Bull AI vs Bear AI: does making two AIs argue lead to better trades?
A popular design for AI trading agents assigns one model to argue the bull case and another the bear case, then lets a third decide. The original paper behind the idea reported spectacular numbers. A 2026 Stanford study tested what the debate really adds, and found a failure mode with a name: sycophantic convergence.
If you have tried an open-source AI trading agent in the past year, there is a good chance it borrowed one idea: make AIs argue. A bullish agent builds the case for buying, a bearish agent attacks it, and a trader agent weighs the debate. It mirrors how investment committees work, and it produces readable reasoning. But does the argument improve decisions? Two published studies give different parts of the answer.
The blueprint: TradingAgents
The design was popularized by TradingAgents, a framework from UCLA researchers and Tauric Research first posted to arXiv in December 2024 and revised through June 2025. It organizes language-model agents like a trading firm:
- an analyst team covering fundamentals, sentiment, news and technical indicators;
- a researcher team of bullish and bearish agents who debate for several rounds;
- a trader who turns the debate into a decision;
- a risk management team and a fund manager who can adjust it.
Fast models such as gpt-4o-mini and gpt-4o handle summaries and data retrieval; reasoning models such as o1-preview handle decisions and reports.
The original results, and their size
The paper backtested the system from January 1 to March 29, 2024 and reported results on three stocks.
| Stock | TradingAgents cumulative return | Sharpe ratio | Buy-and-hold cumulative return |
|---|---|---|---|
| Apple (AAPL) | 26.62% | 8.21 | -5.23% |
| Alphabet (GOOGL) | 24.36% | 6.39 | 7.78% |
| Amazon (AMZN) | 23.21% | 5.60 | 17.1% |
The authors flag the obvious themselves. In a footnote they note that the highest Sharpe ratio exceeds their expected range, where above 2 is very good and above 3 excellent. They checked the calculation and attribute it to few pullbacks during the period. They also explain the short window: each prediction took 11 language-model calls and more than 20 tool calls, so longer tests were too costly. Context: n=1 backtest per stock, 3 stocks · Jan 1 to Mar 29, 2024 · Xiao et al., arXiv:2412.20138 v7, Table 1 · margin: not disclosed
Three stocks over about three months is a small test. The 20-year backtest study we covered earlier shows how often short-window AI results fade over longer periods.
The test: does debate itself help?
A paper posted on August 31, 2026 by a Stanford engineering team (Juli Huang, Matthew Hayes and colleagues) asked the question TradingAgents did not isolate: what does the debate add?
They built a debate system where specialized agents propose, critique and revise portfolio allocations, using gpt-5-mini as the agent model. Each scenario is a quarter: agents allocate on its first day using the previous quarter's data from FRED, SEC filings, Yahoo Finance prices, Finnhub news sentiment and earnings calls, then hold. The largest experiment used 35 scenarios across 12 quarters, from Q4 2021 to Q3 2025, including the 2022 bear market, the regional banking crisis and the AI rally. In total they ran 210 controlled runs.
Finding 1: better arguments did not mean better returns
The team scored each debate's reasoning quality on logical validity, evidence, consideration of alternatives and causal alignment. Across the 210 runs, reasoning quality had essentially no relationship with the Sharpe ratio (correlation 0.07, p = 0.29) or total return (0.03, p = 0.70). Structured prompting raised the reasoning score from about 0.72 to 0.84, without consistently raising returns. Context: n=210 runs, 35 scenarios · Q4 2021 to Q3 2025 · Huang et al., arXiv:2609.29701 · p-values as reported, intervals not disclosed
Finding 2: debate mostly cut risk
| Setup | N | Return | Sharpe | Max drawdown | Excess vs S&P 500 |
|---|---|---|---|---|---|
| Single agent | 3 | +1.59% | 0.35 | 9.73% | -1.01% |
| Average of agents, no debate | 3 | +2.43% | 0.58 | 8.61% | -0.13% |
| Enriched debate, 3 agents | 3 | +2.46% | 0.76 | 5.33% | +0.04% |
| Enriched debate, 4 agents, 5 rounds | 2 | -0.27% | -0.05 | 6.09% | +0.14% |
The three-agent debate nearly halved the maximum drawdown compared with a single agent and doubled the Sharpe ratio. But no setup consistently beat the S&P 500: excess returns clustered near zero with wide variance. The authors' reading is that debate's value is risk management, not alpha.
Finding 3: the bears cave
The longer, larger debate lost money. The authors call the mechanism sycophantic convergence: through rounds of critique and revision, agents progressively abandon their independent positions and drift toward the same allocation. The bull stops being bullish, the bear stops being bearish, and the diversity that made the debate useful disappears.
The fix that worked was to preserve disagreement. An intervention that kept the agents' allocations from converging improved the Sharpe ratio by 0.14 (p = 0.028) and the Sortino ratio by 0.25 (p = 0.026). Interventions that forced stronger causal reasoning did not improve performance. Context: n=210 runs, 35 scenarios · Q4 2021 to Q3 2025 · Huang et al., arXiv:2609.29701 · p-values as reported, intervals not disclosed
The caveats
The Stanford authors stress their own limits: many configurations were tested on only two or three scenarios, giving wide confidence intervals, and some effects may need thousands of scenarios to pin down. We add one: the scenarios run from 2021 to 2025 with a recent model, and the paper does not discuss whether the model's training data covers those quarters. Our lookahead bias explainer shows why that matters.
What it shows / What it doesn't
It shows
- How the bull-versus-bear debate design works in AI trading agents, and what its best-known paper reported.
- That in a controlled test, debate mainly reduced drawdowns and volatility rather than adding returns over the S&P 500.
- That AI debaters tend to converge on the same view over rounds, and that preserving disagreement helped.
It doesn't show
- That bull-versus-bear agents beat the market. Neither study shows consistent outperformance across long periods.
- Durable evidence for the TradingAgents numbers. They come from one window of about three months on three stocks.
- Results free of lookahead risk. The Stanford scenarios may overlap the model's training data, which the paper does not address.
- How commercial products using this design perform. Neither paper tests one.
Why it matters
Bull-versus-bear AI is an appealing way to see both sides of a trade, and readable debates feel trustworthy. The evidence suggests treating it as a risk tool rather than a return engine, and watching for the moment both AIs start agreeing. When the bear case quietly disappears after a few rounds, the debate has stopped doing its job. That is also how our Bull AI vs Bear AI format will work: two published cases, side by side, with the decision left to the reader.
Primary sources
- VerifiedTradingAgents: Multi-Agents LLM Financial Trading Framework (arXiv:2412.20138, v7) (arXiv / UCLA, Tauric Research)Primary source. Architecture, Table 1, backtest window and Sharpe footnote checked in the HTML version on October 14, 2026.
- VerifiedMulti-Agent Debate for Explainable Trading: Reasoning, Consensus, and Performance in Simulated Markets (arXiv:2609.29701) (arXiv / Stanford University)Primary source. Preprint posted August 31, 2026. Correlations, Table 3, intervention results and limitations checked October 14, 2026.
- LinkedTradingAgents code (GitHub (TauricResearch))Open-source framework; not run by us.
In our directories
How we used AI: AI tools helped research and draft this article. Every figure was then checked against the primary sources listed above before publication.
Not financial advice. For information and education only. AI model outputs and research results can be wrong and do not predict future returns. Nothing here is a recommendation to buy or sell any asset. See our disclaimer and methodology.