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

Battle Count: 63

Highly relevant to quantitative trading. The paper presents a complete end-to-end trading system with real-time market scanning, probability estimation, Bayesian aggregation, information-theoretic arbitrage detection (KL/JS divergence), latency arbitrage exploiting CEX-to-DEX price gaps, quarter-Kelly position sizing, and risk management (daily loss limits, position caps). It directly addresses prediction market trading on Polymarket with both paper and live execution modes. The methodology integrates multiple quantitative finance concepts (Kelly criterion, Black-Scholes log-normal pricing, efficient markets hypothesis, Grossman-Stiglitz paradox) with modern AI/ML approaches. The latency arbitrage module and cross-market inefficiency detection are directly applicable to quantitative trading strategies.

High implementation complexity. Requires: (1) multi-provider LLM API integration (OpenAI, Anthropic, Ollama) with rate limiting via asyncio semaphores; (2) 50-persona prompt engineering with chain-of-thought elicitation; (3) confidence-weighted Bayesian aggregation pipeline; (4) KL/JS divergence computation for cross-market analysis; (5) log-normal pricing model for CEX-implied probabilities; (6) Polymarket CLOB API integration via py-clob-client on Polygon blockchain; (7) Binance WebSocket price feed integration; (8) quarter-Kelly position sizing with hard caps and daily loss limits; (9) 5-second scan loop with volume filtering; (10) SQLite async database for persistence; (11) FastAPI REST/WebSocket server; (12) Vue 3 real-time dashboard. The system spans AI/ML, quantitative finance, blockchain, and full-stack web development.

The paper provides full architectural description, implementation details (Python FastAPI backend, Vue 3 frontend, SQLite database, asyncio pipeline), and evaluation methodology. However, no GitHub repository URL is explicitly provided in the extract. The system uses proprietary API providers (OpenAI, Anthropic) which introduce non-determinism. The 50-persona pool definitions, specific prompt templates, and hyperparameter configurations are described but not fully enumerated. Paper/live trading modes are described. The arXiv preprint is available but code availability is uncertain.