Nonparametric Regime Segmentation in Financial Time Series via Hilbert–ICEEMDAN and Penalized Change-Point Inference

This paper presents a novel methodological framework for detecting structural breaks in multivariate financial time series using Hilbert–ICEEMDAN decomposition combined with penalized change-point estimation. The approach decomposes 12 major financial assets over 17+ years (5,150 observations) into piecewise-stationary Intrinsic Mode Functions (IMFs) using ICEEMDAN, enabling well-defined instantaneous frequency extraction via Hilbert spectral analysis. IMF selection employs median period filtering within the 5-60 day range, with cross-asset aggregation through robust median-based combination of amplitude and frequency components. Change-point detection uses the PELT algorithm with least-squares cost function, penalty parameter β=5.0, and minimum segment length of 63 days. The framework identifies 57 distinct regime periods with median duration of 69 days, demonstrating consistency across parameter sensitivity analysis (125 combinations tested) and robustness diagnostics (9/9 validation checks passed). As an empirical illustration, regime-switched trading strategies achieve Sharpe ratios of 0.778 versus 0.381 for equalweight benchmarks, with walk-forward validation showing median Sharpe ratios of 1.155 for 5-year training periods. The system successfully captures major market events including the 2008 financial crisis, 2010 flash crash, 2011 European debt crisis, 2018 volatility spike, 2020 COVID crash, and 2022 Ukraine conflict. MSC 2020: 62M10, 62M20, 91G05, 62L12, 91G70 JEL: C22, C58, C14, C32, G11