Prediction Markets as Calibration Teachers for Real-Time Bidding: Market Pricing Meets Ad Auctions

Abstract

Real-time bidding (RTB) systems hinge on calibrated conversion estimates because small probability errors are amplified into large budget allocation and auction outcomes. We study whether prediction markets can serve as an external, explainable calibration teacher: market prices aggregate heterogeneous information into a probability that can be evaluated with proper scoring rules and translated into a prior for advertising probability calibration. We propose Market-Teacher Calibration (MTC), a Bayesian variant of Platt scaling whose prior is learned from resolved prediction markets. Unlike post-hoc calibration that relies exclusively on scarce advertising labels, MTC regularizes the calibration slope and intercept toward the regime observed in prediction markets, improving stability under label sparsity and distribution shift. We conduct end-to-end experimental evaluations on three publicly available datasets: the Manifold Markets contracts dump (57,333 resolved binary markets), a fixed Polymarket dataset-server snapshot (100 markets; 95 with resolved outcomes) accessed through the Hugging Face dataset-server interface, and the iPinYou advertising logs (3,000,000 impressions) for CTR modeling. Across datasets we report Brier score, log loss, and expected calibration error (ECE), and we simulate auction bidding with a fixed value-per-click policy to measure cost-per-click (CPC) and clicks per unit spend. On iPinYou, a deliberately miscalibrated CTR model trained with negative subsampling is corrected by MTC, reducing test log loss from 0.00767 to 0.00533 and improving clicks per 1000 cost units from 0.81 to 1.05 in a budgeted bidding simulation. Our results support the thesis that prediction market pricing can be repurposed as a transparent probabilistic anchor for ad systems, linking market efficiency to bidding robustness and interpretability.