A new arXiv paper (2508.00554) introduces ContestTrade, a multi-agent trading system built on an internal competition mechanism. The concept is appealing: two teams of LLM agents (Data Team and Research Team) compete through a 'Quantify-Predict-Allocate' process to produce trading decisions. But before getting excited, let's apply Strategy Arena's Anti-2CV methodology to assess the robustness of the reported results.
What the paper says
ContestTrade is inspired by institutional investment workflows. The Data Team transforms massive market data into textual factors optimized for limited LLM context windows. The Research Team generates parallelized multi-path trading decisions via deep research tools. The internal competition mechanism scores agents after market outcomes, predicts future utility from historical scores, and allocates resources accordingly.
Anti-2CV analysis
-
Fees and costs: The paper does not explicitly mention transaction fees, slippage, or LLM API call costs. In a multi-agent system, these costs can quickly accumulate and wipe out theoretical gains.
-
Paper trading vs. live: The reported results come from backtests or paper trading. No evidence of live trading with real capital is provided. Survival bias, look-ahead bias, and period selection bias are not discussed.
-
Monte Carlo and cross-validation: No Monte Carlo sensitivity analysis or temporal cross-validation is reported. Without these tests, it is impossible to know whether the performance is due to luck or overfitting on historical data.
-
Data leaks: The 'Quantify-Predict-Allocate' mechanism uses past scores to predict future utility. If scores are computed on the same data used for training or validation, there is a high risk of look-ahead bias.
Our verdict
ContestTrade is an interesting architecture, but the published results are not yet validated by a rigorous methodology. For such a system to be credible, one would need: - To publish transaction fees and LLM inference costs. - To perform backtests with temporal cross-validation and Monte Carlo simulations. - To demonstrate that the internal competition mechanism does not create selection bias.
## Caveat
This article is based on an academic preprint. The results mentioned come from backtests or paper trading and do not constitute proof of profitability in live conditions. Algorithmic trading carries high risks, including costs, liquidity issues, and market regime changes. Do not make any investment decisions based solely on this analysis.
References - Original paper: ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism - Strategy Arena Anti-2CV methodology: Public anti-2CV methodology: fees, paper trading caveats, MC CV and leak fixes