Physics-informed modeling of persistent predictive penalty from vocal affect in markets

Non-verbal emotion shapes collective decisions, yet its predictive value at scale remains unclear. Using a physics-informed multimodal model that isolates clear vocal emotion from 1,795 dynamic and noisy earnings calls, we show a persistent, counterintuitive effect we term the Predictive Penalty : adding vocal affect to standard predictors makes post-call volatility less predictable. This penalty is small on day one, intensifies over the next two weeks, and persists for about a month, proving strongest during the unscripted Q&A. Vocal emotion does not forecast returns, pointing to sentiment-driven risk rather than new information. The pattern holds across industries, time periods, and volatility regimes and is robust against extensive checks. We recast executive emotion as a quantitative gauge of communication-induced risk and present a reproducible approach. Our work provides a blueprint for quantifying how the non-verbal dimension of communication injects measurable risk into high-stakes environments, from trading floors to treaty rooms.