ForesightFlow: An Information Leakage Score Framework for Prediction Markets

Battle Count: 58

Highly relevant to quantitative trading in prediction markets and market microstructure research. The paper provides a framework for detecting informed flow in real-time, which is directly actionable for trading strategies that follow detected informed flow. The microstructure adaptations (VPIN, Kyle's lambda, variance ratio, two-sidedness) are directly applicable to order-flow analysis. The deadline-ILS extension addresses the specific contract structures where insider trading has been empirically documented. However, the paper is primarily methodological; the detector has not been trained and no trading strategy is backtested. The framework is more immediately relevant to market surveillance, platform integrity, and regulatory monitoring than to direct alpha generation, though the detected signals could inform trading decisions.

High implementation complexity due to: (1) multi-source data pipeline integration (Polymarket APIs, UMA Oracle, GDELT BigQuery, Polygonscan, The Graph subgraph); (2) hierarchical T_news recovery with three tiers and manual review; (3) resolution typology classification requiring LLM-assisted fine-grained categorization; (4) adaptation of classical microstructure measures to discrete bounded-price binary markets; (5) per-category hazard-rate estimation via MLE; (6) wallet-level on-chain feature computation across 796K addresses; (7) multi-window ILS computation with scope-condition filtering; (8) production system architecture with dual backfill/streaming paths, Redis caching, TimescaleDB, and real-time WebSocket streaming. The methodology is well-specified but the full empirical pipeline requires substantial infrastructure work.

All system code, evaluation artifacts, the FFIC inventory, the resolution-typology classification of the 911,237-market corpus, and a live dashboard are released openly at https://github.com/ForesightFlow and https://foresightflow.xyz. Three accompanying datasets are released: polymarket-tnews-tevent-recovery-v1, polymarket-hazard-rates-v1, and polymarket-ils-corpus-v1. The FFIC inventory is released under CC-BY-4.0. The paper provides detailed notation, formal definitions, scope conditions, and a complete data pipeline description. However, the detector has not been trained and the full empirical evaluation is deferred to a companion paper.