agentic-prediction-arbitrage

This project combines natural language processing with arbitrage detection in a novel way. Instead of purely statistical arbitrage (price mismatches), it uses NLP to analyze the credibility of prediction market event descriptions and their resolution sources. The pipeline likely scrapes new markets from Polymarket and UMA, extracts the event description and resolution criteria, then applies a credibility classifier to assess whether the market's implied probability matches the factual likelihood based on the resolution source (e.g., official reports, news articles, oracles). When a significant discrepancy is found—a market priced at 20% but the NLP model estimates 40% probability based on source credibility—it flags an arbitrage opportunity. The Python codebase is small (10KB), suggesting a focused implementation without heavy frameworks. The repo includes web scraping components for both Polymarket and UMA protocols, indicating cross-platform awareness. This approach is particularly valuable for event-driven markets (political elections, scientific outcomes) where narrative and credibility play a huge role in pricing. The main challenge is building a reliable credibility model—NLP uncertainty could lead to false positives. The project is very new (June 2026) and lacks usage examples, but the concept is strong. For developers interested in AI-augmented trading strategies, this offers a fresh direction beyond pure price analysis. Worth cloning to examine the NLP model and contribute improved training data. Suitable for ML engineers and prediction market enthusiasts.