Online Anytime-Valid Volatility Regime Detection with Particle Filtering

Volatility regime shifts must be detected online, without a prespecified horizon and without the multiplicity cost of repeated testing. Anytime-valid inference and e-detectors address this setting, but such a monitor needs an alternative hypothesis to bet against: a running estimate of current volatility. We use sequential Monte Carlo to supply one, coupling a test martingale for volatility shifts to a particle filter that tracks latent stochastic volatility, so that the betting alternative moves toward the new regime once a change occurs. The coupling has to respect predictability, and the filter offers a choice here: only its one-step-ahead predictive moment is measurable with respect to the past, while the filtered posterior mean it reports by default is not. The null is invariant under rescaling, so the same predictive moment can also be used without estimating a baseline volatility at all: the unknown $\sigma$ is quotiented out rather than plugged in. At matched realized false-alarm rates this invariant construction is the fastest of the seven detectors we test, with a median delay of 15 days against 37 for a plug-in baseline. On daily S&P~500 returns over 2025--2026 online SAVI finds six regime transitions and flags the April 2025 tariff shock on the day of each move, while a retrospective PELT scan places changepoints in quiet stretches with no evident catalyst.