PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Alejandro Lopez-Lira, Yuehua Tang

arXiv:2304.07619v6q-fin.STcs.CL

TL;DR

The paper asks whether general-purpose LLMs can predict stock-market reactions to news without explicit financial training, and whether their forecasts reveal information-processing frictions. It evaluates post-cutoff headlines with multiple models and develops a theoretical framework linking LLM capacity, underreaction, and arbitrage limits. GPT-4 predicts initial reaction directions with roughly 90% portfolio-day hit rates and predicts later drift, while predictability declines as LLM adoption rises.

  • Problem

    The paper examines whether LLMs can predict stock-market reactions from news without explicit financial training, addressing limited evidence on their financial decision-making capability.

  • Method

    The paper evaluates LLM assessments of post-knowledge-cutoff news headlines against initial reactions and subsequent drift, supported by a theoretical model of information processing and arbitrage limits.

  • Results

    GPT-4 achieves approximately 90% portfolio-day hit rates for initial reaction direction and significantly predicts subsequent drift, especially for smaller stocks and negative news.

  • Takeaways & Limitations

    Forecasting ability generally increases with model size, while strategy returns decline as LLM adoption rises, consistent with improved price efficiency.

  • Takeaways & Limitations

    Profitable exploitation is feasible mainly for participants with sufficiently low transaction costs, such as market makers.

Abstract

from arXiv · show

We document the capability of large language models (LLMs) like ChatGPT to predict stock market reactions from news headlines without direct financial training. Using post-knowledge-cutoff headlines, GPT-4 captures initial market responses, achieving approximately 90% portfolio-day hit rates for the non-tradable initial reaction. GPT-4 scores also significantly predict the subsequent drift, especially for small stocks and negative news. Forecasting ability generally increases with model size, suggesting that financial reasoning is an emerging capacity of complex LLMs. Strategy returns decline as LLM adoption rises, consistent with improved price efficiency. To rationalize these findings, we develop a theoretical model that incorporates LLM technology, information-processing capacity constraints, underreaction, and limits to arbitrage.

1 Introduction

The paper shows that GPT-4 can anticipate immediate stock-market reactions to news and identify subsequent drift, while using LLM assessments to study market information-processing frictions. Forecasting ability generally rises with model sophistication, whereas predictability declines as adoption increases.

  • GPT-4 and market reactions: Approximately 90% daily portfolio hit rates show GPT-4 correctly identifies the direction of initial price responses to news headlines.The reported hit rates include 93.3% for overnight headlines and 88.8% for intraday headlines.
  • GPT-4 and market reactions: GPT-4 assessments predict subsequent price drift over the next one to two trading days, especially for smaller stocks and negative news.The associated underreaction is strongest where arbitrage constraints are greater.
  • GPT-4 and market reactions: A GPT-4 long-short strategy based on overnight news earns 34 basis points per day before transaction costs.The strategy’s practical profitability is limited because sufficiently low transaction costs are required.
  • Model sophistication: Forecasting performance generally increases with model size, with GPT-4’s drift-strategy Sharpe ratio reaching 2.97 versus 1.66 for GPT-3.5.Basic models show limited initial-reaction capability and no significant subsequent-drift predictability.
  • Adoption and efficiency: The ChatGPT-based strategy’s annualized Sharpe ratio declines from 6.54 in 2021Q4 to 1.22 during January–May 2024 as model capabilities and adoption rise.The paper presents this pattern as suggestive evidence that broader LLM use improves market efficiency.
  • Information-processing frictions: Markets react efficiently to formulaic information but underreact to news requiring complex synthesis, including insider transactions and specialized conference presentations.GPT-4’s strongest advantages over GPT-3.5 occur in these complex-synthesis categories.

2 Institutional Background

This section introduces ChatGPT and the GPT architecture as general-purpose language models trained on large text corpora. It motivates comparing models of different complexity to assess their ability to predict stock returns from textual information.

  • LLM background: ChatGPT is a general-purpose language model based on the Generative Pre-trained Transformer architecture.GPT models use transformer-based mechanisms to process language sequences.
  • LLM background: GPT models are pretrained on large text corpora using unsupervised learning before being adapted to language tasks.The pretraining process develops representations of language syntax and semantics.
  • LLM background: Transformer self-attention helps models focus on relevant parts of long input sequences and generate coherent responses.The architecture supports tasks such as translation, summarization, question answering, and text completion.
  • Financial application: ChatGPT is not explicitly trained to predict stock returns, making its off-the-shelf financial assessment capability an empirical question.The paper evaluates ChatGPT alongside BERT, BART, FinBERT, and Llama2 models.
  • Financial application: Comparing models of different complexity allows the paper to examine how model complexity relates to stock-return prediction from textual data.The study examines twelve LLMs, with an overview provided in the online appendix.

3 Conceptual Framework: LLMs, Information Process-

The conceptual framework models LLMs as information processors operating alongside heterogeneous investors, market frictions, and limits to arbitrage. It explains why forecasting emerges, varies across settings, and may diminish as capable LLMs become widely adopted.

  • LLMs as information processors: The framework treats LLM capacity as depending on model sophistication and news complexity, with λ(c, k) representing interpretation precision.This structure accounts for performance differences across models and news types.
  • LLMs as information processors: Return predictability arises when an LLM processes news faster or more accurately than the marginal human trader setting the current price.Prices can subsequently drift toward a more fully processed information level.
  • Threshold and arbitrage: The model predicts a quality threshold k* below which LLM signals cannot overcome noise trader risk, transaction costs, and information-processing frictions.Only models with k > k* generate robust profitable predictions.
  • Threshold and arbitrage: Predictability is stronger in smaller, less liquid stocks and after negative news because information frictions and arbitrage constraints are greater.Shorting negative-news positions also faces higher costs and regulatory restrictions.
  • Adoption and efficiency: LLM adoption can improve price efficiency by helping inattentive agents and augmenting attentive agents’ information processing.Improved assessments pull equilibrium prices closer to fundamentals.
  • Adoption and efficiency: Widespread adoption can eliminate the LLM’s information advantage because shared signals become incorporated into prices.The paper describes this as the predictive signal containing the seeds of its own obsolescence.

4 Data

The empirical analysis combines firm-level news, stock returns, and intraday market data to evaluate LLM assessments of stock-market reactions. The final matched sample covers thousands of firms over an out-of-sample period after the models’ training cutoff.

  • Data sources and sample: The study combines CRSP returns and prices, news headlines, RavenPack data, and TAQ intraday trading data.TAQ data are cleaned to construct minute-by-minute volume-weighted average prices.
  • Sample design: The sample is out-of-sample because the ChatGPT models’ training data stop in September 2021.The analysis also addresses potential lookahead bias by comparing GPT-4 and GPT-3.5 performance over time.
  • News construction: News is matched with RavenPack identifiers to retain relevant firm-specific headlines and remove matching errors from web scraping.The sample requires highly relevant, complete articles or press releases and excludes stock-gain and stock-loss labels.
  • Data sources and sample: The final sample contains 159,137 firm-headline-date observations covering 4,123 companies from October 2021 through May 2024.These observations represent about 85% of the 4,875 firms in CRSP during the period.
  • News construction: About 82% of headlines are classified as overnight news and 18% as intraday news based on release timing.Overnight news is released before 9 a.m. or after 4 p.m. on a trading day.

5 ChatGPT Prompt

The paper uses reproducible, off-the-shelf prompts to ask ChatGPT whether public headlines are good, bad, or uncertain for a company’s short-term stock price. It converts these responses into GPT-4 scores and compares advanced and basic language models without explicit financial training.

  • Sample construction: The headline sample is cleaned by classifying overnight news at a 9 a.m. cutoff, removing duplicates and overly similar same-company headlines, and accounting for missing timestamps.Alternative morning cutoffs produce similar results; 139 of 159,137 observations lack timing information needed for overnight-versus-intraday classification.
  • Prompt design: The prompt asks ChatGPT to classify a headline’s short-term stock-price implication as “YES,” “NO,” or “UNKNOWN,” then provide one concise explanation.The company name and headline are substituted into the prompt, which treats the headline as the model’s only information source and permits an “UNKNOWN” response.
  • Reproducibility: Temperature is set to 0 so the model selects the highest-probability response and produces more reproducible outputs.The paper defines temperature as controlling response randomness and creativity.
  • Illustration: A Humana example shows GPT-4 interpreting a terminated acquisition pursuit as negative because the stock may lose an acquisition premium.A vendor sentiment score instead rates the same headline positively at 0.65, illustrating differing contextual interpretations.
  • Prompt design: The study applies the prompt to publicly available headlines through OpenAI’s API using the gpt-4-0314 snapshot, whose training data ends in September 2021.The cutoff is used to support an out-of-sample evaluation of later headlines.
  • Scoring: GPT-4 responses are mapped to a numerical score: “YES” equals 1, “UNKNOWN” equals 0, and “NO” equals -1.The sample’s average GPT-4 score is 0.31, while the median is zero.
  • Model comparison: The analysis also evaluates BERT, GPT-1, GPT-2, and Llama2 with a different strategy because basic models cannot follow instructions or answer specific questions.GPT-1 and GPT-2 are described as autocomplete models rather than instruction-following systems.
  • Robustness: A robustness test finds GPT-4 recommendations are very similar when the model reasons before answering rather than answers before reasoning.The comparison uses 1,000 randomly selected headlines.

6 ChatGPT and Market Information Processing

GPT-4 scores align strongly with firms’ immediate market reactions and also predict slower post-announcement drift. The drift is strongest for negative news and smaller stocks, but implementation costs materially constrain profitability.

  • Initial reaction: 1.27% versus –1.79% are the average initial reactions for overnight news with positive versus negative GPT-4 scores; intraday reactions are 1.33% versus –3.11%.Both overnight effects are highly statistically significant, supporting GPT-4’s ability to identify the primary economic impact of headlines.
  • Initial reaction: 93.3% and 88.8% are GPT-4’s long-short portfolio hit rates for initial reactions to overnight and intraday news, respectively.The corresponding mean returns are 3.06% for overnight news and 4.44% for intraday news.
  • Post-announcement drift: 58% and 55% are GPT-4’s drift-strategy hit rates for overnight and intraday news, with mean returns of 0.34% and 0.50%, respectively.Annualized pre-transaction-cost Sharpe ratios are 2.97 and 2.63.
  • Post-announcement drift: Negative news produces stronger subsequent drift predictability: overnight short-leg returns average 26 basis points with a 2.01 Sharpe ratio, versus 8 basis points and 0.78 for the long leg.The paper interprets this pattern as consistent with delayed incorporation and limits to arbitrage.
  • Trading implications: 190% versus 46% is daily turnover under 100% versus 25% rebalancing; at 10 basis points, Sharpe ratios are 1.29 and 1.34, respectively.The lower-rebalancing strategy is more robust to implementation costs despite similar zero-cost Sharpe ratios.
  • Trading implications: GPT-4 predicts initial reactions strongly but predicts no significant return potential during the first 1-to-15 minutes after intraday release.The score coefficient is insignificant for that interval, while later windows show predictive value.

Appendix A.23

Across model families, forecasting accuracy generally rises with model complexity, with GPT-4 performing best on initial reactions and subsequent drift. The largest gains occur for news requiring deeper analytical, quantitative, medical, or scientific interpretation.

  • Model complexity: GPT-4 achieves a 93.3% initial-reaction hit rate, compared with 87% for the next reported models, while its drift returns exceed those of GPT-3.5 and DistilBart-MNLI.GPT-3.5 and DistilBart-MNLI produce 0.29% and 0.14% average daily drift returns versus 0.34% for GPT-4.
  • Model complexity: 2.97 is GPT-4’s annualized drift-strategy Sharpe ratio, compared with 1.26 for DistilBart-MNLI and 1.66 for GPT-3.5.BART-large and Llama2-70b are weaker still, with Sharpe ratios of 1.05 and 0.97.
  • Model complexity: Basic models have initial-reaction hit rates below 65%, insignificant positive drift correlations, and negative drift-strategy Sharpe ratios.The paper attributes these results to insufficient information-processing capacity for correctly interpreting economic implications.
  • Model complexity: GPT-4 is the most accurate model considered, and the ability to assess news implications generally increases as model size grows.The paper describes anticipating market responses from text as an emerging capacity of more complex language models.
  • News heterogeneity: GPT-4’s initial-reaction hit rates remain 89–91% across simple and complex headlines and across journalist and firm sources.Its drift hit rates cluster between 55% and 59%, indicating more stable initial-reaction performance than drift performance.
  • News heterogeneity: GPT-4’s drift Sharpe ratio is 2.23 for news articles versus 1.96 for press releases, while drift hit rates are 59% versus 56%.The paper reports consistently high 91% initial-reaction hit rates for both source types.

8 Speed of Price Response and Market Efficiency

The evidence indicates that firm-specific news is absorbed into prices over roughly two days, while return predictability declines as LLM adoption grows. This pattern is consistent with improved price efficiency, although other market conditions may also contribute.

  • Speed of Price Response: 34 bps on the first trading day and 19 bps on the second are the average daily returns for the overnight-news strategy.The strategy does not show significant predictability afterward.
  • Speed of Price Response: 32 bps during the post-release session and 30 bps the next day are the significant intraday-news returns, with no significant returns afterward.The close-to-close intraday strategy is implementable but captures limited, noisier drift.
  • Speed of Price Response: About two days is the estimated window for firm-specific news to be absorbed into market prices.GPT-4 scores predict returns over the next two days, but not afterward.
  • LLM Adoption and Efficiency: 6.54 in 2021Q4 to 1.22 in January–May 2024 is the decline in the overnight-news strategy’s annualized Sharpe ratio.The intermediate values are 3.68 in 2022 and 2.33 in 2023.
  • LLM Adoption and Efficiency: The decline in strategy performance coincides with LLM proliferation and is consistent with reduced market underreaction and improved market efficiency.Changing market conditions could also contribute to the decline, so the evidence is suggestive rather than definitive.
  • Heterogeneity in Market Efficiency: Large-cap stocks show no drift after negative news, whereas small-cap stocks continue to drift despite GPT-4 assessments aligning with initial reactions across sizes.This pattern is presented as variation in market efficiency across market sizes.

Panel A: Overnight News

Panel A evaluates overnight-news portfolios formed from LLM scores, reporting initial-reaction and drift hit rates, long-short Sharpe ratios, returns, and model-size comparisons. GPT-4 performance is also compared with text-embedding approaches and examined across news complexity, source, and topics.

  • 1.326*** is the reported GPT-4 score coefficient in the overnight-news performance table.
  • GPT-4’s long-short portfolio achieves a 93.3% hit rate.The hit rate measures the percentage of days with positive portfolio returns.
  • The portfolios report initial-reaction and drift hit rates, long-short Sharpe ratios, daily returns, portfolio-leg sizes, and Fama–French five-factor statistics.Initial reactions cannot be traded without advance knowledge of news timing, whereas drift strategies are evaluated with annualized Sharpe ratios.
  • The analysis decomposes performance by headline complexity and news source, with categories determined at the individual-headline level before firm-day portfolio aggregation.High and low complexity are defined relative to the daily median Flesch-Kincaid readability score.
  • Overall, GPT-4 exceeds GPT-3.5 by 12.2*** basis points for initial reactions and 4.9*** basis points for drift, while Llama 70B exceeds Llama 7B by 58.6*** and 2.2*** basis points.The model-size comparison uses directional overnight headlines and reports differences by news topic.

B Theoretical Model

The theoretical model treats LLMs as information processors operating alongside attentive and inattentive investors, bounded processing, and limits to arbitrage. It explains when LLM signals generate predictability, why effects differ across markets and news sentiment, and how adoption can reduce mispricing.

  • The framework models LLMs as information processors with varying capabilities, extending noisy rational-expectations models to study return predictability and market efficiency.It explicitly introduces model size and information capacity as economic features of LLM technology.
  • Return predictability arises from temporary divergence between LLM assessments and slower-reacting human beliefs, moderated by attentive participation and arbitrage costs.The model includes attentive and inattentive agents with different information-processing abilities.
  • Mispricing decreases with attentive-agent share, either agent’s information capacity, and total trader volume, but increases with risk aversion and noise-trader risk.
  • A unique threshold k* exists such that, for fixed market parameters and news complexity, only LLMs with k > k* predict returns profitably on average.The threshold represents the minimum model sophistication needed to overcome market noise and information-processing frictions.
  • LLM-return predictability is larger in markets with fewer attentive participants, such as smaller or illiquid stocks, and with negative news.The model attributes the sentiment asymmetry to slower adjustment and higher barriers affecting short positions.
  • When capable LLMs are broadly adopted, improved information processing reduces mispricing and can substantially diminish, or eliminate, initial-signal predictability.The predicted efficiency gains arise when LLM capabilities approach those of attentive agents or exceed inattentive traders’ capabilities.

3. Results

The results sample is concentrated in larger, more liquid, and more volatile firms than the broader CRSP universe. The authors note that it represents a substantial share of market capitalization and trading activity but is not representative of the average stock.

  • The median market capitalization is $1.99 billion in the news sample versus $0.56 billion in CRSP, approximately 3.6 times larger.The concentration is also visible in the mean and 75th-percentile market capitalization.
  • The median trading volume is 827.47 thousand shares in the news sample versus 265.32 thousand in CRSP, approximately 3.1 times higher.The difference persists across quantiles, indicating preferential coverage of more liquid stocks.
  • Daily-return standard deviation is 14.12% for the news sample versus 5.70% for CRSP, approximately 2.5 times higher.Mean daily returns are 0.21% versus -0.03%, respectively, and both are close to zero.
  • The authors attribute higher sample volatility partly to news events coinciding with unusual developments and larger price movements.
  • The sample is not representative of the average CRSP stock, although it covers a large fraction of overall market capitalization and trading activity.

Online Appendix B: Additional Figures and Tables

The appendix reports additional data descriptions, prompting analyses, portfolio results, topic categories, transaction-cost analyses, size comparisons, and robustness checks. It also examines GPT-4’s changing advantage over GPT-3.5 and addresses lookahead bias.

  • News data and topics: The news dataset is summarized by year, timing, news type, covered firms, trading days, and coverage intensity.Timing distinguishes overnight news from intraday news, while news type distinguishes press releases from full articles.
  • Prompting analysis: Prompting ChatGPT 4 to answer first or reason first produces a distribution of predictive recommendations across 1,000 randomly selected headlines.The analysis compares the standard answer-first prompt with a reason-first ordering.
  • News data and topics: 81,700 overnight news headlines are organized into 15 topic categories using K-Means clustering followed by hierarchical merging.GPT-4o-generated labels and representative keywords describe the final categories.
  • Portfolio performance and costs: Small-stock short-only portfolios achieve a 3.17 Sharpe ratio, whereas large-stock short-only portfolios reach -0.61.This asymmetry helps explain the difference between equal-weighted and value-weighted performance.
  • Portfolio performance and costs: 186% average daily one-way turnover makes the overnight strategy highly sensitive to transaction costs.The rapid entry and exit of stocks reflects the predominance of single-day news events.
  • Robustness over time: GPT-4’s initial-reaction advantage remains between 5.1 and 10.2 basis points across years, while its overnight-drift advantage rises from -0.4 basis points in 2021 to 8.9 in 2023.The overnight-drift advantage moderates to 3.7 basis points in 2024.
  • Robustness over time: The increasing overnight-drift advantage over time is presented as evidence against lookahead bias from memorized information.The reported increase from -0.4 basis points in 2021 to 8.9 basis points in 2023 has a t-statistic of 3.35.

D Proposition B.1 - Mispricing

Proposition B.1 characterizes how mispricing responds to market risk, trader composition, information-processing capacity, and trading volume. The comparative statics show that greater risk and noise increase mispricing, while broader or better processing reduces it.

  • Comparative statics: Mispricing decreases with the quantity of attentive agents and with attentive or inattentive agents’ information-processing capacity.The corresponding derivatives are negative.
  • Comparative statics: Mispricing also decreases as total trader volume increases.The derivative with respect to total volume V is negative.
  • Comparative statics: Mispricing increases with risk aversion and noise-trader volatility.The derivatives with respect to risk aversion and noise-trader volatility are positive.

E Theorem 1

Theorem 1 analyzes LLM strategy profits and their profitability threshold as model quality and market conditions vary. It predicts that profits rise with model size but that market sophistication and lower noise reduce return predictability.

  • Profitability: LLM strategy profits increase with model size because the model-size derivative is positive.The proof states that both the derivative’s numerator and denominator are positive.
  • Profitability: LLM strategy profits are positive if and only if model size exceeds a threshold k > k∗.The threshold is defined through the increasing LLM signal-quality function.
  • Threshold properties: The profitability threshold rises with inattentive-agent capacity, attentive-agent capacity, the proportion of attentive agents, and total trader volume.The corresponding threshold derivatives are positive.
  • Threshold properties: The profitability threshold falls with risk aversion and noise-trader standard deviation.The derivatives with respect to these two variables are negative.
  • Predictability: When profitability is positive, more attentive agents, higher trading volume, better inattentive processing, or lower noise risk reduces return predictability and mispricing.These comparative statics are stated for the region k > k∗.
  • Predictability: Markets with higher costs of being attentive have a lower proportion of attentive agents.This links attention costs to the equilibrium composition of market participants.

F Theorem 2

Theorem 2 extends the model to LLM adopters who obtain a private signal earlier than other agents and later reveal it through prices. The framework uses heterogeneous information-processing capacities and noise trading to study price-based learning and persistent predictability.

  • Private information and price learning: LLM adopters receive a private signal η that represents information processed more quickly and revealed publicly with a lag.The private signal gives adopters higher total information precision than attentive agents using only the public signal.
  • LLM technology: LLM technology quality k increases the private-signal precision λ(c, k), with λ(c, 0) = 0 and limk→∞λ(c, k) = 1.The model restricts the private signal to be no more precise than the raw public signal.
  • Agent behavior and timing: The equilibrium has three trading periods: partial public-signal processing, public revelation of the private signal, and final asset payoff.LLM adopters trade using the private signal before it becomes publicly available.
  • Agent behavior and timing: The model assumes inattentive agents process only the public signal, do not learn from the equilibrium price, and do not engage in speculative trading.These assumptions represent limited information-processing and strategic sophistication.
  • Private information and price learning: The price-learning extension requires private information and noise trading so prices aggregate information without fully revealing the private signal.Without private information, prices add no information beyond the public signal; without noise, non-adopters can infer η exactly.
  • LLM technology: The extension formalizes heterogeneous posterior beliefs using Bayesian updating from public and private signals.Posterior precision is the sum of prior and signal precisions.

H.3 The Case of Partial Adoption by Attentive Agents

The model studies a market in which some attentive agents adopt LLM technology and others infer information from prices. Partial adoption can form a stable equilibrium because prices reveal adopters’ private signals.

  • Partial adoption: The model permits partial adoption equilibria in which some sophisticated investors rely on price discovery rather than directly acquiring the technology.As more agents adopt LLMs, prices become more informative about private signals, making non-adoption potentially viable.
  • Partial adoption: A fraction ϕ of attentive agents adopts LLM technology and receives a private signal η with precision τη > 0.Non-adopting attentive agents may infer information about the fundamental from the equilibrium price.
  • Equilibrium structure: In the linear equilibrium, adopters’ private signal enters the price alongside the public signal and noise.The price function is p1 = a + bs + gη + χu1, with coefficients determined by market clearing.
  • Learning from prices: Non-adopting attentive agents observe the public signal and price, then construct an effective signal about the fundamental.This learned signal is a noisy measure of adopters’ private signal and adds information to their posterior.
  • Learning from prices: τlearn ≤ τη, with equality if and only if there is no noise trading in Period 1.Noise trading reduces the precision with which non-adopters can learn from prices.
  • Trading decisions: Agents’ optimal demands account for heterogeneous information sets and the possibility of retrading before the asset pays its dividend.The derivation uses backward induction under CARA-normal preferences.

General Result

The general result characterizes a unique linear rational-expectations equilibrium using a closed-form noise-to-signal ratio. Greater LLM adoption or information quality increases market efficiency and can support partial adoption.

  • Existence and uniqueness: For ϕ > 0, τη > 0, and finite σ2, a linear rational-expectations equilibrium exists and is unique.Uniqueness follows because coefficient matching forces the same χ/g and therefore the same equilibrium coefficients.
  • Closed-form solution: The key fixed point collapses to the single ratio χ/g, which is solved in closed form and depends only on adopter-side objects and primitives.Non-adopter terms and aggregate precision weights cancel from the ratio.
  • Closed-form solution: The equilibrium can be solved sequentially from adopter objects through learning precision, non-adopter objects, aggregate weights, and price coefficients.This procedure avoids iterative fixed-point computation.
  • Efficiency and comparative statics: Market efficiency increases with both the fraction of LLM adopters ϕ and the quality of their information τη.Efficiency is measured through forecast variance, with efficiency increasing if and only if TA∗ increases.
  • Efficiency and comparative statics: As ϕ rises, more adopters trade on η and reduce χ/g, making prices more informative about the private signal.Higher τη also lowers χ/g, although the effect saturates as τη →∞ when payoff noise remains irreducible.
  • Efficiency and comparative statics: Strategic complementarity can support partial adoption because improved price informativeness benefits non-adopters through information aggregation.Large irreducible payoff noise limits how aggressively adopters trade and therefore limits price informativeness.

2. Results

The results show that GPT-4-based predictability is economically meaningful but highly sensitive to implementation choices. Predictability is concentrated in smaller stocks, especially through continued negative-news drift.

  • Rebalancing and costs: Reducing rebalancing from 100% to 50% cuts turnover from 190% to 94% and gross exposure from 196% to 101%.Partial rebalancing reduces signal exposure while maintaining similar risk-adjusted returns per unit of exposure.
  • Rebalancing and costs: Reducing rebalancing from 100% to 50% cuts mean daily returns from 0.34% to 0.17% and volatility from 1.80% to 0.91%.The Sharpe ratio remains nearly unchanged: 2.97 for 100% rebalancing versus 2.95 for 50%.
  • Rebalancing and costs: At 20 basis points of transaction costs, the overnight 100% rebalancing strategy becomes unprofitable, with returns of -0.04%.The 25% rebalancing strategy degrades more gradually and remains positive through 10 basis points.
  • Portfolio weighting and size: Equal-weighted portfolios outperform value-weighted portfolios, with overnight Sharpe ratios of 2.97 versus the lower VW result at zero costs.The evidence indicates that predictability is concentrated in smaller stocks and may be difficult for large institutions to exploit.
  • Rebalancing and costs: The 25% rebalancing strategy reduces daily turnover by approximately 76%, from 190% to 46% for overnight news.This makes it a more realistic implementation alternative than full rebalancing.
  • News, drift, and size: After negative news, small caps fall 2.33% initially and drift another -0.70%, whereas large caps fall 1.12% initially and drift +0.06%.The short side therefore works for small and medium caps but not large caps, reflecting continuation versus mean reversion.
  • News, drift, and size: Short-only strategies have Sharpe ratios of 3.17 for small caps versus -0.61 for large caps.Positive signals show positive drift across all size groups, though the magnitude declines for larger stocks.
  • News, drift, and size: The size pattern is consistent with smaller stocks lacking support mechanisms that allow large-cap prices to partially recover after negative news.Uninformed investors gradually process the information, allowing negative drift to continue in small caps.
Loading 2304.07619v6…