Aug 2026· Journal of Risk and Financial Management· Vol 19, pp. 612· 0 citations· 28 references
TL;DR
This work frames 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.
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.
Stock price forecasting is fundamentally difficult because of market non-stationarity, high volatility, and the weak predictability of intraday returns. While kernel methods with random features provide a theoretically grounded framework for capturing non-linear relationships, they are typically applied with a single,...
Yan-Zhu Guan· Applied and Computational En...· 0 citations
Non-stationary forecasting is not the same as forecasting in average terms; forecasting performance depends on the ability to forecast regime changes rather than the ability to fit perfectly in average cases. Regime shifts, calendar effects, policy interventions, and unreliable instrumentation constantly disrupt energy...
P. Shelar· Journal of Intelligent Decis...· 0 citations
Change points that arrive in quick succession, or near the boundaries of the sample, are prone to type-II error: the few observations trapped between successive breaks carry too little information for global procedures to analyze. This article develops a test statistic, $I_1$, for which we derive the limiting null dist...
Nachiketas Waychal, A. Laha, Ankur Sinha· 0 citations
Financial markets fluctuate under several relatively persistent conditions, and returns, volatility, correlation, and risk patterns vary depending on the conditions. Because market regimes and their boundaries are not directly observable and widely recognized labels are unavailable, unsupervised learning provides a sui...
Ru-Yue Xu· Theoretical and Natural Scie...· 0 citations
Time-series foundation models are increasingly adapted to new domains through fine-tuning on target data, under the implicit assumption that more target data yields better forecasts. We show that this assumption can fail in financial forecasting, where individual price changes are difficult to predict, but large moves...
Manh Nguyen, M. Nguyen, H. Nguyen et al.· arXiv.org· 0 citations
Learning a forecasting model's errors offers a modular way to improve its predictive distribution while retaining its underlying dynamics. We study this approach with Flowz, a conditional flow model of direct multi-step price-level residuals fitted using random features and ridge regression. Each EWMA, ARIMA, ARFIMA, a...
G. Vercellino· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.