TL;DR
Better AI helps individual traders profit by extracting subtle patterns from public data, but this creates a surprising side effect: it can make market prices less revealing about new, private information. As AI adoption grows, the market becomes better at sharing sophisticated analysis with everyone but worse at discovering new facts, ultimately reducing the private incentive to invest in even more powerful computation.
What Problem Does This Paper Address?
Economists have long viewed market prices as "information aggregators"—tools that collect scattered knowledge from thousands of individuals and boil it down into a single number. The assumption has been that as trading technology improves, prices should become more "efficient" and informative.
However, this paper identifies a critical distinction: information can come from discovery (new facts not yet in the record) or interpretation (better analysis of what is already public). While AI is incredibly good at interpretation, the author argues that this very strength might crowd out discovery. If an AI can already guess a company's success by analyzing its public filings and past stock patterns, a trader who has a "private clue" (like visiting a supplier) has less incentive to trade aggressively on that clue. Consequently, that unique private information never makes it into the price.
Methods
The researcher uses a multi-pronged approach to prove this "information substitution" effect:
- Theoretical Modeling: A competitive market model where traders have "resource-constrained" forecasting abilities. It uses a CARA (Constant Absolute Risk Aversion) utility framework to determine how traders choose positions based on their AI's forecasts.
- Neural Market Experiments: The author built a simulated market where agents are powered by Gated Recurrent Units (GRUs) and State-Space Models (SSMs). These agents "learn" to trade by observing price histories and private signals.
- Empirical Analysis: Using U.S. stock market data to compare how "limited" models (Ridge regression) and "richer" models (Gradient Boosting) improve when given current market snapshots.
- LLM Benchmarking: A novel experiment using GPT-5.2 and GPT-6.1 Sol to see how language models change their earnings forecasts when given a "private" research note versus a long history of public stock returns.
Datasets
- Synthetic Data: Large-scale simulations of an eight-factor dynamic economy used for the neural market experiments.
- U.S. Market Data: A massive dataset of 58,807 firm-quarter observations (2018–2025) combining stock prices from CRSP, accounting data from Compustat, and analyst expectations from I/B/E/S.
- LLM Testbed: 50 selected historical U.S. firm-quarter cases used to prompt language models with varying levels of financial history.
Key Insights and Ideas
- The Substitution Effect: AI treats public data analysis as a substitute for private observation. When the AI's "public" forecast is highly accurate, the "private" clue adds less marginal value, so the trader reacts to it less.
- More Sharing, Less Discovery: As AI adoption increases, prices become more useful to "limited observers" (people without fancy AI) because the AI's sophisticated analysis is now "baked into" the price. However, the price becomes less informative about newly discovered facts.
- The "Compute" Trap: In a world of high AI adoption, the private benefit of upgrading to a faster, bigger model actually falls. If your competitors' AIs have already put the analysis into the price, there’s less profit left for you to find by doing the same math.
- Retraining Feedback: When traders who didn't upgrade retrain their old models on a market now dominated by new AI, they inadvertently amplify the loss of private information by changing how they interpret the new price signals.
Key Results
- Discovery Decline: In the population model, the precision of new information in prices fell by about 16% after advanced AI was deployed and models were retrained.
- Upgrade Value: In neural market experiments, the benefit of upgrading from a small to a large AI model dropped by 44.2% as the larger model became widespread.
- Predictive Value: For a "limited" observer, the value of seeing the market price increased by 12.4% as AI adoption rose, while it decreased by 6.8% for an advanced observer.
- LLM Response: GPT-6.1 Sol (the more advanced model) responded 6% less to a private research clue than GPT-5.2 when historical market returns were available, confirming that better "public" analysis reduces the weight given to "private" signals.
Limitations and Future Work
The author acknowledges several boundaries to this study:
- Holding Horizons: The model primarily looks at one-period claims; the dynamics might shift for long-lived assets where resale prices matter more.
- Gaussian Assumptions: Much of the mathematical proof relies on Gaussian (normal) distributions, though the neural and empirical tests suggest the findings are robust even when those assumptions are relaxed.
- Compute Costs: The study evaluates the gross benefit of AI but does not factor in the literal cost of electricity and hardware, which would further decrease the net incentive to upgrade.
- Future Direction: A key area for future research is whether these effects lead to market instability or "flash" events, as the "sharing" of information might lead to more correlated trading behaviors.
Summary updated