When Drift Breaks: Particle-Based Real-Time Regime Detection

Engineering notes

Key topics: autonomous driving. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Whereas existing approaches to financial regime detection calibrate thresholds to historical price-and-return statistics, we propose a framework built on immunisation: a particle filter whose observation model is calibrated to the distributional shape of market stress rather than specific historical episodes. The particle filter—with theoretical foundations in probabilistic robotics and autonomous driving, and adapted to financial markets—forms the inferential backbone. Its observation model stacks distributional shape descriptors—skewness, tail asymmetry, kurtosis, and the share of leading sector-eigenvalue energy in cross-asset return covariance—computed across a hierarchy of temporal windows and injected as structured distributional archetypes, replacing the random initialisation of Thrun and Burgard, and of Reisinger. Applied to S&P 500 across four distinct crises (Dotcom 2002, Lehman 2009, COVID-19 2020, and the 2022 inflation-driven bear market), the descriptors show pre-crisis discrimination, with effect sizes (Cohen’s d) of 2.9 or more for realised volatility, Bowley downside skewness, and tail-quantile features, and 1.3 for sector concentration (λ1 ratio). Out of sample (2015–2026), the pipeline confirms endogenous regime transitions with lead-time before the market trough. With the flexibility of particle filters, the richness of the observation model is the primary enabler of real-time regime detection. We frame the system as a distributional-shape monitor that detects regime transitions in real time, not a pre-peak forecaster: highly sensitive, it registers deformation as stress becomes measurable, with the attendant sensitivity–specificity trade-off. On the abrupt COVID-19 shock, it confirms the transition sixteen trading days before the trough, without claiming pre-peak detection; confirmation timing scales with each crisis’s own duration, from roughly two to three weeks for the fastest episodes to several months for the slowest.

Links and sources

Comments