, ECON-10), multiple testing (EDIT-07, ADVE-10), event study day 0 and five benchmarks (ECON-08, EDIT-06, ADVE-02); duplicated sentence and unreferenced equations fixed (EDIT-11)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed (Nov 5, 2026) parent 9740288 Browse files

2 changed files +366 −212

2 changed files
Unified Split
modified paper/sections/data.tex +200 −131
@@ -1,39 +1,52 @@
1 1 % ============================================================================
2 +% Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, sections/macros_stageE.tex.
3 +% Exhibits: self-contained floats generated by scripts/r3k_20_tables_v2.py (tables/) and
4 +% scripts/r3k_19_figures_v2.py (figures_r3k/), placed here behind \ExhibitOnce; appendix-bound
5 +% exhibits (tables/EXHIBIT_MAP.json) are printed by tables/appendix_exhibits.tex.
2 6 \section{Data and Variable Construction}
3 7 \label{sec:data}
4 8 % ============================================================================
5 9
6 −\subsection{The scored filing corpus}
10 +\subsection{The scored filing corpus and the encoder's input window}
7 11 \label{subsec:corpus_data}
8 12 The backbone of the study is the UQO EDGAR Research Database, an infrastructure built by the
9 −authors that downloads, parses and semantically scores SEC filings. Its first release (June 2026)
10 −covered the 500 current S\&P~500 constituents and had scored only their Form~8-K current reports
11 −and the Form~10-Q reports of 84 firms. On 27--28 October 2026 the database was rebuilt on a broad
12 −issuer universe (Section~\ref{subsec:universe}) with every Form~10-K, 10-Q and 8-K filed from
13 −January 2010 to 27~October 2026: \mnFilingsCorpus{} filings of \mnIssuersCorpus{} issuers, of which
14 −all but sixteen were parsed. The parsing layer splits each document into its canonical sections
15 −where a section template exists (the annual report into Items 1 to 9B, the quarterly report into
16 −its financial statements, MD\&A, quantitative disclosures, controls, legal proceedings, risk
17 −factors and unregistered sales; the current report is kept as a single body) and then into
18 −overlapping chunks of about 500 words. The embedding layer maps every chunk---\num{10720890} in
19 −all---to a 384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence transformer
20 −\citep{reimers2019sbert}. Table~\ref{tab:corpus} describes the corpus. Unlike the earlier panel, the
21 −firm-year signal now draws on all three forms for essentially every firm-year; a typical firm-year
22 −contains about fifteen filings and three hundred chunks. The companion methodological paper (UQO
23 −Working Paper No.~0001) documents the acquisition, parsing and extraction pipeline, the universe
24 −reconciliation and the SPID project layer in detail.
13 +authors that downloads, parses and semantically scores SEC filings \citep{uqowp0001}. Its first
14 +release (June 2026) covered the 500 current S\&P~500 constituents and had scored only their
15 +Form~8-K current reports and the Form~10-Q reports of 84 firms. On 27--28 October 2026 the
16 +database was rebuilt on a broad issuer universe (\cref{subsec:universe}) with every Form~10-K,
17 +10-Q and 8-K filed from January 2010 to 27~October 2026: \mnFilingsCorpusN{} filings of
18 +\mnIssuersCorpusN{} issuers (\mnFilingsTenK{} annual, \mnFilingsTenQ{} quarterly and \mnFilingsEightK{}
19 +current reports), all but sixteen of them parsed. The parsing layer splits each document into its
20 +canonical sections where a section template exists (the annual report into Items 1 to 9B, the
21 +quarterly report into its financial statements, MD\&A, quantitative disclosures, controls, legal
22 +proceedings, risk factors and unregistered sales; the current report is kept as a single body)
23 +and then into overlapping chunks of \mChunkWordsMedian{} words (median; \mChunkShareFull\% of
24 +chunks are at the cap). The embedding layer maps every chunk---\mnChunksCorpus{} in all,
25 +\mnSectionsCorpus{} sections---to a 384-dimensional vector with the \texttt{all-MiniLM-L6-v2}
26 +sentence transformer \citep{reimers2019sbert}. \Cref{tab:corpus} describes the corpus; a typical
27 +firm-year contains about fifteen filings and three hundred chunks, and the signal of essentially
28 +every firm-year now draws on all three forms.
25 29
26 −\begin{table}[htbp]\centering\small
27 −\caption{\textbf{Corpus composition.} Filings downloaded and parsed by form for the issuer universe,
28 −and the share of estimation-sample firm-years whose signal includes at least one scored filing of
29 −that form.}\label{tab:corpus}
30 −\input{tables/tab_corpus}
31 −\end{table}
30 +One property of the encoder, undocumented in the first version, conditions everything that
31 +follows. \texttt{all-MiniLM-L6-v2} was run with its default maximum sequence length of
32 +\mMaxSeq{} WordPiece tokens (the tokenizer allows \mTokenizerMax), and a chunk of
33 +\mChunkWordsMedian{} words tokenizes to a median of \mTruncTokensMedian{} tokens. In a random
34 +sample of \mTruncSampleN{} chunks, \mTruncShare\% are longer than the window, so that the encoder
35 +reads about the first \mTruncWindowWords{} words of a typical chunk and \mTruncMassDropped\% of
36 +the token mass is never seen. The consequence for the AI score is direct: among chunks that
37 +contain at least one core AI term, \mTruncHitsBeyond\% of the occurrences lie beyond the window,
38 +and chunks whose AI terms all lie beyond it score \mTruncScoreBeyond{} on the adoption anchor
39 +against \mTruncScoreInside{} for chunks whose terms lie inside it. The measure is therefore a
40 +score of the opening of each chunk; re-chunking the corpus to the window, or a second encoder,
41 +would require a new embedding pass that was not executed in this revision
42 +(\cref{subsec:limits}). \Cref{subsec:construct} quantifies what the score captures.
43 +
44 +\ExhibitOnce{tab:corpus}{\input{tables/tab_corpus}}
32 45
33 46 \subsection{Anchor concepts and semantic scores}
34 47 \label{subsec:semantic}
35 48 Each of eight economic concepts is defined by three short paragraphs written in the register of a
36 −filing; the full texts are reproduced in Appendix~\ref{app:anchors}. The AI-adoption anchor, for
49 +filing; the full texts are reproduced in \cref{app:anchors}. The AI-adoption anchor, for
37 50 instance, describes a firm that ``internally uses artificial intelligence, machine learning,
38 51 generative AI, large language models, automation, predictive analytics \dots\ to improve
39 52 productivity, operations, efficiency, decision-making.'' The three paragraphs of a concept are
@@ -45,10 +58,17 @@ $\bar{\mathbf a}_a$. The score of chunk $c$ on concept $a$ is the cosine similar
45 58 \end{equation}
46 59 where $\mathbf e_c$ is the embedding of the chunk. Because similarity is computed in the embedding
47 60 space, a passage on ``deploying neural networks to forecast demand'' scores highly on AI adoption
48 −even though it shares no keyword with the anchor. The eight concepts comprise three AI
61 +even though it shares no keyword with the anchor; in a probe of written sentences, paraphrases
62 +without any AI keyword score \mProbeParaphraseAdopt{} on the adoption anchor, above the
63 +\mProbeParaphrasePct{}th percentile of corpus chunks. The eight concepts comprise three AI
49 64 facets---\emph{AI adoption}, \emph{AI infrastructure} and \emph{AI disruption risk}---and five
50 −placebos: \emph{ESG commitment}, \emph{climate risk}, \emph{cybersecurity}, \emph{automation/labour}
51 −and \emph{macro risk}.
65 +placebos: \emph{ESG commitment}, \emph{climate risk}, \emph{cybersecurity},
66 +\emph{automation/labour} and \emph{macro risk}. Two features of the anchors matter for
67 +interpretation: the adoption anchor lists cybersecurity and automation among the uses of AI, and
68 +two placebo anchors (cybersecurity and automation/labour) mention AI explicitly, so that the AI
69 +and placebo scores share vocabulary by construction (\cref{app:anchors}). The anchors were
70 +written by the authors in June 2026 with the vocabulary of that date and applied to filings back
71 +to 2010.
52 72
53 73 Chunk scores are aggregated to the filing level in three ways. The simple mean averages all chunks
54 74 of filing $f$; the top-5\% tail averages the chunks at or above the 95th percentile of the filing's
@@ -61,68 +81,124 @@ they appear,
61 81 where $\sigma(c)$ is the section of chunk $c$ and the weights emphasize MD\&A and risk-factor text
62 82 for the relevant concept (for AI adoption and infrastructure, MD\&A receives weight 2.0 and risk
63 83 factors 1.2; for AI disruption risk, risk factors 2.0 and MD\&A 1.5; financial statements 0.5; the
64 −undivided body of an 8-K 1.0). Filing-level scores are averaged, without weights, over all scored
65 −filings dated in calendar year $Y$:
84 +undivided body of an 8-K 1.0). Filing-level scores of \cref{eq:wmean} are averaged, without
85 +weights, over all scored filings dated in calendar year $Y$:
66 86 \begin{equation}
67 87 \text{AI}_{i,Y,a} \;=\; \frac{1}{|\mathcal F_{i,Y}|}\sum_{f\in\mathcal F_{i,Y}} \bar s^{\,w}_{f,a}.
68 88 \label{eq:annual}
69 89 \end{equation}
70 −Because a firm-year now contains annual, quarterly and current reports, equation~\eqref{eq:annual}
71 −blends three kinds of text; Section~\ref{subsec:corpus} rebuilds the signal separately from each
72 −form and shows that they are far from interchangeable. Our three primary signals are
73 −\texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}), \texttt{ai\_infrastructure\_wmean}
74 −(\textbf{AI\,Infra}) and \texttt{ai\_disruption\_risk\_wmean} (\textbf{AI\,Risk}); we also form a
75 −standardized composite, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$.
90 +Because a firm-year contains annual, quarterly and current reports, \cref{eq:annual} blends three
91 +kinds of text, and because current reports are the most numerous, the unweighted average is
92 +dominated by them: the composite correlates \mcorrEightK{} with a composite built from 8-K
93 +filings alone and \mcorrTenK{} with one built from 10-K filings alone. \Cref{subsec:corpus}
94 +rebuilds the signal separately from each form and adds an equal-form-weight composite. Our three
95 +primary signals are \texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}),
96 +\texttt{ai\_infrastructure\_wmean} (\textbf{AI\,Infra}) and \texttt{ai\_disruption\_risk\_wmean}
97 +(\textbf{AI\,Risk}); the headline composite is
98 +$\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$, which excludes the
99 +disruption-risk facet. Because that facet turns out to carry the largest spreads, every headline
100 +table also reports AI\,Risk and the first principal component of the three facets
101 +(\mpcaEVR\% of their variance), and the verdicts are stated for each.
76 102
77 103 \subsection{A second measure: disclosed AI projects}
78 104 \label{subsec:spid}
79 −The SPID project database, built on the same corpus, submits every eligible filing section to a
80 −large-language-model reader with a strict JSON schema and extracts discrete corporate
81 −projects---named initiatives with a type drawn from a 21-category taxonomy, a status, cited years,
82 −locations, partners and amounts. We use two firm-year counts from it: the number of mentions of
83 −technology projects (AI initiatives, data centers, cloud migrations, automation, digital
84 −transformation, ERP implementations and technology deployments; \textbf{SPID\,Tech}, entered as
85 −$\log(1+n)$ and standardized within year), and an indicator for at least one AI-initiative project
86 −or AI-tagged technology (\textbf{Any AI project}). The two measures capture different objects: the
87 −semantic score measures how much a firm's text \emph{resembles} AI language, the project count
88 −how many concrete AI-related initiatives it \emph{describes}. Projects are sparse: only
89 −\mspidAnyShare\% of firm-years contain an AI-initiative mention (\mspidAnyPostShare\% in the post-launch
90 −cohorts), and about a quarter contain any technology project, so the SPID\,Tech quintile sort
91 −effectively separates firms that describe such projects from those that do not.
105 +The SPID project database, built on the same corpus \citep{uqowp0001}, submits every eligible
106 +filing section to a large-language-model reader (\mSpidModel, temperature zero, JSON-object
107 +output, run on 27--28 October 2026; the system prompt and settings are reproduced in
108 +\cref{app:spid}) and extracts discrete corporate projects---named initiatives with a type drawn
109 +from a 21-category taxonomy, a status, cited years, locations, partners and amounts. We use two
110 +firm-year counts from it: the number of mentions of technology projects (AI initiatives, data
111 +centers, cloud migrations, automation, digital transformation, ERP implementations and technology
112 +deployments; \textbf{SPID\,Tech}, entered as $\log(1+n)$ and standardized within year), and an
113 +indicator for at least one AI-initiative project or AI-tagged technology (\textbf{Any AI
114 +project}). The two measures capture different objects: the semantic score measures how much a
115 +firm's text \emph{resembles} AI language, the project count how many concrete AI-related
116 +initiatives it \emph{describes}. Projects are sparse: \mspidAnyShare\% of firm-years contain an
117 +AI-initiative mention (\mspidAnyPostShare\% in the post-launch cohorts), about a quarter contain
118 +any technology project, and the median and 75th percentile of the technology-project count are
119 +zero. A quintile sort on such a variable assigns its lower quintiles by row order, so the headline
120 +project tests are binary contrasts (any AI project, any technology project) with their minimum
121 +detectable effects; the quintile rows are kept only for comparison with the first version
122 +(\cref{subsec:spidres}). The precision and recall of the extractor have not been validated
123 +against human labels; a stratified sampling frame of \mSpidSampleN{} mentions with their source
124 +text has been drawn for that purpose and no label---human or machine---exists yet
125 +(\cref{app:spid}).
92 126
93 127 \subsection{Universe and survivorship}
94 128 \label{subsec:universe}
95 129 The universe is the union of the S\&P~500 membership of May 2026 and the equity holdings of the
96 −iShares Russell~3000 ETF as of 23~October 2026, mapped to SEC registrants: 2,564 issuers, of which
97 −2,499 file the three forms we use (the remainder are foreign private issuers filing 20-F or 40-F).
98 −The ETF holdings are a current-membership proxy for the Russell~3000, not a point-in-time
99 −constituent list, and the panel therefore carries the survivorship and look-ahead biases of a
100 −current universe \citep{browngoetzmannross1995}: firms that were delisted before 2026 are absent,
101 −and \mnEntrantFirms{} of the \mnFirms{} firms in the estimation sample have price histories that
102 −begin after July 2009 (they earn \mentrantPrem{} points more per year than the \mnSeasoned{}
103 −seasoned firms). The bias is more severe than in the S\&P~500 sample because small firms exit more
104 −often; Section~\ref{subsec:entrants} shows that entrants are spread evenly across the AI quintiles
105 −and that the results hold among seasoned firms. Within the universe we distinguish two overlapping
106 −partitions. The \emph{index partition} separates the 490 S\&P~500 members with a valid forward
107 −return (\mnSPfy{} firm-years) from the \mnonSPfirms{} non-members, so that the earlier study is
108 −nested in the present one. The \emph{size partition}, re-assigned each formation year, separates
109 −the largest thousand firms by market capitalization (\mnLarge{} firm-years, \mnLargeFirms{} distinct
110 −firms), a Russell~1000 proxy, from the remaining firms (\mnSmall{} firm-years, \mnSmallFirms{}
111 −firms), a Russell~2000 proxy.
130 +iShares Russell~3000 ETF (IWV) as of 23~October 2026, mapped to SEC registrants: 2,564 issuers,
131 +of which \mnIssuersCorpus{} file the three forms we use (the remainder are foreign private issuers
132 +filing 20-F or 40-F). The ETF holdings are a current-membership proxy for the Russell~3000, not a
133 +point-in-time constituent list, and the panel therefore carries the survivorship and look-back
134 +selection of a current universe \citep{browngoetzmannross1995}: firms that were delisted before
135 +October 2026 are absent, the cohorts grow from \mnObsFirstCohort{} firms in 2010 to
136 +\mnObsLastCohort{} in 2024 (\cref{tab:coverage}), and \mnEntrantFirms{} of the \mnFirms{} firms
137 +have price histories that begin after July 2009 (they earn \mentrantPremW{} points more per year,
138 +winsorized, than the \mnSeasoned{} seasoned firms). \Cref{subsec:survivorship} measures the
139 +selection with archived IWV holdings: of the constituents of May 2019, \mattritionSmallNineteen\%
140 +of the smallest market-value tercile and \mattritionLargeNineteen\% of the largest are absent from
141 +the 2025 membership, and the annualised exit rate of the small tercile is \mExitRatioMin{} to
142 +\mExitRatioMax{} times that of the large tercile across the 2019--2024 snapshots. The selection
143 +affects levels for certain and spreads by an amount that cannot be signed on this corpus, because
144 +delisted firms have no AI scores: the ``look-ahead-free'' property of the design refers to the
145 +timing of the signal relative to the return, not to the selection of the universe, and every
146 +level and spread below is conditional on survival to October 2026. Within the universe we
147 +distinguish two overlapping partitions. The \emph{index partition} separates the \mspFirms{}
148 +S\&P~500 members with a valid forward return (\mnSPfy{} firm-years) from the \mnonSPfirms{}
149 +non-members, so that the earlier study is nested in the present one. The \emph{size partition},
150 +re-assigned each formation year, separates the largest thousand firms by market capitalization
151 +(\mnLarge{} firm-years, \mnLargeFirms{} distinct firms), a Russell~1000 proxy, from the remaining
152 +firms (\mnSmall{} firm-years, \mnSmallFirms{} firms), a Russell~2000 proxy; because the panel grows
153 +over time the top-1,000 cut covers most of the 2010 cohort and less than half of the 2024 cohort,
154 +so a fixed within-cohort percentile partition (largest 42\%) is reported beside it and
155 +\cref{tab:coverage} gives the counts by cohort. \mnNoCap{} firm-years (\mShareNoCapVtwo\%) have
156 +no capitalization and belong to neither size group; they remain in the all-firm sorts.
112 157
113 −\subsection{Price, share and factor data}
158 +\subsection{Price, share and market-capitalization data}
114 159 \label{subsec:price}
115 −We download monthly adjusted-close prices for all tickers from Yahoo Finance over June 2009 to
116 −October 2026 (2,499 tickers with at least one monthly return), together with the SPDR S\&P~500 ETF
117 −(SPY) and the iShares Russell~3000 ETF (IWV) as benchmarks, and the Fama--French five factors
118 −\citep{famafrench2015}, the momentum factor \citep{carhart1997} and the risk-free rate from the
119 −Kenneth R.\ French data library (through September 2026). Monthly simple returns are computed from
120 −adjusted prices, which incorporate dividends and splits; a handful of non-positive prices and
121 −impossible returns are set to missing. Market capitalization at formation is the price times the
122 −number of shares outstanding most recently reported on the cover of a 10-K or 10-Q filed on or
123 −before the formation date, taken from the SEC's XBRL company-facts API; this replaces the
124 −constant-shares proxy of the earlier study and is available for every firm. Daily prices from June
125 −2021 to June 2024 are used for the event study, with IWV as the market index.
160 +We download monthly prices for all tickers from Yahoo Finance over June 2009 to October 2026
161 +(\mnIssuersCorpus{} tickers with at least one monthly return), together with the SPDR S\&P~500 ETF
162 +(SPY) and IWV as benchmarks, and the Fama--French five factors \citep{famafrench2015}, the momentum
163 +factor \citep{carhart1997} and the risk-free rate from the Kenneth R.\ French data library (through
164 +September 2026). Monthly simple returns are computed from adjusted closes, which incorporate
165 +dividends and splits; the pre-review version removed only returns below $-99\%$, non-finite values
166 +and non-positive prices. The review found that this filter let price splices through---the worst
167 +being a Yahoo series that restarts after a chapter-11 reorganization and prints a monthly return
168 +of several hundred times---so every monthly return above 300\% (\mhygieneMonthlyFlagged{}
169 +observations) and every firm-year forward return above 500\% (\mhygieneFYFlagged) was checked
170 +against a second, licensed daily price source (FinData OS, FirstRate data). \mHygieneConfirmedMonthly{}
171 +monthly and \mHygieneConfirmedFY{} firm-year returns are genuine and kept; \mhygieneMonthlyBreaks{}
172 +monthly and \mhygieneFYBreaks{} firm-year observations are series breaks and are set to missing
173 +together with every forward return whose window contains them (\mhygieneFYRemoved{} firm-years;
174 +\cref{tab:hygiene} lists each one). The panel is keyed on the SEC registrant (CIK) and the
175 +formation year rather than on the ticker, which removes \mnDedupDropped{} duplicate rows (a filer
176 +listed under two tickers or two share classes), and a firm receives a forward return only if it
177 +was listed at formation.
178 +
179 +Market capitalization at formation was rebuilt. The pre-review variable multiplied a split- and
180 +dividend-adjusted Yahoo close by the as-reported cover-page share count, so that the product was
181 +wrong by the cumulative factor of every later split or reverse split and by raw share-count
182 +errors; its logarithm ranged up to \mLogCapMaxVone{}, Apple's June-2011 capitalization was
183 +\mAaplCapVone{} billion dollars (rank \mAaplRankVone{} in its cohort), and \mnCapsGtThreeTrillionVone{}
184 +firm-years exceeded three trillion dollars. Yahoo's unadjusted close is itself split-adjusted, so
185 +the corrected variable is the unadjusted month-end close from the second price source times the
186 +point-in-time number of shares outstanding reported on the cover of the most recent 10-K or 10-Q
187 +filed on or before the formation date (SEC XBRL company-facts, summed across share classes,
188 +with cover-page counts flagged as thousand-fold unit errors set to missing and the previous valid
189 +count carried forward), on the same split basis as the price. Screens remove capitalizations more
190 +than ten times or less than a tenth of the firm's trailing twelve-month median and those not
191 +observed within three months of formation. The corrected logarithm ranges from \mlogCapMin{} to
192 +\mlogCapMax{}, Apple's June-2011 capitalization is \mAaplCapVtwo{} billion dollars (rank
193 +\mAaplRankVtwo), \mnCapsGtThreeTrillionVtwo{} firm-years exceed three trillion dollars (all of
194 +them Apple, Microsoft and Nvidia in 2022--24), the share of firm-years without a capitalization
195 +falls from \mShareNoCapVone\% to \mShareNoCapVtwo\%, and \msizeGroupChange\% of firm-years change
196 +size group. The value-weighted exhibits of the first version were single-stock artefacts---the
197 +largest weight in a cohort-by-quintile cell was \mvwMaxWeightVone\%---and are replaced
198 +(\cref{subsec:sorts}); the largest weight is now \mvwMaxWeight\%, which is genuine mega-cap
199 +dominance, so a weight-capped variant is reported beside the value-weighted rows. Daily prices
200 +from January 2015 to June 2024 are used for the event study and its placebo launches, with IWV as
201 +the market index and the daily Fama--French factors for the calendar-time benchmark.
126 202
127 203 \subsection{Timing and the construction of tradeable returns}
128 204 \label{subsec:timing}
@@ -134,17 +210,21 @@ forward return is the buy-and-hold return from July of $Y{+}1$ through June of $
134 210 r_{i,Y+1} \;=\; \prod_{m=\text{Jul}\,Y+1}^{\text{Jun}\,Y+2}\bigl(1+r_{i,m}\bigr)-1 ,
135 211 \label{eq:fwd}
136 212 \end{equation}
137 −computed when at least six monthly returns are available. The gap between the last filing of year
138 −$Y$ and portfolio formation is at least six months, which rules out look-ahead bias. The signal of
139 −the 2022 cohort includes filings dated after the ChatGPT launch of 30~November 2022; the
140 −difference-in-differences design of Section~\ref{sec:robustness}, which fixes exposure at its
141 −2019--2021 level, is immune to this. The holding period of the 2021 cohort straddles the launch; we
142 −treat it as pre-launch and show that moving it does not matter. After requiring a valid forward
143 −return we obtain \mnObs{} firm-years (\mnFirms{} firms) over formation years 2010--2024; the
144 −forward return of the 2024 cohort runs to June 2026, and filing years 2025--2026 contribute
145 −signals but no forward return.
213 +computed when at least six monthly returns are available; only \mnLtTwelve{} firm-year of the
214 +estimation sample has fewer than twelve, and requiring twelve leaves the headline spread at
215 +\mspreadFullStrict{} points. The gap between the last filing of year $Y$ and portfolio formation
216 +is at least six months, which rules out look-ahead in the timing of the signal. The signal of
217 +the 2022 cohort includes filings dated after the ChatGPT launch of 30~November 2022; they are
218 +\mcohortTwentyTwoPostShare\% of that cohort's chunks, and the 2022 spread computed on pre-launch
219 +filings only is \mcohortTwentyTwoPreOnly{} points against \mcohortTwentyTwoFullSame{} on the
220 +full-year text. The holding period of the 2021 cohort (July 2022 to June 2023) straddles the
221 +launch; we treat it as pre-launch and report the alternative split, which matters for the
222 +pre-launch premium (\cref{subsec:breakdef}). After requiring a valid forward return we obtain
223 +\mnObs{} firm-years (\mnFirms{} firms) over formation years 2010--2024; the forward return of the
224 +2024 cohort runs to June 2026, and filing years 2025--2026 contribute signals but no forward
225 +return.
146 226
147 −\subsection{Controls and standardization}
227 +\subsection{Controls, standardization and winsorization}
148 228 \label{subsec:controls}
149 229 For each formation year we compute cross-sectional $z$-scores of every signal,
150 230 \begin{equation}
@@ -152,58 +232,47 @@ For each formation year we compute cross-sectional $z$-scores of every signal,
152 232 \label{eq:z}
153 233 \end{equation}
154 234 so that coefficients are expressed per cross-sectional standard deviation and the secular rise of
155 −AI language is removed. In the S\&P~500 analyses the standardization is redone within the index
156 −sample, so that each panel of the two-sample tables is a self-contained replication. We control for
157 −\emph{size} (log market capitalization at formation), \emph{momentum} (the own return over the
158 −twelve months ending at formation, skipping the last month) and \emph{volatility} (the annualized
159 −standard deviation of monthly returns over the trailing year). In regressions, returns and
160 −continuous controls are winsorized at the 1st and 99th percentiles; portfolio sorts use raw returns
161 −and report winsorized spreads as a check. Appendix~\ref{app:variables} gives precise definitions.
235 +AI language is removed (\cref{eq:z} is applied to every signal and control). In the S\&P~500
236 +analyses the standardization is redone within the index sample, so that each panel of the
237 +two-sample tables is a self-contained replication. We control for \emph{size} (log market
238 +capitalization at formation), \emph{momentum} (the own return over the twelve months ending at
239 +formation, skipping the last month) and \emph{volatility} (the annualized standard deviation of
240 +monthly returns over the trailing year). The first version stated that regressions used
241 +winsorized returns and controls while its code used raw returns throughout; the revision applies
242 +the stated rule and reports both. Forward returns and continuous controls are winsorized at the
243 +1st and 99th percentiles \emph{within formation year}, so that the bounds of early cohorts do not
244 +depend on later ones; every sort table reports the raw, winsorized and median spreads side by
245 +side, every regression table the raw and winsorized slopes, and the text quotes the winsorized
246 +specification unless stated otherwise. \Cref{app:variables} gives precise definitions.
162 247
163 248 \subsection{Summary statistics}
164 249 \label{subsec:summary}
165 −Table~\ref{tab:summ} reports summary statistics for the estimation sample and Table~\ref{tab:corr}
166 −the correlations of the standardized signals. The three AI signals are collinear (the correlation of
167 −adoption with infrastructure is \mcorrAdoptInfra{} and with disruption risk \mcorrAdoptRisk),
168 −three lenses on one latent construct that we also summarize with their first principal component
169 −(\mpcaEVR\% of common variance). The SPID project count is only weakly related to the language
170 −scores (correlation \mcorrSpid{} with AI adoption): firms that describe concrete technology
171 −projects are not, in the main, the firms whose text most resembles the AI anchors. The signal is
172 −persistent: the year-over-year cross-sectional rank correlation of the composite is \mrankAuto{}
173 −and at a five-year lag \mrankAutoFive, which explains why firm fixed effects absorb most of it and
174 −motivates the pre-determined-exposure design.
250 +\Cref{tab:summ} reports summary statistics for the estimation sample, including the minimum and
251 +maximum of every variable, and \cref{tab:corr} the correlations of the standardized signals. The
252 +three AI signals are collinear (the correlation of adoption with infrastructure is
253 +\mcorrAdoptInfra{} and with disruption risk \mcorrAdoptRisk), three lenses on one latent construct
254 +summarized by their first principal component. The SPID project count is only weakly related to
255 +the language scores (correlation \mcorrSpid{} with AI adoption): firms that describe concrete
256 +technology projects are not, in the main, the firms whose text most resembles the AI anchors.
257 +The cybersecurity placebo correlates \mcorrCyber{} with the composite. The signal is persistent:
258 +the year-over-year cross-sectional rank correlation of the composite is \mrankAuto{} and at a
259 +five-year lag \mrankAutoFive, which explains why firm fixed effects absorb most of it.
175 260
176 −\begin{table}[htbp]\centering\small
177 −\caption{\textbf{Summary statistics, estimation sample.} Firm-years with a valid forward return,
178 −formation years 2010--2024. Semantic signals are section-weighted cosine similarities; returns are
179 −simple twelve-month buy-and-hold returns.}\label{tab:summ}
180 −\adjustbox{max width=\textwidth}{\input{tables/tab_summary}}
181 −\end{table}
182 −
183 −\begin{table}[htbp]\centering\small
184 −\caption{\textbf{Correlations of within-year standardized variables.}}\label{tab:corr}
185 −\input{tables/tab_corr}
186 −\end{table}
261 +\ExhibitOnce{tab:summ}{\input{tables/tab_summary}}
262 +% tab:corr, tab:coverage and tab:topfirms are appendix-bound (tables/EXHIBIT_MAP.json) and printed by tables/appendix_exhibits.tex.
187 263
188 264 \subsection{The secular rise of AI language and the firms that score highest}
189 265 \label{subsec:trend}
190 −Figure~\ref{fig:trend} plots the cross-sectional mean of the AI-adoption score by filing year for
191 −all firms, for the largest thousand and for the rest. AI language rises through the 2010s and
266 +\Cref{fig:trend} plots the cross-sectional mean of the AI-adoption score by filing year for all
267 +firms, for the largest thousand and for the rest. AI language rises through the 2010s and
192 268 inflects after 2022 in every size group, and larger firms score higher throughout; the within-year
193 269 design compares each firm to its contemporaneous peers and removes both facts.
194 −Table~\ref{tab:topfirms} lists the firms with the highest average composite over at least five
195 −years: a mixture of enterprise-technology and software firms, data-intensive services, and
196 −industrial firms, with the smaller technology names of the Russell~2000 well represented.
197 −
198 −\begin{figure}[htbp]\centering
199 −\includegraphics[width=0.9\textwidth]{fig4_trend.pdf}
200 −\caption{\textbf{The secular rise of AI language.} Cross-sectional mean of the section-weighted
201 −AI-adoption score by filing year, for all firms in the estimation sample, the largest thousand by
202 −market capitalization and the remaining firms.}\label{fig:trend}
203 −\end{figure}
270 +\Cref{tab:topfirms} lists the firms with the highest average composite over at least five years:
271 +enterprise-technology and software firms, data-intensive services and industrial firms, but also
272 +two electric utilities (IDACORP, rank \mIdaRank, and Hawaiian Electric, rank \mHeRank), whose
273 +filings describe data-center load growth and generation capacity rather than AI. Utilities are
274 +the sector with the \mUtilSectorRank{}nd-highest mean composite (\mSectorZUtil{} standard
275 +deviations, after Information Technology at \mSectorZIT) although \mKwUtilShareZero\% of their
276 +firm-years contain no core AI term; \cref{subsec:construct} returns to this.
204 277
205 −\begin{table}[htbp]\centering\small
206 −\caption{\textbf{Firms with the highest average AI composite.} Mean within-year standardized
207 −composite over the firm's formation years (at least five).}\label{tab:topfirms}
208 −\input{tables/tab_topfirms}
209 −\end{table}
278 +\ExhibitOnce{fig:trend}{\input{figures_r3k/fig4_trend}}
modified paper/sections/methodology.tex +166 −81
@@ -1,22 +1,23 @@
1 1 % ============================================================================
2 +% Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, sections/macros_stageE.tex.
2 3 \section{Methodology}
3 4 \label{sec:methodology}
4 5 % ============================================================================
5 6
6 −We bring six research designs to bear, ranging from portfolio sorts to a market-model event study, and
7 −we apply each of them to the full universe, to the largest thousand firms, and to the remaining smaller firms.
8 −A seventh element, new to this version, is a second, independent measure of AI engagement: the number of
9 −technology projects (AI initiatives, data centers, cloud migrations, automation, digital transformation,
10 −ERP and technology deployments) that the SPID project database extracts from the same filings with a
11 −large-language-model reader, which we use both as a stand-alone signal and as a control.
12 −Each carries its own identifying assumptions and vulnerabilities: a sort is transparent but silent
13 −on risk; a factor regression adjusts for risk but imposes a linear factor structure; a panel with
14 −fixed effects controls for unobserved heterogeneity but discards the cross-sectional variation that
15 −may matter most; an event study buys clean timing at the cost of a single window. A conclusion that
16 −survives all six is more credible than one resting on any one of them, but only if the inference
17 −within each design is sound, and this section is explicit about where it is fragile.
7 +We bring six research designs to bear, from portfolio sorts to a market-model event study, and
8 +apply each of them to the full universe, to the largest thousand firms and to the remaining
9 +smaller firms. A seventh element is the second, independent measure of AI engagement---the count
10 +of technology projects that the SPID database extracts from the same filings---used both as a
11 +stand-alone signal and as a control. Each design carries its own assumptions and vulnerabilities:
12 +a sort is transparent but silent on risk; a factor regression adjusts for risk but imposes a
13 +linear factor structure; a panel with fixed effects controls for unobserved heterogeneity but
14 +discards the cross-sectional variation that may matter most; an event study buys clean timing at
15 +the cost of a single window and a benchmark. None of the designs identifies a causal effect of
16 +disclosure on returns; they establish associations and their timing, and the section is explicit
17 +about where the inference within each design is fragile.
18 18
19 −\subsection{Portfolio sorts}
19 +\subsection{Portfolio sorts and inference on cohort series}
20 +\label{subsec:m_sorts}
20 21 Each formation year $Y$ we sort firms into quintiles of a within-year standardized signal and
21 22 compute the equal-weighted (EW) and value-weighted (VW) forward return of each quintile. The annual
22 23 high-minus-low spread and its time-series average are
@@ -25,21 +26,30 @@ high-minus-low spread and its time-series average are
25 26 \overline{\text{LS}} \;=\; \frac{1}{N_Y}\sum_{Y}\text{LS}_{Y},
26 27 \label{eq:spread}
27 28 \end{equation}
28 −and we test $\overline{\text{LS}}=0$ with Newey--West standard errors with three lags
29 −\citep{neweywest1987} on the series of annual spreads. The appeal of the sort is that it imposes no
30 −functional form; the EW--VW comparison indicates whether the effect is concentrated in smaller index
31 −members. The EW--VW comparison indicates whether the effect is concentrated in smaller firms; because the
32 −universe now spans the largest 1,000 firms and some 1,500 smaller ones, we also report the sorts within
33 −the two size groups and a size-neutral sort in which quintiles are formed within size group and year.
34 −The limitation is the number of cohorts: the full sample has $N_Y=15$ annual spreads, the
35 −pre-period twelve, and the post-period three. A Newey--West $t$-statistic computed on three
36 −observations has no useful sampling distribution, so for the post-period we report it only for
37 −completeness and rely instead on the cohort-by-cohort values, on leave-one-out means, on pooled
38 −firm-level regressions with standard errors clustered by firm, and on the monthly long--short
39 −series described next.
29 +and we test $\overline{\text{LS}}=0$ on the series of annual spreads of \cref{eq:spread}. The
30 +appeal of the sort is that it imposes no functional form; its weakness is the number of cohorts:
31 +the full sample has $N_Y=15$ annual spreads, the pre-period twelve, and the post-period three.
32 +Three rules govern the inference on these series. First, every mean is reported with a
33 +Newey--West $t$-statistic with three lags \citep{neweywest1987}, the plain $t$ that treats the
34 +cohorts as independent, and a 95\% $t$-interval over cohorts (a cohort-resampling bootstrap
35 +interval is reported where the two differ); with fifteen observations the Newey--West statistic
36 +is compared with a simulated Kiefer--Vogelsang fixed-$b$ critical value of \mfixedBcv{} rather
37 +than with 1.96. Second, a Newey--West statistic computed on the three post-launch cohorts has no
38 +useful sampling distribution; it is printed with its $n$ for completeness, never used to rank
39 +results, and the post-period evidence rests instead on the cohort-by-cohort values, leave-one-out
40 +means, and pooled firm-level regressions of the top- and bottom-quintile forward returns on a
41 +top-quintile indicator with cohort fixed effects, for which we report firm-clustered, two-way
42 +(firm $\times$ cohort) clustered \citep{petersen2009,thompson2011,cameron2011} and
43 +wild-cluster-bootstrap inference by cohort (Webb six-point weights, 9,999 draws). Third, every
44 +sort is reported on raw returns, on returns winsorized within formation year, and as the median
45 +spread within cohort, together with the share of the raw spread earned by the top one percent of
46 +returns. Because the universe spans the largest 1,000 firms and some 1,500 smaller ones, we also
47 +report the sorts within the two size groups and a size-neutral sort in which quintiles are formed
48 +within size group and year, and we test the difference between groups directly with the contrasts
49 +of \cref{subsec:m_contrasts}.
40 50
41 51 \subsection{Factor-model spanning of a long--short portfolio}
42 −A raw spread could be compensation for bearing known risks. We build an annually rebalanced,
52 +A raw spread could be compensation for bearing known risks. We build an annually selected,
43 53 equal-weighted, self-financing long--short portfolio (top minus bottom quintile), record its
44 54 monthly return $\text{LS}_m$, and estimate
45 55 \begin{equation}
@@ -51,18 +61,16 @@ monthly return $\text{LS}_m$, and estimate
51 61 \label{eq:alpha}
52 62 \end{equation}
53 63 with heteroskedasticity- and autocorrelation-consistent standard errors (three lags). We report the
54 −CAPM version, the Fama--French five factors plus momentum (FF6; \citealp{famafrench2015,carhart1997}),
55 −and FF6 augmented with a technology-sector factor $\text{TECH}_m$, the equal-weighted monthly return
56 −of the Information Technology constituents of the panel minus the equal-weighted return of all
57 −constituents. The sector factor is not a priced risk factor in the usual sense; it is a diagnostic
58 −that asks whether the AI long--short is distinguishable from a bet on the technology sector during
59 −the sample. The intercept $\alpha$ is the average monthly return beyond what the loadings explain.
60 −Because the monthly series has about 180 observations (36 in the post-period), it supports inference that
61 −the three annual post-period spreads cannot; we also report the Sharpe ratio of the series in each
62 −sub-period. A placebo on the break date complements the sub-period comparison: for every candidate
63 −first post-cohort between 2013 and 2022 we compute the difference between the mean spread of the
64 −cohorts from that year on and the mean of the earlier cohorts, and we report where 2022 ranks among the
65 −ten candidates.
64 +CAPM version of \cref{eq:alpha}, the Fama--French five factors plus momentum (FF6;
65 +\citealp{famafrench2015,carhart1997}), and FF6 augmented with a technology-sector factor
66 +$\text{TECH}_m$, the equal-weighted monthly return of the Information Technology constituents of
67 +the panel minus the equal-weighted return of all constituents. The sector factor is not a priced
68 +risk factor in the usual sense; it is a diagnostic that asks whether the AI long--short is
69 +distinguishable from a bet on the technology sector during the sample. Because the monthly series
70 +has about 180 observations (36 in the post-period), it supports inference that the three annual
71 +post-period spreads cannot; we report the Sharpe ratio of the series in each sub-period and, for
72 +the headline portfolio, the deflated Sharpe ratio that accounts for the number of long--short
73 +series examined in the paper.
66 74
67 75 \subsection{Two-way fixed-effects panel regressions}
68 76 We estimate
@@ -70,42 +78,56 @@ We estimate
70 78 r_{i,Y+1} = \beta\, \text{AI}_{i,Y} + \gamma' X_{i,Y} + \alpha_i + \delta_Y + \varepsilon_{i,Y},
71 79 \label{eq:panel}
72 80 \end{equation}
73 −where $r_{i,Y+1}$ is the winsorized forward return of equation~\eqref{eq:fwd}, $\text{AI}_{i,Y}$ the
81 +where $r_{i,Y+1}$ is the winsorized forward return of \cref{eq:fwd}, $\text{AI}_{i,Y}$ the
74 82 standardized signal, $X$ the standardized controls, and $\alpha_i,\delta_Y$ firm and year fixed
75 −effects. Standard errors are double-clustered by firm and year
76 −\citep{petersen2009,thompson2011,cameron2011}; with fifteen year clusters the time dimension of
77 −the clustering is thin, and we report firm-clustered versions where the distinction matters. Firm
78 −fixed effects absorb all time-invariant cross-firm differences, so $\beta$ identifies whether
79 −within-firm changes in AI language predict returns; year effects absorb common shocks. Contrasting
80 −this estimate with the cross-sectional ones isolates whether the signal lives across or within
81 −firms (H4).
83 +effects. Standard errors in \cref{eq:panel} are double-clustered by firm and year; with fifteen
84 +year clusters the time dimension of the clustering is thin, and we report firm-clustered versions
85 +where the distinction matters. Firm fixed effects absorb all time-invariant cross-firm
86 +differences, so $\beta$ measures whether within-firm changes in AI language are associated with
87 +returns; year effects absorb common shocks. Contrasting this estimate with the cross-sectional
88 +ones shows whether the signal lives across or within firms (H4).
82 89
83 −\subsection{Fama--MacBeth regressions}
90 +\subsection{Fama--MacBeth regressions and the within-sector estimators}
84 91 We also run annual cross-sectional regressions and average the slopes \citep{famamacbeth1973},
85 92 \begin{equation}
86 93 r_{i,Y+1} = a_Y + b_Y\, \text{AI}_{i,Y} + c_Y' X_{i,Y} + e_{i,Y}, \qquad
87 94 \hat b = \frac{1}{N_Y}\sum_Y \hat b_Y ,
88 95 \label{eq:fm}
89 96 \end{equation}
90 −with Newey--West errors on the time series of $\hat b_Y$. This estimator weights each year equally
91 −and yields an interpretable per-standard-deviation premium; plotting the annual slopes reveals the
92 −time profile of the premium directly, and adding GICS-sector dummies to the cross-sectional
93 −regression gives the within-sector version of the slope.
97 +with Newey--West and plain standard errors on the time series of $\hat b_Y$. The estimator of
98 +\cref{eq:fm} weights each year equally and yields an interpretable per-standard-deviation
99 +premium; plotting the annual slopes reveals the time profile of the association directly. The
100 +within-sector version of the claim is estimated three ways, because the three disagree: GICS-sector
101 +dummies in the cross-sectional regression of \cref{eq:fm}; a pooled regression with sector-by-year
102 +fixed effects and firm, two-way and wild-cluster inference; and a sector-neutral quintile sort in
103 +which quintiles are formed within sector and year, complemented by a residual sort on returns
104 +orthogonalized to sector-by-year effects and the controls. Each is reported with and without the
105 +Energy sector, whose pre-launch high-minus-low difference is the largest of any sector.
94 106
95 −\subsection{Structural-break and difference-in-differences designs}
96 −To test regime dependence (H2) we split the sample at the ChatGPT launch and compare sub-period
97 −spreads, and we estimate the pooled interaction model
107 +\subsection{Break tests, difference-in-differences and the event-time profile}
108 +\label{subsec:m_break}
109 +To examine regime dependence (H2) we split the cohorts at the ChatGPT launch and compare
110 +sub-period spreads, and we estimate the pooled interaction model
98 111 \begin{equation}
99 112 r_{i,Y+1} = \beta_1\, \text{AI}_{i,Y} + \beta_2\, (\text{AI}_{i,Y}\times \text{Post}_Y)
100 113 + \gamma' X_{i,Y} + \delta\,\text{Post}_Y + \varepsilon_{i,Y},
101 114 \qquad \text{Post}_Y=\mathbb{1}[Y\ge 2022],
102 115 \label{eq:break}
103 116 \end{equation}
104 −where $\beta_2$ is the change in the premium per standard deviation of the signal, with standard
105 −errors clustered by formation year. We complement this with a difference-in-differences design in
106 −which treatment is the firm's pre-determined exposure $\overline{\text{AI}}_i^{\,19\text{--}21}$, the
107 −average of its composite over the filings of 2019--2021, standardized across firms, so that it
108 −cannot respond to the event:
117 +where $\beta_2$ is the change in the slope per standard deviation of the signal, with firm,
118 +two-way and wild-cluster-bootstrap inference, since $\text{Post}_Y$ varies only at the cohort
119 +level. Three tests ask whether 2022 is special among break dates. The placebo of the first
120 +version---the post-minus-pre difference in mean spreads for every candidate first post-cohort
121 +from 2013 to 2022 and the rank of 2022 among the ten---compares nested windows and is kept only as
122 +an illustration. The formal tests are an exact permutation of the 2022--24 block against all
123 +\mPermSubsets{} three-cohort subsets of the fifteen cohorts and against the twelve contiguous
124 +three-cohort blocks, and an Andrews-type sup-Wald mean-shift test over all admissible break dates with 15\% trimming
125 +and a simulated finite-sample critical value.
126 +
127 +We complement the break tests with a difference-in-differences specification in which treatment
128 +is the firm's pre-determined exposure $\overline{\text{AI}}_i^{\,19\text{--}21}$, the average of
129 +its composite over the filings of 2019--2021, standardized across firms, so that it cannot respond
130 +to the event:
109 131 \begin{equation}
110 132 r_{i,Y+1} = \theta\,\bigl(\overline{\text{AI}}_i^{\,19\text{--}21}\times\text{Post}_Y\bigr)
111 133 + \gamma' X_{i,Y} + \alpha_i + \delta_Y + \varepsilon_{i,Y},
@@ -113,50 +135,113 @@ cannot respond to the event:
113 135 \end{equation}
114 136 estimated with firm and year fixed effects and double-clustered errors, and with a binary version
115 137 in which treatment is above-median pre-determined exposure. An event-time version replaces the
116 −single interaction with interactions of the fixed exposure with each formation-year dummy relative
117 −to 2021,
138 +single interaction of \cref{eq:did} with interactions of the fixed exposure with each
139 +formation-year dummy,
118 140 \begin{equation}
119 − r_{i,Y+1} = \sum_{\tau\neq 2021}\theta_\tau\,\bigl(\overline{\text{AI}}_i^{\,19\text{--}21}\times\mathbb{1}[Y=\tau]\bigr)
141 + r_{i,Y+1} = \lambda\,\overline{\text{AI}}_i^{\,19\text{--}21}
142 + + \sum_{\tau\neq 2021}\theta_\tau\,\bigl(\overline{\text{AI}}_i^{\,19\text{--}21}\times\mathbb{1}[Y=\tau]\bigr)
120 143 + \gamma' X_{i,Y} + \delta_Y + \varepsilon_{i,Y},
121 144 \label{eq:evt}
122 145 \end{equation}
123 −which traces the dynamic response and, in the years before 2022, tests the parallel-trends
124 −assumption. Because each $\theta_\tau$ is identified within a single year, standard errors for
125 −equation~\eqref{eq:evt} are clustered by firm.
146 +where the main effect $\lambda$ makes each $\theta_\tau$ a difference from the reference year 2021.
147 +The first version omitted $\lambda$, so that its event-time coefficients were absolute slopes per
148 +year rather than differences, and its figure labelled ``relative to 2021'' was not. Two further
149 +caveats make \cref{eq:evt} a descriptive event-time profile rather than an identification
150 +design. The 2019--21 exposure is measured \emph{after} the outcomes of 2014--2020, so the
151 +pre-launch coefficients mix look-back selection with any pre-trend; a variant with trailing
152 +exposure (the mean composite over the three years before each outcome year) removes this. And the
153 +profile is estimated on a current-membership universe. Because each $\theta_\tau$ is identified
154 +within a single year, standard errors are clustered by firm.
155 +
156 +\subsection{Tested contrasts and minimum detectable effects}
157 +\label{subsec:m_contrasts}
158 +The first version compared the post-launch spread of one group with that of another and read the
159 +difference off their separate $t$-statistics. The revision tests the differences. For the size
160 +and index partitions we estimate, on the top- and bottom-quintile firm-years,
161 +\begin{equation}
162 +\begin{aligned}
163 + r_{i,Y+1} \;=\;& \phi_1\,\text{hi}_{i,Y} + \phi_2\,(\text{hi}_{i,Y}\times\text{Post}_Y)
164 + + \phi_3\,(\text{hi}_{i,Y}\times G_{i,Y}) \\
165 + &+ \phi_4\,(\text{hi}_{i,Y}\times\text{Post}_Y\times G_{i,Y})
166 + + \text{lower-order terms} + \delta_Y + \varepsilon_{i,Y},
167 +\end{aligned}
168 + \label{eq:contrast}
169 +\end{equation}
170 +where $G$ is the group indicator (largest thousand, or S\&P~500 member) and $\phi_4$ is the
171 +difference between the groups' post-minus-pre changes in the high-minus-low spread, with
172 +firm-clustered, two-way and wild-cluster inference and a cohort-block bootstrap of the same
173 +difference computed from the cohort spreads. Every contrast, every post-minus-pre difference and
174 +every project-count estimate is accompanied by its minimum detectable effect at 80\% power
175 +(2.8 times the firm-clustered standard error), so that a non-rejection can be read as a bound
176 +rather than as evidence of equality.
177 +
178 +\subsection{Multiple testing and the specification curve}
179 +\label{subsec:m_mht}
180 +The paper reports on the order of two hundred $t$-statistics. Following \citet{harveyliuzhu2016},
181 +we define the families within which the tests are exchangeable---the full-sample and post-period
182 +spreads of \cref{tab:sortsboth,tab:prepostboth}, the eight placebo interactions of
183 +\cref{tab:placebo}, the ten break dates, the event sub-samples and horizons, the three filing
184 +forms, and the robustness rows of \cref{tab:robust}---and report Holm, Benjamini--Hochberg and
185 +Romano--Wolf step-down adjusted $p$-values within each family, the latter on shared
186 +cohort-resampling or wild-cluster bootstrap draws. We also report which statistics exceed the
187 +$|t|>3$ hurdle, the deflated Sharpe ratio of the headline long--short given the \mdsrTrials{}
188 +long--short series examined, and a specification curve of the full-sample equal-weighted
189 +composite spread over winsorization, weighting, horizon, filing form, aggregation and size filter
190 +(\mspecN{} specifications), with the share that is positive and significant.
126 191
127 192 \subsection{Market-model event study}
193 +\label{subsec:m_event}
128 194 Finally, we run an event study \citep{mackinlay1997} around the ChatGPT launch, which converts the
129 195 annual-frequency portfolio evidence into a daily, cleanly dated test. The launch was a
130 196 consumer-product event, unanticipated by equity markets, that raised the salience of AI without
131 −itself changing any firm's fundamentals overnight. For each stock we estimate the market model
197 +itself changing any firm's fundamentals overnight. Day~0 is the first trading day after the
198 +launch, 1~December 2022; the first version's window began on 1~November 2022, so that its
199 +``60-day'' statistic contained twenty pre-launch days and every horizon ended a month earlier
200 +than labelled, and the block $[-20,-1]$ is now reported separately. For each stock we estimate
201 +the market model
132 202 \begin{equation}
133 203 r_{i,d}=a_i+b_i\, r^{m}_{d}+e_{i,d}
134 204 \label{eq:mm}
135 205 \end{equation}
136 −over the estimation window 3~January--31~October 2022 (209 trading days), with the iShares Russell~3000
137 −ETF (IWV) as the market index, and
138 −compute abnormal and cumulative abnormal returns over the event window 1~November 2022--30~June
139 −2024,
206 +over the estimation window 3~January--31~October 2022 (209 trading days), with IWV as the market
207 +index, and compute abnormal and cumulative abnormal returns over the event window,
140 208 \begin{equation}
141 209 \text{AR}_{i,d}=r_{i,d}-(\hat a_i+\hat b_i r^m_d), \qquad
142 210 \text{CAR}_i(T)=\sum_{d=1}^{T}\text{AR}_{i,d}.
143 211 \label{eq:car}
144 212 \end{equation}
145 213 CARs are compared across terciles of pre-event AI exposure measured from the filings of calendar
146 −year 2021, all dated before the launch, within the full universe and within size groups, and firm-level
147 −CARs are regressed on exposure with size, beta, and sector controls. A firm-level Welch $t$-test treats the
148 −firms as independent, which they are not when they share an event date
149 −\citep{brownwarner1985,kolari2010}. We therefore also treat the equal-weighted High-minus-Low
150 −abnormal return as a single daily series $\text{AR}^{HL}_d$ and compute
214 +year 2021, all dated before the launch, within the full universe and within size groups (terciles
215 +formed within group as well as across), and firm-level CARs are regressed on exposure with size,
216 +beta and sector controls. A firm-level Welch $t$-test treats the firms as independent, which they
217 +are not when they share an event date \citep{brownwarner1985,kolari2010}; the first version
218 +reported Welch statistics for every sub-sample while computing a dependence-robust statistic
219 +only for the full sample. We therefore treat the equal-weighted High-minus-Low abnormal return as
220 +a single daily series $\text{AR}^{HL}_d$ and compute, for every sample and horizon,
151 221 \begin{equation}
152 222 t_{\text{CDA}}(T)=\frac{\sum_{d=1}^{T}\text{AR}^{HL}_d}{\hat\sigma_{HL}\sqrt{T}},
153 223 \label{eq:cda}
154 224 \end{equation}
155 225 where $\hat\sigma_{HL}$ is the daily standard deviation of the same portfolio series in the
156 226 estimation window (the crude dependence adjustment of \citealp{brownwarner1985}) or, to allow for
157 −event-induced variance \citep{boehmer1991}, in the event window. We also report the average
158 −pairwise correlation of estimation-window residuals, which governs the overstatement of the
159 −independence-based test \citep{kolari2010}. Horizons of 250 and 395 trading days are long for an
160 −event study; at such horizons the market model's intercept, estimated in the 2022 bear market,
161 −makes all CARs drift negative, and only differences between terciles are interpretable
162 −\citep{fama1998,kothariwarner2007}.
227 +event-induced variance \citep{boehmer1991}, in the event window; the standardized-CAR statistic of
228 +\citet{boehmer1991} is reported for reference. The average pairwise correlation of
229 +estimation-window residuals, $\bar\rho=\mevRho$, implies a Kolari--Pynn\"onen deflation factor of
230 +\mevKP{} for a single-portfolio independence-based test \citep{kolari2010}; it is not applied to
231 +the High-minus-Low statistic because the common component that the factor corrects for nets out
232 +of the difference portfolio, which is why the portfolio statistic of \cref{eq:cda} and a
233 +KP-deflated firm-level $t$ differ.
234 +
235 +The benchmark is the weak point of a long-horizon event study, and the single-index model of
236 +\cref{eq:mm} is estimated in the 2022 bear market, in which the high-exposure tercile has a
237 +higher beta, a larger capitalization and a larger technology share than the low tercile. We
238 +therefore report the one-year High-minus-Low statistic under five additional benchmarks: a
239 +beta-only adjustment ($r-\hat b\,r^m$), a market-adjusted return ($r-r^m$), the market model
240 +estimated over calendar 2021, a size-and-momentum-matched benchmark (5 $\times$ 5 cells of the
241 +2021-cohort panel; book-to-market is not in the panel), and the calendar-time alpha of the
242 +High-minus-Low portfolio on the daily five Fama--French factors plus momentum over the 250 days
243 +after the launch. Seven placebo launches (30~November of 2015 to 2021) run the identical design.
244 +Horizons of 250 and 395 trading days are long for an event study; at such horizons the
245 +market-model intercept makes all CARs drift and only differences between terciles are
246 +interpretable \citep{fama1998,kothariwarner2007}, and the benchmark dependence of those
247 +differences is itself a result.
163 248
Menu

Theme
↑↓ navigate ↵ open esc close / or ⌘K anywhere