Time-partitioning Gaussian Process Mixture Model for Persistent Regime Identification in Time Series
ORIGINAL / Time-partitioning Gaussian Process Mixture Model
TPGPMM is a Bayesian nonparametric framework combining an adaptive Pitman-Yor prior on contiguous segmentations with regime-specific Gaussian processes to identify persistent regimes and model nonstationary dependence without Markovian assumptions. Simulations recover regime structure; applications to VIX and 10-year Treasury yield identify regimes aligned with market events.
01 ABSTRACT
The authors propose TPGPMM, a Bayesian nonparametric model for regime identification and nonstationary dependence modeling in time series. Simulations demonstrate recovery of regime structure, and real-data applications identify regimes aligned with market events. The authors note that variants estimating observation noise improve predictive accuracy but can yield less sparse or less event-aligned segmentations; a regime-specific mean variant fits better than a stationary Gaussian process with comparable forecast accuracy.
02 KEY FINDINGS
- TPGPMM combines an adaptive Pitman-Yor prior with regime-specific Gaussian processes, learning the discount parameter to adapt to homogeneous or heterogeneous regime lengths.
- A split-merge Metropolis-within-Gibbs sampler with exact Gibbs boundary updates jointly learns change points and hyperparameters while quantifying uncertainty.
- Applied to VIX and 10-year Treasury yield, it identifies regimes aligned with major market and monetary-policy events, revealing a trade-off between interpretability and predictive accuracy.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG