A1282
Title: Detecting structural breaks in financial time series
Authors: Tommaso Venturino - ETH Zurich (Switzerland)
Patrick Cheridito - ETH Zurich (Switzerland)
Urban Ulrych - ETH Zurich (Switzerland) [presenting]
Abstract: Financial time series are subject to frequent structural instability, which complicates forecasting, risk measurement, and real-time monitoring. A unified framework is proposed that integrates adaptive state-space diagnostics with global offline segmentation and strictly causal online regime recognition. Offline, a Kalman filter with Dempster-Shafer covariance adaptation is combined with penalized exact segmentation to localize structural breaks across heterogeneous data-generating processes. Online, break detection is formulated as a filtration-respecting probabilistic classification problem and implemented using a LightGBM learner based on strictly causal features derived from both raw returns and model diagnostics. Across synthetic processes with known ground truth, the framework achieves balanced break identification offline and stable short-horizon recognition relative to variance-based and Bayesian benchmarks. Applied to a cross-asset panel of financial series, it identifies economically interpretable regime shifts while maintaining parsimonious segmentation. The results highlight the complementarity of adaptive state-space monitoring and causal machine learning for structural break analysis in financial econometrics.
