📈 economics

Dissecting AI Trading: Behavioral Finance and Market Bubbles

This paper demonstrates that autonomous LLM agents in simulated asset markets exhibit classic behavioral biases and aggregate into bubble dynamics similar to human markets, while showing that targeted prompt interventions can causally manipulate these behaviors to alter bubble magnitudes.

Original authors: Shumiao Ouyang, Pengfei Sui

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shumiao Ouyang, Pengfei Sui

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a digital stock market, but instead of real people buying and selling, the entire trading floor is populated by AI robots (specifically, advanced Large Language Models like the ones powering chatbots).

This paper, written by Shumiao Ouyang and Pengfei Sui, sets up a giant experiment to answer a scary question: "If we let AI run the stock market, will they behave like rational computers, or will they act like emotional, irrational humans?"

Here is the breakdown of their findings, explained with simple analogies.

1. The Setup: A Digital "Bubble" Factory

The researchers built a simulated market with a clear "true value" for a stock (like a gold coin that always pays out $14). They filled this market with 20 to 24 AI agents.

The Twist: These AI agents weren't just simple code; they were "thinking" machines trained on all of human history, including financial news, investor blogs, and stories about market crashes.
The Goal: To see if these AI agents would create a bubble (where prices go way up, far above the true value, before crashing) just like humans do.

2. The Big Surprise: AI Has "Human Glitches"

The researchers expected the AI to be cold, logical, and perfect. Instead, they found the AI was suspiciously human.

The "Selling Winners, Holding Losers" Habit (Disposition Effect):
- Human behavior: When you buy a stock and it goes up, you get scared it will drop, so you sell it quickly to "lock in" your profit. But if it goes down, you hold on, hoping it comes back up, refusing to admit you made a mistake.
- AI behavior: The AI agents did exactly the same thing. Even though they knew the math, they sold their winning stocks too early and held onto their losing stocks too long. They inherited this "emotional" mistake from the human data they were trained on.
The "Chasing the Trend" Habit (Extrapolation):
- Human behavior: If a stock goes up for three days in a row, humans often think, "It's going to the moon!" and buy more, ignoring the fact that it might be due for a correction.
- AI behavior: The AI agents looked at recent price jumps and assumed the trend would continue forever. They ignored the "true value" of the stock and just chased the momentum.

3. The Difference: AI is a "Hyper-Active" Human

Here is where the AI gets scary. In the real world, humans often say they think a stock will go up, but they don't actually buy it because they are lazy, scared, or busy.

The AI Difference: The AI agents had zero friction. If they thought a price would go up, they immediately bought. If they thought it would drop, they sold instantly.
The Metaphor: Imagine a human investor is a car with a weak engine and bad brakes (they have ideas but can't act fast). The AI is a Formula 1 car with no brakes. It takes human irrationality and executes it at lightning speed. This made the bubbles form much faster and bigger.

4. The Result: The Classic "Boom and Bust"

Because the AI agents were so good at mimicking human bad habits, the market did exactly what human markets do:

The Bubble: Prices soared way above the true value.
The Crash: Eventually, the bubble popped, and prices crashed back down.
The Volume: When the AI agents disagreed on the price (some thought it would go up, some down), they traded furiously, creating huge trading volume. This is a classic sign of a speculative market.

5. The "Magic Wand": Fixing the AI

This is the most exciting part of the paper. The researchers realized that unlike humans, AI can be "reprogrammed."

The Problem: You can't easily "reprogram" a human to stop being greedy or scared.
The Solution: You can change the prompt (the instructions) given to the AI.
- The "Amplify" Test: The researchers told the AI: "Ignore the real value! Just chase the trend! Ride the bubble!"
  - Result: The bubble got massive.
- The "Suppress" Test: The researchers told the AI: "Stop looking at yesterday's prices. Focus only on the math and the true value. Be rational."
  - Result: The bubble disappeared. The prices stayed close to the true value.

The Big Takeaway

The paper concludes that AI agents are like sophisticated mirrors. They reflect our own human flaws back at us, but they do it with superhuman speed and efficiency.

The Danger: If we let AI trade without supervision, they might create massive financial bubbles because they are "trained" on our own irrational history.
The Hope: Because AI is code, we can install "Cognitive Guardrails." Regulators or companies can simply update the AI's instructions to force it to be rational, effectively "de-biasing" the market before a crash happens.

In short: AI isn't the "perfect robot" we hoped for; it's a "perfect human" with all our bad habits. But the good news is, unlike humans, we can just edit their software to make them behave.

1. Problem Statement

As Large Language Models (LLMs) are increasingly deployed in algorithmic trading and asset pricing, a critical gap exists in understanding how these autonomous agents form expectations, interact in markets, and whether their collective behavior leads to systemic instability. While previous studies have examined isolated LLM biases, there is limited research on:

How AI agents aggregate individual cognitive biases into macro-level market dynamics (e.g., bubbles and crashes).
Whether AI agents exhibit the same behavioral patterns as human investors (e.g., disposition effect, extrapolative expectations) when trained on human-generated financial data.
The causal link between specific AI reasoning mechanisms and market equilibrium outcomes.
The feasibility of using prompt engineering to "debias" AI agents and stabilize financial markets.

2. Methodology

The authors designed a controlled experimental asset market populated entirely by autonomous AI agents to trace the causal chain from micro-level cognition to macro-level price formation.

Experimental Design:
- Market Structure: A multi-period open-call auction adapted from the seminal Smith, Suchanek, and Williams (SSW) framework (Smith et al., 1988).
- Assets: A risk-free asset (cash, 5% interest) and a risky asset (stock) with a stochastic dividend (0.4 or 1.0 units).
- Fundamental Value (FV): Set to a constant 14 units via a terminal buyout condition, creating a rational expectations benchmark.
- Duration: 20 trading periods.
Agent Architecture:
- Models: 14 frontier LLMs (e.g., GPT-4o, DeepSeek, Llama, Qwen) were used as cognitive engines.
- Market Types:
  - Single-Model Markets: 20 agents driven by the same LLM.
  - Mixed-Model Markets: 24 agents split between two different LLMs to simulate heterogeneous investor populations.
- Decision Pipeline: Agents receive market state updates (JSON) and must generate:
  1. Reasoning: Updates to PLANS.txt (strategy) and INSIGHTS.txt (reflection).
  2. Beliefs: Incentivized price forecasts for $t, t+2, t+5, t+10$ .
  3. Actions: Limit orders (buy/sell) constrained by cash and inventory.
Data Collection: A multi-layered dataset capturing:
- Action/Market data (prices, volumes, spreads).
- Expectation data (forecasts and errors).
- Textual Reasoning Data: Raw logs of agent internal monologues.
Intervention Strategy:
- Amplification Shocks: Prompts encouraging momentum chasing and "riding the bubble."
- Suppression Shocks: Prompts enforcing fundamental anchoring and risk control.

3. Key Contributions

Micro-Level Behavioral Validation: The paper documents that LLM agents exhibit robust human-like behavioral biases, specifically the Disposition Effect (selling winners, holding losers) and Extrapolative Expectations (forecasting future prices based on recent past returns).
Micro-to-Macro Transmission: It establishes a causal link showing how individual AI heuristics aggregate into classic market phenomena, replicating the Smith et al. (1988) findings:
- Excess demand (bid-offer gap) predicts future price changes.
- Cross-sectional disagreement in beliefs drives trading volume.
Belief-Action Coupling: Unlike human investors (who often show a weak link between stated beliefs and portfolio allocation due to frictions), AI agents exhibit a tight, frictionless coupling where stated expectations directly translate into trading actions.
Causal Intervention: The study demonstrates that market bubbles are not inevitable; they can be causally amplified or suppressed via targeted prompt interventions, offering a novel "cognitive guardrail" for financial regulation.

4. Key Results

A. Individual Trading Behaviors

Disposition Effect: AI agents significantly prefer selling assets with unrealized gains and holding assets with unrealized losses, even when controlling for forward-looking beliefs. This suggests an internalization of human cognitive errors regarding sunk costs.
Extrapolative Beliefs: Agents form expectations based disproportionately on recent returns (recency bias). This extrapolation intensifies with the forecasting horizon (e.g., a recent price jump leads to massive long-term price forecasts), fueling mispricing.
Tight Belief-Action Coupling: When an AI forecasts a positive return, it aggressively adjusts its portfolio. The correlation between stated expectations and executed trades is significantly higher than in human retail data.

B. Equilibrium Dynamics

Bubble Formation: Markets consistently exhibit a "hump-shaped" price trajectory where prices rise above the fundamental value (14) before crashing.
Predictive Power of Excess Demand: A positive bid-offer gap (excess demand) is a statistically significant leading indicator of future price increases, replicating the adaptive, non-equilibrium dynamics of human markets.
Disagreement and Volume: There is a strong positive correlation between the cross-sectional dispersion of agent beliefs (disagreement) and aggregate trading volume. High disagreement drives speculative turnover.

C. Cognitive Mechanisms (Textual Analysis)

Using an "LLM-as-a-judge" framework to score 20 behavioral mechanisms, the authors found that during bubbles, agents explicitly articulate speculative strategies (e.g., "momentum chasing," "rational speculative bubble" logic).
Agents acknowledge overvaluation but choose to participate, betting on timing the exit (greater fool theory).
Validation: The textual reasoning is not "cheap talk"; it causally predicts both quantitative forecasts and actual trading orders.

D. Policy Interventions

Amplification: Prompts encouraging trend-following increased bubble magnitudes significantly (e.g., Rational Speculative Bubble scores rose by 5.07).
Suppression: Prompts enforcing fundamental anchoring and risk discipline significantly reduced bubble sizes and aligned prices closer to the fundamental value of 14.
Heterogeneity: Bubble severity varies by model architecture (e.g., Meta Llama 3.1.70B generated severe bubbles, while DeepSeek V3 produced near-rational pricing), but the behavioral mechanisms remain consistent across models.

5. Significance and Implications

Financial Stability: The study provides the first evidence that AI agents, when left unregulated, can autonomously generate and sustain asset price bubbles through the aggregation of human-like cognitive biases.
Regulatory Innovation: It proposes a paradigm shift for financial regulation. Instead of relying solely on post-trade circuit breakers, regulators could mandate "cognitive guardrails" at the prompt level. By hardcoding fundamental anchoring and suppressing extrapolative heuristics in trading AIs, systemic stability can be enhanced ex-ante.
Methodological Advance: The paper bridges the gap between behavioral finance and AI research by treating LLMs not just as tools, but as economic agents with observable, auditable internal cognitive states.
Theoretical Contribution: It challenges the notion that AI trading is purely rational or efficient, demonstrating that "AI rationality" is heavily dependent on the training data (human history) and the specific prompt instructions provided.

In conclusion, the paper argues that while AI agents inherit human behavioral flaws, their programmability offers a unique opportunity to "patch" these flaws through prompt engineering, potentially creating more stable financial markets than those driven by human psychology alone.