How Wise is the Crowd? Bias and Edge in Prediction Markets

Prediction markets are increasingly relied upon as real-time probability oracles, yet their predictive signals remain polluted by structural inefficiencies. While prior literature documents anomalies like the favorite-longshot bias at an aggregate level, the microstructural origins of these distortions—specifically who generates and exploits them—remain unstudied in modern ecosystems. To investigate this, we engineer a scalable, multi-threaded data architecture capable of synchronously ingesting and persisting tick-level order flow, decentralized wallet histories, and user commentary across Polymarket and Kalshi. Moving beyond traditional discrete quantile-binning, we introduce a continuous, multivariate framework to isolate aggregate market bias from individual trader edge across the probability spectrum and contract lifecycle. Our findings challenge the idea that favorite-longshot bias is present in every prediction market. In the markets we find it to be present, such as Mention Markets, the classic favorite-longshot bias may in fact be a statistical artifact masking a pervasive “Yes Bias”, driven by extreme temporal volatility and not controlling for the time to market completion in previous methodologies. Furthermore, we find that “Whales”, or the most capitalized players, are not the most sophisticated. By dynamically reconstructing participant positions, we demonstrate that Whales, on average, systematically bleed expected value to small-order traders. Rather than acting as sharp informed players, these large actors likely trade on ideological conviction, structurally overpaying for specific narratives and suffering from adverse selection against smaller participants. Utilizing a Natural Language Inference (NLI) stance-detection framework, we see to evident interaction between vocality of a trader and realized market performance. We find no statistically significant correlation between sentiment intensity and informational edge, indicating that the most prominent voices in these ecosystems function primarily as sources of communicative noise. Ultimately, this research provides a reproducible methodology for actively denoising prediction market probabilities by adjusting for temporal lifecycles, liquidity environments, and the ideological biases held by heavily capitalized, unsophisticated traders.