PolySwarm: A Multi-Agent Large Language Model Framework for Prediction Market Trading and Latency Arbitrage

In one sentence

Summary

This paper addresses the limitations of single-model LLMs in financial forecasting, specifically hallucination, miscalibration, and prompt sensitivity, by proposing PolySwarm, a multi-agent system for real-time prediction market trading. The authors argue that swarm intelligence can suppress idiosyncratic errors through diversity, positioning the framework as a solution to the epistemic uncertainty inherent in transformer-based inference for high-stakes financial decisions. The system targets decentralized platforms like Polymarket, leveraging their liquidity and binary outcome structures to test automated forecasting and arbitrage strategies. The methodology involves deploying a swarm of 50 heterogeneous LLM personas with distinct analytical priors (e.g., macro, technical, contrarian) that operate independently to avoid anchoring. Individual probability estimates are aggregated using confidence-weighted Bayesian model averaging, combined with market-implied probabilities via a 70/30 linear mixture. The system employs an information-theoretic engine using Kullback-Leibler and Jensen-Shannon divergences to detect cross-market inefficiencies and negation pair mispricings. A latency arbitrage module exploits stale prices by deriving CEX-implied probabilities from a log-normal model. Experiments evaluate the system on Polymarket using Brier scores, log-loss, and calibration analysis, benchmarking against human superforecasters and single-model baselines. Results demonstrate that swarm aggregation consistently outperforms single-model baselines in probability calibration, validating the wisdom-of-crowds effect in LLM ensembles. The system executes trades using quarter-Kelly position sizing for risk control. Key limitations include the risk of agent hallucination, high computational costs at scale, regulatory exposure in decentralized finance, and feedback-loop risks. The paper concludes that while multi-agent systems improve calibration, they require careful design to mitigate correlated errors and ensure robustness in live trading environments, suggesting future work on adaptive calibration and human-AI collaboration.

Research problem

Single-model LLMs suffer from hallucination, miscalibration, and high variance in financial forecasting, leading to unreliable trading signals. The paper addresses how to aggregate diverse LLM predictions to improve calibration and detect market inefficiencies in prediction markets.

Core contributions

• PolySwarm system design: A production-ready multi-agent LLM trading terminal with 50 diverse personas on Polymarket.
• Confidence-weighted Bayesian aggregation: A principled ensemble method combining swarm consensus with market probabilities.
• Information-theoretic analysis engine: Novel application of KL and JS divergence for cross-market inefficiency detection.
• Latency arbitrage module: A CEX-to-DEX arbitrage pipeline using log-normal pricing models to exploit stale prices.
• Empirical evaluation: Systematic benchmarking using Brier scores and calibration analysis against human superforecasters.

Method

• Deploy 50 heterogeneous LLM personas with distinct analytical priors and information access patterns.
• Independent parallel sampling to maximize prediction diversity and prevent inter-agent anchoring.
• Confidence-weighted Bayesian model averaging to aggregate individual agent probability estimates.
• Linear Bayesian mixture (70/30) combining swarm consensus with market-implied probabilities.
• Kullback-Leibler and Jensen-Shannon divergence analysis to detect cross-market inefficiencies.
• Log-normal pricing model to derive CEX-implied probabilities for latency arbitrage signals.
• Quarter-Kelly position sizing for risk-controlled trade execution based on edge estimates.
• Asynchronous execution pipeline with LLM caching to reduce computational latency and API costs.

Results & findings

• Swarm aggregation consistently outperforms single-model baselines in probability calibration on Polymarket tasks.
• The 70/30 Bayesian mixture of swarm consensus and market probabilities optimizes risk-adjusted returns.
• KL and JS divergence effectively identify cross-market inefficiencies and negation pair mispricings.
• Latency arbitrage module exploits stale Polymarket prices within human reaction-time windows.
• Quarter-Kelly position sizing provides risk-controlled execution, limiting exposure to tail events.
• Heterogeneous persona design reduces correlated errors compared to homogeneous agent pools.

Data & experiments

• Data source: Polymarket prediction markets (binary outcomes) and CEX price data for arbitrage signals.
• Asset scope: Political, economic, and scientific outcome categories on decentralized platforms.
• Benchmarks: Human superforecaster performance, single-model LLM baselines, and naive market-implied probabilities.
• Metrics: Brier scores, log-loss, calibration analysis, and position sizing via quarter-Kelly criterion.
• Experimental design: Asynchronous execution pipeline with paper and live trading modes.

Limitations & caveats

• Hallucination in agent pools remains a critical risk, potentially leading to erroneous trading signals.
• High computational cost at scale due to concurrent inference across 50 diverse LLM personas.
• Regulatory exposure in decentralized prediction markets, which may lack clear legal frameworks.
• Feedback-loop risk where trading activity influences market prices, potentially distorting signals.
• Sensitivity to the choice of persona diversity and aggregation weights, requiring careful hyperparameter tuning.
• Limited generalizability to highly liquid, efficient traditional markets where arbitrage opportunities are fleeting.

Curator notes

Discovery status: institutional_trading_focus_promoted. Source links: execution agent large language model trading