PROPHET: A Multi-Modal Ensemble Framework for Calibrated Probability Forecasting in Decentralized Prediction Markets

Abstract

Can machine learning reliably outperform the collective wisdom of prediction market crowds? We present PROPHET, a three-tower ensemble that jointly exploits temporal price dynamics (TFT + TCN), event-level language understanding (RAG-augmented GPT-4 + FinBERT), and order-book microstructure (Graph Attention Network). These heterogeneous signals are fused through a Bayesian model averaging meta-learner over 26 interpretable features, with final calibration provided by Regularised Adaptive Prediction Sets (RAPS). On 47,832 resolved Polymarket contracts spanning January–October 2024—to our knowledge, the largest such benchmark reported in the literature—PROPHET achieves a Brier Score of 0.098 ± 0.004, a 19.0% reduction over the efficient-market baseline and 29.4% over the strongest single tower. Multi-horizon experiments at {6 h, 24 h, 3 d, 7 d} confirm these gains are robust and not driven by near-resolution information leakage. In a paper-trading simulation with realistic friction, the system yields a Sharpe ratio of 2.47; a subsequent 37-day live on-chain deployment produces a Sharpe of 2.14, independently verified by a third-party analytics platform. To the best of our knowledge, this constitutes the first publicly verifiable live validation in the prediction market ML literature. Live trading wallet: https://polymarket.com/ profile/0xe007275693587b5c472c374052d10f894b66a797. Keywords: Prediction markets Ensemble learning Probabilistic calibration Graph attention networks Temporal Fusion Transformer Conformal prediction

Ähnliche Arbeiten