Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models

1 Introduction

• GPT-4 scores predict continued price drift for one to two trading days, generating 34 basis points per day before transaction costs, especially for small stocks and negative news.A long-short strategy based on overnight-news assessments is most feasible for participants with sufficiently low transaction costs, such as market makers.Source: §1 · p. 3
• The paper introduces a theoretical framework combining LLM technology, information-processing constraints, underreaction, and limits to arbitrage to explain these predictability patterns.Its off-the-shelf approach requires no labeled data or explicit financial training and provides interpretable assessments of headline implications.Source: §1 · pp. 2–3
• Markets process news heterogeneously: earnings, partnerships, and clinical-trial announcements show little drift, whereas insider transactions and specialized conference presentations exhibit significant underreaction.GPT-4 aligns with initial reactions across these topics, but subsequent drift is concentrated where markets are slower to incorporate the information.Source: §1 · p. 4
• Forecasting ability generally increases with model size: GPT-4’s drift strategy has a 2.97 Sharpe ratio versus 1.66 for GPT-3.5 and negative Sharpe ratios for most basic models.GPT-1, GPT-2, and Llama2-7b have low initial-reaction hit rates and no significant drift predictability, while GPT-3.5 and DistilBART-MNLI remain weaker than GPT-4.Source: §1 · pp. 4–5
• Within model families, larger models perform best on information requiring complex synthesis, with GPT-4 advantages of 25.2 bps for insider-transaction initial reactions and 13.8 bps for drift.Healthcare conference participation shows a 20.4 bps drift advantage, while Llama2-70b outperforms Llama2-7b by 316 bps for earnings announcements.Source: §1 · p. 5
• A Sharpe ratio of 2.63 for intraday news contrasts with 1.71 for OpenAI embeddings, showing that GPT-4 maintains performance when embedding models deteriorate with limited training data.The comparison uses 26,109 intraday observations for the embedding result and rolling six-month training windows; GPT-4’s zero-shot reasoning does not depend on massive samples.Source: §1 · p. 5
• 6.54 in 2021Q4 falls to 3.68 in 2022, 2.33 in 2023, and 1.22 in January–May 2024 as GPT models’ capabilities and adoption increase.The paper presents this decline as suggestive evidence that LLM adoption improves information processing and reduces return predictability.Source: §1 · pp. 5–6
• GPT-4 outperforms GPT-3.5 on complex synthesis tasks but performs equally on formulaic announcements, mapping market information-processing bottlenecks to scarce reasoning capacity.Markets process transparent, quantifiable information efficiently but underreact to information requiring complex synthesis; greater model sophistication benefits topics requiring deeper reasoning or specialized knowledge.Source: §1 · p. 6
• The paper argues that a model-complexity threshold and supervised-learning limitations indicate underreaction reflects scarce semantic reasoning capacity rather than merely insufficient statistical sophistication.GPT-4 succeeds where simpler models and domain-specific approaches fail, suggesting that the analytical sophistication needed to anticipate market responses is relatively rare among investors.Source: §1 · p. 6

Related Literature

• The paper tests off-the-shelf LLMs’ accessible and interpretable extraction of economic content, uses them to study market information processing, and provides a conceptual framework for their market impact.It distinguishes this conversational approach from supervised embedding-based forecasting, which offers greater customization and optimization for specific applications.Source: Related Literature · p. 7
• The paper contributes a first theoretical framework linking LLMs to market dynamics, information incorporation, and financial decision-making, while documenting an accessible financial task where off-the-shelf LLMs perform well.Its interpretability analysis examines LLM performance across topics and may inform development of models tailored to financial information processing and return forecasting.Source: Related Literature · pp. 7–8

Model background

• ChatGPT is an OpenAI large-scale language model based on the GPT Transformer architecture, which achieves advanced performance across natural language processing tasks.GPT uses a multi-layer neural network pretrained on extensive text data and fine-tuned for specific language tasks.Source: §2 · p. 8
• GPT’s transformer blocks use self-attention to process long text sequences, identify relevant contextual information, and generate more accurate and coherent responses.The architecture models language structure and patterns through deep learning and unsupervised pretraining on large text corpora.Source: §2 · pp. 8–9

Research scope

• ChatGPT is designed for general-purpose language tasks and is not explicitly trained to predict stock returns.Its capabilities include translation, summarization, question answering, and coherent text generation.Source: §2 · p. 9

Comparative models

• The study compares ChatGPT with BERT, BART, DistilBart-MNLI, FinBERT, and Llama2 to assess whether model complexity affects textual stock-return prediction.Appendix A provides an overview of the 12 LLMs studied.Source: §2 · p. 9

3.1 Key Ingredients

• The formal framework combines heterogeneous information-processing capacity, delayed belief updating, and noise-trader risk to explain temporary deviations from fully informed prices.Attentive agents process news quickly, whereas inattentive agents initially grasp only part of its implications; noise trading makes arbitrage risky and allows mispricing to persist.Source: §3.1 · p. 10

3.2 LLMs as Information Processors

• LLM information-processing precision rises with model sophistication k but falls with news complexity c, represented by λ(c, k), explaining heterogeneous performance across models and news.The framework explicitly treats LLMs as information processors whose interpretation precision depends on both model characteristics and the complexity of the information.Source: §3.2 · p. 10

3.3 Mechanism of Predictability

• LLM predictability arises when an LLM interprets news more accurately or quickly than the marginal human trader, causing prices to drift toward the LLM’s more accurate post-news valuation.Noise-trader risk prevents convergence from being instantaneous or riskless for attentive arbitrageurs.Source: §3.3 · pp. 10–11

3.4 Key Theoretical Insights

• Only models above a critical sophistication threshold k∗ generate robust profitable predictions, with the threshold higher in markets where attentive participation or trading volume accelerates information incorporation.Noisier markets with greater risk aversion or noise trading can leave profitable opportunities for moderately capable LLMs.Source: §3.4 · p. 11
• Predictability is stronger where information frictions or arbitrage constraints are greater, including smaller or less liquid stocks and negative-news events.Fixed participation costs reduce sophisticated trading in small or illiquid stocks, while shorting costs and restrictions impede correction of negative-news mispricing.Source: §3.4 · pp. 11–12
• Widespread capable-LLM adoption improves market efficiency by empowering inattentive and attentive agents, but eventually diminishes predictability as shared signals become incorporated into prices.The resulting price-impact effect shrinks the gap between LLM assessments and market prices, making the LLM advantage disappear as adoption becomes widespread.Source: §3.4 · p. 12

• The dataset combines CRSP daily returns and prices, scraped company news, RavenPack identifiers, and cleaned TAQ intraday observations.TAQ data are used to construct minute-by-minute volume-weighted average prices.Source: §4 · p. 13
• Headlines are matched with RavenPack information to retain relevant news and correct web-scraping errors involving news, timestamps, and entity names.The matching procedure covers headlines collected from major news agencies and financial news websites.Source: §4 · p. 13
• 159,137 firm-headline-date observations cover 4,123 companies from October 2021 to May 2024, representing about 85% of CRSP firms.About 67.5% of headlines are press releases and 32.5% are news articles.Source: §4 · pp. 13–14
• 82% of headlines, or 129,953 observations, are classified as overnight news released before 9 a.m. or after 4 p.m.; the remaining 18% are intraday.The distinction reflects different entry times when forming trading strategies, with later analyses emphasizing overnight news.Source: §4 · p. 14
• The sample requires RavenPack relevance scores of 100, excludes stock-gain and stock-loss labels, and imposes event similarity days above 90 to capture new information.Only complete articles and press releases are retained.Source: §4 · p. 14

5 ChatGPT Prompt

• The researchers use the publicly available headlines with an expert-role prompt that substitutes each firm name and headline into the query.The prompt is designed specifically for financial analysis and stock recommendations.Source: §5 · p. 15
• The prompt asks ChatGPT to evaluate a headline’s short-term impact on a company’s stock price, answer “YES,” “NO,” or “UNKNOWN,” and provide one concise explanatory sentence.The headline is the only information source, and “UNKNOWN” is allowed when it lacks sufficient information.Source: §5 · p. 15
• Temperature 0 maximizes reproducibility by making the model select the highest-probability word conditional on the prompt.Temperature governs response randomness and creativity.Source: §5 · p. 15
• The Humana example shows ChatGPT 4 judging the news negatively because ending Cigna’s bid could eliminate a potential acquisition premium.This contrasts with the data vendor’s positive sentiment score of 0.65 and illustrates the role of contextual reasoning.Source: §5 · p. 16

Appendix details the prompts we use for these models.

• The study uses gpt-4-0314, whose training data ends in September 2021, to ensure an out-of-sample evaluation.Responses are converted into GPT-4 scores: “YES” = 1, “UNKNOWN” = 0, and “NO” = -1; the sample mean is 0.31 and median is zero.Source: §5 · p. 16, Appendix details the prompts we use… · p. 16
• For BERT, GPT-1, GPT-2, and Llama2, the researchers use a different strategy because these basic models cannot follow instructions or answer specific questions.GPT-1 and GPT-2 are auto-complete models, and Llama-2’s later September 2022 cutoff may overstate cross-family comparisons.Source: §5 · p. 16, Appendix details the prompts we use… · p. 16
• A robustness test on 1,000 headlines finds very similar GPT-4 recommendations when the model reasons before answering rather than answers before reasoning.The main strategy asks for the answer first and reasoning second.Source: Appendix details the prompts we use… · p. 16

6 ChatGPT and Market Information Processing

• GPT-4 scores predict returns for the next one to two trading days because prices drift in the predicted direction, particularly after negative news.This underreaction creates a post-announcement trading window, although subsequent-drift predictability is more modest than immediate-reaction predictability.Source: §6.1 · p. 18

6.1 Can ChatGPT Forecast Stock Price Movements?

• GPT-4’s long-short portfolios predict initial market reactions with 93.3% and 88.8% hit rates for overnight and intraday news, translating into mean returns of 3.06% and 4.44%, respectively.Source: §6.1 · p. 18

6.2 Post-Announcement Drift and Predictive Regressions

• Around 700% cumulative return from 2021m10 to 2024m5 shows the zero-transaction-cost long-short strategy substantially outperformed market portfolios, especially through negative-news positions.Both long and short legs contribute, but the effect is stronger among stocks with negative news.Source: §6.2 · pp. 19–20
• Over 300% cumulative return remains at 5 bps round-trip costs and above 100% at 10 bps, but 20 bps makes the strategy unprofitable.Implementation therefore requires minimizing transaction costs, particularly price impact, which is challenging for large or unsophisticated investors.Source: §6.2 · p. 20
• At 10 basis points per round-trip, the 100% daily rebalancing strategy’s Sharpe ratio falls to 1.29, whereas the 25% strategy maintains a Sharpe ratio of 1.34.Reducing rebalancing from 100% to 25% lowers turnover from approximately 190% to approximately 46% per day.Source: §6.2 · p. 20
• For overnight news, switching GPT-4’s score from −1 to 1 predicts a 2.65 percentage-point increase in initial reaction and a 32.2-basis-point increase in subsequent drift.Source: §6.2 · p. 21
• Drift predictability is substantially stronger among smaller stocks, with GPT-4 interaction coefficients of 0.404 for overnight news and 0.683 for intraday news.The results are interpreted as evidence of underreaction and limits to arbitrage in firm-specific news.Source: §6.2 · p. 22
• A neutral-to-positive GPT score predicts a 30 bps return when trading begins 15 minutes after intraday news and ends at the next day’s close.The score is insignificant over the first 1–15 minutes but significant from 15 minutes to the news-day close and from the news-day close to the next close.Source: §6.2 · pp. 22–23

6.3 Information Processing Across News Types

• The news-type analysis clusters headline embeddings into topics, constructs GPT-weighted initial-reaction and drift performance scores, and regresses them on topics.This decomposition separates topics with initial underreaction from those whose information is incorporated efficiently.Source: §6.3 · pp. 23–24
• GPT-4 shows strong initial alignment when assessing earnings and revenue reports, strategic partnership announcements, and clinical trial announcements, with respective Init×GPT values of 336.7, 92.7, and 308.5 bps.Source: §6.3 · p. 24
• For insider transactions, dividends, and healthcare conferences, GPT-4 predicts both initial alignment and significant drift, revealing delayed market incorporation and information-processing frictions.Init×GPT equals 71.4, 45.5, and 103.5 bps, while Drift×GPT equals 26.3, 22.3, and 34.2 bps, respectively, with drift results significant at the 5% level or lower.Source: §6.3 · p. 24

7 Comparing Different Large Language Models

• The comparison examines FinBERT, return-trained embeddings, and LLMs of varying complexity, with a conceptual framework predicting that only models above a complexity threshold can correctly interpret news in ways aligned with markets.Source: §7 · p. 25

Appendix A.23

• GPT-4 matches GPT-3.5’s 93% initial-reaction hit rate and exceeds its 56% subsequent-drift hit rate with 58%; DistilBart-MNLI records 87% and 56%, respectively.Source: Appendix A.23 · p. 26

7.2 News Complexity and News Types

• GPT-4 predicts initial market reactions across all categories with 89–91% hit rates, while drift hit rates remain tightly clustered at 55–59% regardless of news complexity or source.Source: §7.2 · p. 28

7.3 Which Topics Larger LLMs Understand Better

• The topic analysis compares GPT-4 with GPT-3.5 and Llama2-70B with Llama2-7B to identify domains where greater model sophistication provides the largest advantage.Table 7 presents differential performance across news topics, including a bottom-row comparison.Source: §7.3 · p. 29
• GPT-4 outperforms GPT-3.5 by 12.2 basis points for initial reactions and 4.9 basis points for drift, while Llama2-70B exceeds Llama2-7B by 58.6 basis points initially.The gains indicate substantial benefits from parameter scale for financial text interpretation, especially for tiny stocks.Source: §7.3 · pp. 29–30
• GPT-4’s largest topic gains include a 25.2-basis-point insider-stock-transaction advantage for initial reactions, a 20.4-basis-point participation-announcement drift advantage, and 15.4-basis-point analytical-complexity advantages.Insider-stock-transaction drift also favors GPT-4 by 13.8 basis points; dividend announcements show virtually no performance difference at 0.2 basis points, and no topics significantly favor GPT-3.5.Source: §7.3 · p. 30
• Llama2-70B’s initial-reaction advantage reaches 315.7 basis points for earnings and revenue announcements, with further advantages of 139.6 for stock ratings and 85.0 for index inclusion.Across model families, the largest benefits occur for deeper analytical reasoning, medical or scientific knowledge, and complex financial interpretation.Source: §7.3 · p. 30

7.4 ChatGPT vs Embeddings

• GPT-4’s intraday drift-strategy Sharpe ratio is 2.63, versus 1.71 for OpenAI embeddings and 0.79 for MiniLM, despite supervised monthly retraining.These represent declines of 35% and 70% from GPT-4’s performance, respectively.Source: §7.4 · p. 32
• The intraday comparison uses 26,109 observations versus 105,742 overnight observations, indicating that supervised embeddings require much more data to capture post-announcement drift.GPT-4’s zero-shot semantic reasoning remains consistent despite training-data limitations.Source: §7.4 · p. 32
• GPT-4 achieves a 93.3% initial-reaction hit rate versus 84.6% for OpenAI embeddings and 80.2% for MiniLM, while its drift Sharpe ratio reaches 2.97 versus 2.53 and 1.42.The comparison uses embedding models trained with rolling windows, whereas GPT-4 operates zero-shot.Source: §7.4 · pp. 31–32

8.1 Speed of Price Response

• GPT-4 scores predict returns for two days after news, with overnight-news returns of 34 bps and 19 bps, and intraday returns of 32 bps and 30 bps, but not afterward.The evidence indicates that firm-specific news is absorbed into market prices in about two days.Source: §8.1 · p. 34

8.2 Market Efficiency

• GPT-4 strategy Sharpe ratios fell from 6.54 in 2021Q4 to 1.22 in January-May 2024, consistent with reduced underreaction and improved market efficiency as LLM adoption grew.The authors note that changing market conditions could also contribute to the decline, although its timing coincides with LLM proliferation.Source: §8.2 · p. 35
• GPT-4 assessments reveal asymmetric drift: large-cap stocks show no drift after negative news, whereas small-cap stocks continue drifting despite aligned initial reactions.The authors interpret this pattern as evidence that large-cap markets adjust efficiently while small-cap markets underreact.Source: §8.2 · p. 35

9 Conclusion

• GPT-4 captures approximately 90% of initial portfolio reactions without direct financial training, predicts one to two days of drift, and outperforms smaller models as financial reasoning emerges with scale.The conclusion also reports topic-dependent efficiency, including efficient reactions to earnings reports but underreaction to insider transactions, and presents a model incorporating LLM technology, information capacity constraints, underreaction, and limits to arbitrage.Source: §9 · pp. 35–36
• Future research can further enhance the information-processing capacity and predictive power of AI-driven models in financial economics.Source: §9 · p. 37

Panel A: Overnight News

• GPT-4 Score coefficients are 1.326***, 1.018***, 1.083***, 0.161***, 0.158***, and 0.087*** across the reported regression specifications.The regressions include date and firm fixed effects with standard errors clustered by date and firm.Source: Panel A: Overnight News · p. 52, Table 3 · p. 54
• The timeline decomposition separates initial reaction, subsequent drift, and total reaction, alongside GPT-4 alignment measures for immediate agreement, later adjustment, and ultimate vindication.It uses directional overnight headlines, excludes neutral scores, clusters headlines into BERTopic topics, and double-clusters standard errors by firm and date.Source: Panel A: Overnight News · p. 55, Table 4 · p. 55
• Appendix performance comparisons report HR-I, HR-D, SharpeLS, portfolio returns, portfolio leg sizes, Fama–French five-factor alpha, its t-statistic, and R2 for U.S. common stocks covered by news.HR-I measures initial-reaction hit rate, HR-D measures drift hit rate, and SharpeLS is the annualized long-short Sharpe ratio.Source: Panel A: Overnight News · p. 56
• Model performance is additionally partitioned by headline complexity and source, with high versus low complexity defined by the daily median Flesch-Kincaid readability score and sources split into news articles and press releases.Each category reports initial-reaction hit rate, drift hit rate, and annualized long-short Sharpe ratio.Source: Table 6 · p. 57, Panel A: Overnight News · p. 57
• Overall differences were 12.2*** and 4.9*** for GPT-4 versus GPT-3.5, and 58.6*** and 2.2*** for Llama 70B versus 7B, respectively, for initial and drift prediction.Across topics, larger models generally outperform smaller models, although some topic-specific differences are negative.Source: Panel A: Overnight News · p. 58
• Embedding baselines use MiniLM and OpenAI embeddings with multinomial logistic regression, rolling-window retraining, and separate models for initial reaction and post-announcement drift.The comparison evaluates equally weighted, daily rebalanced long-short portfolios using hit rate, mean return, and drift-strategy Sharpe ratio.Source: Table 8 · p. 59, Panel A: Overnight News · p. 59

A Model Summaries

• The model appendix surveys 12 models spanning GPT, BERT, BART, DistilBART, Llama2, and FinBERT, including GPT-3.5, GPT-4, and Llama2 variants of 7B, 13B, and 70B parameters.The study uses pre-cutoff models for the sample period; Llama2 is retained for within-family scale comparisons despite its September 2022 cutoff.Source: A Model Summaries · pp. 60, 62, 63

B Theoretical Model

• LLM signals become informative only beyond a sophistication threshold relative to news complexity, while greater adoption accelerates price incorporation and can reduce mispricing despite persistent frictions.The model links predictability to temporary divergence between LLM assessments and slower human beliefs; its magnitude depends on attentive-agent prevalence and arbitrage costs.Source: B Theoretical Model · p. 64

B.1 Agents

• Attentive and inattentive agents share CARA utility and risk aversion but differ in attention, allowing the model to study LLM adoption, return predictability, and market efficiency.The asset is in zero net supply, and attentive agents comprise fraction πA while inattentive agents comprise πI = 1 −πA of the participating population V.Source: §B.1 · p. 65

B.2 Time Dynamics

• Across three short-horizon periods, inattentive agents partially process news initially and complete updating later, enabling attentive agents to trade on temporary price discrepancies under nonfundamental risk.Inattentive agents process fractions ω and (1−ω) of attentive agents’ signal precision across the first and second periods.Source: §B.2 · p. 65

B.3 Equilibrium Before News

• Before news, equilibrium pricing is determined by market clearing around uncertain fundamentals and random payoff variation, with agents holding a correct prior and the asset in zero net supply.The model separates uncertainty about fundamentals from real payoff variation and adds noise traders only in the final period for tractability.Source: §B.3 · p. 66

B.4 Information Structure

• The model introduces LLM information capacity λ(c, k), increasing with model size k and accounting for news complexity c, with diminishing returns to size and greater benefits for complex news.It assumes λ(c, 0) = 0 and limk→∞λ(c, k) = 1, providing a microfoundation for heterogeneous LLM performance.Source: §B.4 · p. 68
• Agents process only part of the available information, while LLMs process signal precision initially and defer remaining precision to a second period when necessary.A sufficiently large LLM can exceed attentive agents’ information capacity.Source: §B.4 · p. 68

B.5 Information Update

• Bayesian updating gives each agent type a signal-conditioned fundamental-value expectation and posterior variance, determining its conditional payoff variance.Explicit formulas are provided in Online Appendix D.Source: §B.5 · p. 69

B.6 LLMs for Forecasting

• Without LLM trading, the model studies out-of-sample LLM reactions before ChatGPT-4’s release, then characterizes price dynamics and forecasting before allowing agents to use LLMs.Attentive agents trade on initial underreaction, while inattentive agents trade toward the final payoff without fully understanding the news impact.Source: §B.6 · p. 69

B.6.1 Second Period

• In the second period, attentive and inattentive agents share the same fundamental expectation, and equilibrium price equals consensus fundamentals plus noise-trading shocks.For attentive agents, future-price variance is amplified by fundamental risk and risk aversion but dampened by total trader volume.Source: §B.6.1 · pp. 69–70

B.6.2 Period One

• Higher short-selling costs reduce attentive participation after negative news, slowing price adjustment and generating stronger negative-news predictability than positive-news predictability.This mechanism produces Proposition B.2’s asymmetric predictability across news sentiment.Source: §B.6.2 · p. 72

Market Efficiency

• Mispricing decreases with attentive-agent participation, either agent type’s information capacity, and total trading volume, but increases with risk aversion and noise-trader risk.More attentive trading lowers mispricing.Source: Market Efficiency · p. 73
• The model defines greater market efficiency as less mispricing, while trader volume and lower trading risk reduce noise-trader price impact.Mispricing is not directly observable because the subjective information set is unavailable, so LLMs serve as an imperfect proxy.Source: Market Efficiency · p. 73

B.6.3 Predictability with LLMs

• LLM-strategy profits arise when the model’s valuation relative to the current price correctly signs the subsequent movement toward attentive agents’ valuation.The framework maps this profit condition to the empirical strategy of trading on LLM signals.Source: §B.6.3 · p. 74
• LLM-based strategies become profitable only when model size exceeds a threshold k*, with predictability larger for smaller or illiquid stocks and negative news.The threshold k* increases with information capacity, attentive-agent prevalence, and agent volume, but decreases with risk aversion and noise-trader variance.Source: §B.6.3 · p. 74

B.7.1 LLMs Better Than Inattentive Agents

• When LLMs outperform inattentive traders, broader adoption or higher model quality reduces market mispricing by accelerating news incorporation.If LLM capabilities approach attentive-agent capacity and all inattentive agents adopt them, initial-signal return predictability can vanish because of price impact.Source: §B.7.1 · p. 76

B.7.2 LLMs Better than Attentive Agents

• When LLMs outperform attentive agents, increasing model size improves overall market efficiency by enhancing already sophisticated traders’ information processing.Partial-adoption scenarios generally reinforce the conclusion that LLM adoption enhances market efficiency.Source: §B.7.2 · p. 77

C Prompts for Other LLMs

• Alternative-model prompts adapt to model capabilities: GPT-1 and GPT-2 use autocomplete sentiment, BERT completes a masked word, and BART performs zero-shot news classification.The resulting signals assign positive, negative, or neutral scores using model-specific output rules.Source: C Prompts for Other LLMs · p. 78

Online Appendix A: Sample Characteristics

• The sample-characteristics appendix compares firms in the news sample with the broader CRSP universe because news coverage is not random.These differences matter for interpreting the representativeness of the results.Source: Online Appendix A: Sample… · p. 80
• News-covered firms are larger and more liquid than other firms, but their returns are more volatile during news periods.Source: Online Appendix A: Sample… · p. 80

1. News Sources

• 153,892 unique news stories from 17 sources comprise the coverage, with the four largest sources accounting for 94.2% of stories and the top ten for 99.9%.Source: §1 · p. 80

2. Methodology

• The methodology compares all eligible U.S. common stocks in CRSP with the news-matched subset, using firm-day returns, market capitalization, and trading volume.The final news sample contains only share code 11 companies and summarizes each characteristic using means, standard deviations, quartiles, and medians.Source: §2 · pp. 80–81

3. Results

• The news sample’s median market capitalization is $1.99 billion versus $0.56 billion for CRSP, approximately 3.6 times larger, with differences across mean and 75th percentile.The comparison indicates that news coverage disproportionately focuses on large-cap stocks.Source: §3 · p. 82, Table OA2 · p. 83
• The news sample’s median trading volume is 827.47 thousand shares versus 265.32 thousand for CRSP, approximately 3.1 times higher across quantiles.This pattern suggests media preferentially cover more liquid stocks with greater trading activity.Source: §3 · p. 82, Table OA2 · p. 83
• Daily-return volatility is 14.12% in the news sample versus 5.70% for CRSP, approximately 2.5 times higher, while mean returns are 0.21% versus -0.03%.Both mean-return levels are close to zero, as expected for daily returns.Source: §3 · p. 82, Table OA2 · p. 83

4. Implications

• The sample is not representative of the average CRSP firm because news media naturally focus on companies of greater interest to investors and the public.Accordingly, the sample contains larger and more liquid stocks than the typical CRSP firm.Source: §4 · p. 82
• Higher news-sample volatility is partly mechanical because media coverage occurs during unusual developments that naturally coincide with larger price movements.The volatility difference therefore partly reflects when firms receive news coverage, not only persistent firm characteristics.Source: §4 · p. 82
• Although the sample is not representative of the average CRSP stock, its concentration in larger firms gives it substantial market coverage.The results apply to a significant portion of the equity market because the sample represents a large fraction of overall market capitalization and trading activity.Source: §4 · p. 82

Online Appendix B: Additional Figures and Tables

• The appendix also evaluates alternative prompting orders and intraday strategies, including entry windows from one minute after announcement through the next day’s close.The prompting table compares answering first with reasoning first, while the intraday figure reports cumulative returns for three entry-and-exit schedules.Source: Online Appendix B: Additional Figures… · pp. 86, 93, Table OA4 · p. 86, Figure OA1 · p. 93
• The news sample contains 81,700 overnight headlines from 2021–2024, with earnings-related news and insider-trading disclosures the two most prevalent event categories.BERTopic uses K-Means clustering with K=50 followed by hierarchical merging into 15 final categories, named by GPT-4o.Source: Online Appendix B: Additional Figures… · pp. 85, 88
• Small-cap short-only portfolios achieved a 3.17 Sharpe ratio, whereas large-cap short-only portfolios reached -0.61, explaining the gap between equal-weighted and value-weighted performance.Equal-weighted and value-weighted portfolios had Sharpe ratios of 2.97 and 0.56, respectively, because value weighting emphasizes large stocks where short strategies fail.Source: Online Appendix B: Additional Figures… · p. 92, Table OA10 · p. 92
• The overnight strategy averaged 186% daily one-way turnover, making its performance highly sensitive to transaction costs.The strategy held averages of 74 stocks in its long leg and 21 in its short leg, reflecting rapidly changing news availability.Source: Online Appendix B: Additional Figures… · pp. 95–96, Figure OA3 · p. 95, Figure OA4 · p. 96

Over Time

• GPT-4’s advantage over GPT-3.5 in predicting overnight drift increased from -0.4 basis points in 2021 to 8.9 basis points in 2023, arguing against lookahead bias.Its initial-reaction advantage remained between 5.1 and 10.2 basis points, while the drift advantage moderated to 3.7 basis points in 2024.Source: Over Time · pp. 97–98, Figure OA5 · p. 98, Figure OA6 · p. 99

Online Appendix D: Model Proofs

• The model-proofs appendix verifies theoretical sign inequalities, derivatives, and expectations using Mathematica 14.0.These proofs support the theoretical results developed in the appendix.Source: Online Appendix D: Model Proofs · p. 100

A Equilibrium Price in Period 2 (p2)

• In period 2, informed agents’ aggregate demand absorbs the noise-trading shock, with the negative sign indicating trading against the shock when it pushes prices in its direction.Attentive and inattentive agents share the same conditional expectation and variance in this period, and the equilibrium price follows from market clearing.Source: A Equilibrium Price in Period 2 (p2) · pp. 100–101

B Bayesian Information Update Formulas

• Bayesian updating assigns agent-specific effective signal precisions, including τA = γAτS, τI = ωτA, and τL = λ(c, k)τS, and decomposes payoff variance into real and fundamental uncertainty.The conditional payoff variance is σ^2ξ + σ^2µ,j|s.Source: B Bayesian Information Update Formulas · p. 101

C Attentive Agent’s Demand

• Substituting the specified coefficients a, b, and c yields the expression used to characterize attentive-agent demand.Source: C Attentive Agent’s Demand · p. 103
• A closed-form solution is obtained using Lemma C.1 for a normally distributed variable, under the condition 1 − 2cσ^2 > 0.The omitted proof follows by completing the square in the probability density function.Source: C Attentive Agent’s Demand · p. 103
• The demand function is finalized by applying the first-order condition and rewriting the resulting expression in demand form.Source: C Attentive Agent’s Demand · pp. 104–105

Comparative statics

• Mispricing decreases with the quantity and information capacity of attentive agents, and with the information capacity of inattentive agents.The attentive-agent derivatives are negative because their numerators are negative and denominators positive; the inattentive-agent derivative is also negative.Source: D Proposition B.1 - Mispricing · pp. 105–107
• Mispricing decreases with total trader volume but increases with risk aversion and noise-trader volatility.The risk-aversion derivative is positive, as is the derivative with respect to noise-trader volatility.Source: D Proposition B.1 - Mispricing · p. 107

E Theorem 1

• LLM strategy profits increase with model size because the profit derivative is positive.The derivative is positive because both its numerator and denominator are positive.Source: E Theorem 1 · p. 108
• LLM strategy profits are positive if and only if model size exceeds the threshold k∗.For fixed news complexity, λc(k) is strictly increasing and invertible, implying profitability precisely when k > k∗.Source: E Theorem 1 · pp. 108–109

E.1 Threshold Properties

• The profitability threshold rises with inattentive-agent capacity ω, attentive-agent capacity γA, attentive-agent proportion πA, and total trader volume V.These comparative statics are all positive.Source: §E.1 · p. 109
• The profitability threshold falls as risk aversion or noise-trader standard deviation increases.Both derivatives are negative.Source: §E.1 · p. 110

E.2 Proposition B.2

• When k > k∗, greater attentiveness lowers return predictability within the region of positive profitability.More costly attentive participation reduces the proportion of attentive agents.Source: §E.2 · pp. 110–111

E.3 Proposition B.3

• When profitability is positive, a higher attentive-agent proportion lowers both LLM profitability and mispricing.This extends the earlier comparative-static result for πA.Source: §E.3 · p. 111
• When k > k∗, higher trading volume, better inattentive-agent information processing, or lower noise risk reduces return predictability and mispricing.These effects are established by the corresponding derivatives.Source: §E.3 · p. 111

F Theorem 2

• Greater LLM adoption and larger model size each reduce equilibrium mispricing.The derivatives of mispricing with respect to inattentive-agent LLM adoption θ and model size k are both negative.Source: F Theorem 2 · p. 113

H Proof for Partial Adoption by Attentive Agents

• The partial-adoption extension models LLM adopters as receiving a private signal η that is processed faster and publicly revealed with a lag.The private signal raises total information precision from τd + τA to τd + τA + τη, representing superior processing capacity.Source: H Proof for Partial Adoption by… · p. 116

H.1.1 Primitive Assumptions

• The equilibrium assumes competitive agents, a single risky asset, public signal s, private signal η, staged trading, and inattentive agents who neither learn from prices nor speculate.LLM adopters’ private-signal precision is governed by λ(c, k), while all agents observe η after its Period 2 disclosure.Source: §H.1.1 · pp. 117–118
• Price learning requires both private information and noise: η adds non-common information, while u1 prevents prices from revealing η perfectly.Without private information, prices reveal no new fundamental information; without noise, non-adopters can deduce η exactly and lose adopters’ informational advantage.Source: §H.1.1 · p. 119

H.2.1 Bayesian Updating

• Bayesian updating gives LLM adopters, attentive non-adopters, and inattentive agents type-specific posterior beliefs based on their respective signal precisions.Posterior precision sums prior and signal precisions, while posterior variance includes fundamental uncertainty and payoff noise.Source: §H.2.1 · p. 120

H.2.2 Key Definitions for Equilibrium Analysis

• Equilibrium analysis defines attentive non-adopters’ post-learning precision as TA∗ ≡ τd + τA + τlearn, where τlearn measures information extracted from prices.The extension also introduces type-specific speculative precisions following the main model’s logic.Source: §H.2.2 · pp. 120–121
• By period 2, all agents trade on the unified information set (s, η), while LLM adopters’ remaining price uncertainty comes solely from future noise u2.Their speculative precision is therefore based on variance from u2 alone.Source: §H.2.2 · p. 121
• Non-adopting attentive agents observe (s, p1) but not η, so their period-2 uncertainty combines uncertainty about µL|s,η with future noise u2.The incremental uncertainty is defined as ∆ ≡ Var(µL|s,η|s,p1).Source: §H.2.2 · p. 121
• Effective trading precisions express optimal retrading demands under unbiased price expectations, with additive precision for LLM adopters and a parallel variance operator for non-adopters.The future-price-risk variance p,L is derived in Lemma H.4.Source: §H.2.2 · pp. 121–122
• Inattentive agents neither learn from prices nor engage in the corresponding retrading process under Assumption H.5.Their behavior is defined separately from the attentive-agent precision framework.Source: §H.2.2 · p. 122

H.3 The Case of Partial Adoption by Attentive Agents

• Learning from prices creates strategic complementarity: greater LLM adoption makes prices more informative about private signals, supporting stable partial adoption.Non-adopters can infer adopters’ private signal from the equilibrium price, making non-adoption potentially viable.Source: §H.3 · p. 122

Case 1: LLM Adopters

• LLM adopters observe (s, η) and obtain total information precision TL = τd + τA + τη, exceeding the precision available without the private signal.A fraction ϕ ∈ (0, 1] of attentive agents adopts the technology and receives the private signal η with precision τη > 0.Source: §H.3 · pp. 122–123

Case 2: Non-Adopting Attentive Agents

• Non-adopting attentive agents observe (s, p1), extract a noisy signal about adopters’ private signal, and update their beliefs about the fundamental.Their price-based learning precision satisfies τlearn ≤ τη, with equality if and only if there is no Period-1 noise trading.Source: §H.3 · pp. 123–124

H.3 The Case of Partial Adoption by Attentive Agents

• Market efficiency increases with both the fraction of LLM adopters, ϕ, and LLM signal quality, τη, as measured by forecast precision.The forecast variance is Var(µd | s, p1) = 1/TA∗, so efficiency increases if and only if TA∗ increases.Source: §H.3 · p. 123

Case 2: Non-Adopting Attentive Agents

• Non-adopters’ demand is solved by substituting the Period-2 market-clearing law into their Period-1 optimization, eliminating problematic consensus terms and yielding a closed-form solution.The derivation uses factorization because the random Period-2 consensus and noise are independent; when ∆ = 0, the effective variance is identical to that of LLM adopters.Source: Case 2: Non-Adopting Attentive Agents · pp. 126–128

General Result

• A linear REE exists and is unique for ϕ > 0, τη > 0, and finite σ2, because coefficient matching fixes χ/g and all equilibrium objects.The construction proceeds through adopter objects, learning precision, non-adopter objects, aggregate precision weights, and the price coefficients.Source: General Result · p. 130
• The key fixed point collapses to χ/g, which has a closed-form solution depending only on adopter-side primitives because non-adopter terms and aggregate precision cancel.LLM adopters create price informativeness by trading on η against noise, while non-adopters scale g and χ proportionally without changing their ratio.Source: General Result · pp. 131–132
• Increasing either the adoption rate ϕ or LLM quality τη makes prices more informative and improves market efficiency, with the quality effect eventually saturating under large payoff noise.Greater adoption increases adopter mass and lowers χ/g; higher τη also lowers χ/g, but χ/g approaches a positive constant as τη →∞ when irreducible payoff noise is large.Source: General Result · p. 133

Online Appendix E: Interpretability Methodology

• The interpretability analysis uses BERTopic to categorize headlines into coherent semantic topics without imposing predefined topic structures.The approach combines modern embeddings with clustering algorithms.Source: Online Appendix E: Interpretability… · p. 135

Sample and Data Preparation

• The sample contains 81,700 overnight headlines with directional GPT-4 predictions, after removing YES, NO, and UNKNOWN labels from explanations to prevent mechanical clustering.Neutral scores of 0 are excluded.Source: Sample and Data Preparation · p. 135

Technical Implementation

• The implementation embeds headlines in 384 dimensions, reduces them to 5 with UMAP, initializes 50 K-Means clusters, reduces topics hierarchically to 20, and labels them with GPT-4o.UMAP uses 15 neighbors and minimum distance 0; topic labels are generated from representative headlines and keywords while avoiding firm-specific and generic terms.Source: Technical Implementation · pp. 135–136
• The technical labeling procedure requires concise, specific, descriptive topic labels that avoid firm names, tickers, and generic financial terms.Labels should interpret the concise topic and focus on the specific theme represented by the explanations and keywords.Source: Technical Implementation · p. 136

Methodological Advantages

• Embedding-based K-Means clustering followed by hierarchical merging yields 20 interpretable topics from dense news headlines.The approach improves semantic similarity over bag-of-words, assigns every headline to a topic, balances granularity with interpretability, and uses GPT-4o for consistent labeling.Source: Methodological Advantages · p. 137
• The topic assignments decompose GPT-4’s performance and market reactions across 15 corporate-news categories.The underlying regressions omit firm and date fixed effects to preserve topic interpretability while using standard errors double-clustered by firm and date.Source: Methodological Advantages · p. 137

1. Methodology

• Drift returns use open-to-close trading for overnight news and close-to-close trading for intraday news.Overnight positions enter at the market open and exit at the close; intraday positions enter at the close and exit at the next day’s close.Source: §1 · p. 138
• 50% rebalancing produces approximately 100% gross exposure, versus 200% under full rebalancing.Partial rebalancing makes positions converge only partway toward changing targets, reducing turnover and portfolio exposure.Source: §1 · p. 138
• Transaction costs are modeled as fixed round-trip costs of 0, 5, 10, or 20 basis points.The costs represent conditions ranging from highly sophisticated investors to retail traders and capture commissions and price impact at opening and closing auctions.Source: §1 · pp. 138–139

2. Results

• Reducing rebalancing from 100% to 50% halves turnover from 190% to 94% and gross exposure from 196% to 101%, while leaving Sharpe ratios nearly unchanged.For overnight equal-weighted portfolios at zero costs, mean returns fall from 0.34% to 0.17% and volatility from 1.80% to 0.91%, while Sharpe ratios are 2.97 versus 2.95.Source: §2 · p. 139

3. Turnover and Transaction Costs

• At 20 basis points, overnight equal-weighted Sharpe falls from 2.97 to -0.39 under 100% rebalancing, while 25% rebalancing falls from 2.89 to -0.21.The 100% strategy’s mean daily return becomes -0.04%, whereas 25% rebalancing degrades more gradually and remains profitable through 10 basis points.Source: §3 · p. 140

4. Equal-Weighted vs Value-Weighted Portfolios

• Equal-weighted portfolios outperform value-weighted portfolios: overnight Sharpe is 2.97 versus an unspecified lower VW value, and intraday Sharpe is 2.63 versus an unspecified lower VW value.The results indicate that predictability is concentrated in smaller stocks, while VW overnight performance becomes unprofitable at 5 basis points with Sharpe -0.40 versus 2.13 for EW.Source: §4 · pp. 140–141

5. Implications for Tradability

• The 25% rebalancing strategy reduces overnight daily turnover by approximately 76%, from 190% to 46%, making the GPT-4 signal more practically tradable.The analysis emphasizes that exploiting predictability requires careful attention to implementation costs.Source: §5 · p. 141

5. Remains profitable at moderate cost levels (5-10 basis points)

• Predictability is concentrated where arbitrage is most costly: smaller stocks face higher transaction costs and less institutional attention, allowing mispricings to persist longer.This concentration creates practical concerns for large institutional investors but supports the limits-to-arbitrage mechanism.Source: §5 · p. 141, §4 · p. 141

6. Understanding the Size Effect: Reaction vs Drift

• Small-cap short-only strategies achieve a 3.17 Sharpe ratio versus -0.61 for large caps because negative-news drift continues in small and medium stocks but mean-reverts in large stocks.After negative signals, small caps move from an initial reaction of -2.33% to drift of -0.70%, whereas large caps move from -1.12% to +0.06%.Source: §6 · p. 142
• Positive signals generate positive drift across all size groups, with declining magnitude for larger stocks; long-only Sharpe ratios range from 0.30 to 1.09.Large-cap support mechanisms may cause recovery after negative news, whereas small caps lack such support and continue drifting downward as investors process information.Source: §6 · p. 142