Real-Time Distributional Regime Monitoring: Online Wasserstein Detection of Market Stress

Abstract

We develop a real-time framework for monitoring market stress through the geometry of rolling return distributions. A Wasserstein score locates each current distribution relative to low- and high-stress prototypes learned from prior observations; thresholds are calibrated before 2007 and frozen. Across four equity indices from 2000 through June 2026, raw DRP detects 21 of 28 market–event cases, compared with 15 for realized volatility, 14 for EWMA, and 13 for fixed-parameter GARCH, at similar mean offwindow alarm-day shares of 3.1%–3.5%. These coverage differences remain positive after each of the seven episodes is removed in turn. A direct distribution-sensitive comparison uses rolling Kolmogorov– Smirnov monitors with past-only reference samples. With training alarm budgets aligned, DRP and the expanding-reference KS monitor each detect 21 cases. Among their 17 jointly detected cases, DRP has a median paired lead-time advantage of seven trading days; off-window alarm-day shares are 3.37% and 2.59%, respectively. The comparison identifies the response-time and alarm-duration dimensions of prototype-based stress alignment. Controlled experiments establish the underlying information channel: under a variance-matched shape change, raw and shape-only DRP achieve AUCs of 0.842 and 0.854, versus 0.524 for realized volatility. The framework extends financial-stress surveillance from the magnitude of fluctuations to the evolving configuration of the return distribution, linking geometric state measurement to event coverage, response time, and risk review.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.