Fuzzy Jump Models for Soft and Hard Clustering of Multivariate Time Series Data

Abstract

Statistical jump models have been recently introduced to detect persistent regimes by clustering temporal features and discouraging frequent regime changes. However, they are limited to hard clustering and thereby do not account for uncertainty in state assignments. This work presents an extension of the statistical jump model that incorporates uncertainty estimation in cluster membership. Leveraging the similarities between statistical jump models and the fuzzy c-means framework, our fuzzy jump model sequentially estimates time-varying state probabilities. Our approach offers high flexibility, as it supports both soft and hard clustering through the tuning of a fuzziness parameter, and it naturally accommodates multivariate time series data of mixed types. Through a simulation study, we evaluate the ability of the proposed model to accurately estimate the true latent-state distribution, demonstrating that it outperforms competing approaches under high cluster assignment uncertainty. We further demonstrate its utility on two empirical applications: first, by automatically identifying co-orbital regimes in the three-body problem, a novel application with important implications for understanding asteroid behavior and designing interplanetary mission trajectories; and second, on a financial dataset of five assets representing distinct market sectors (equities, bonds, foreign exchange, cryptocurrencies, and utilities), where the model accurately tracks both bull and bear market phases.