The Earnings Call Risk and Confidence Analyzer (ERCA)

Zero-days-to-expiration (0DTE) options have transformedthe daily risk profile of equity markets, with much ofthe activity concentrated around quarterly earnings calls.Retail participants systematically buy short-dated im-plied volatility (IV) before the announcement and ab-sorb disproportionate losses from the IV collapse thatfollows the release, even when the directional bet iscorrect. We refer to this mechanism as the Volatil-ity Trap, and develop ERCA, a stochastic architec-ture designed to detect, measure, and exploit it in realtime. The framework integrates a three-channel filtration(FM ∨FOff ∨FSoc), a Hawkes self-exciting process for so-cial media activity with an O(1) recursive update suitablefor sub-300 ms latency, a Latent Profile Analysis withK = 8 archetypes including an irrational-exuberance cor-rection (β8 =−0.67), a sentiment-coupled jump-diffusionSDE under a Girsanov risk-neutral transformation, andthe Zshort(t) Sentiment Velocity Divergence indicator castas an optimal stopping problem. A DQN ensemble se-lector dynamically allocates among Neural CDE, Multi-Transformer, Bi-Transformer, and Online SVR forecast-ing heads as a function of the volatility regime, achiev-ing RMSE = 0.031 on live data, a 35.4% improvementon equal-weight averaging. Position sizing follows Frac-tional Kelly with c= 0.25 under a 15% drawdown circuitbreaker. Empirical validation on 500 real S&P 500 earn-ings events (2020–2026) yields a Spearman ρ = 0.4773(p < 0.0001) between Zmaxshort and the absolute EPS sur-prise, and a two-sample t = 10.976 (p = 2.99 ×10−25)separating high- from low-surprise events. The frame-work is also deployed as an open-source Streamlit ap-plication (https://erca-live.streamlit.app) backedby live Yahoo Finance and SEC EDGAR data, with auser-visible DQN training loop, allowing the results in this paper to be reproduced directly in the browser.