Streaming hidden Markov models

Fri, 10 Apr 2026 00:00:00 +0000

Introduction

A classical method to tackle the problem of time-series regime detection and segmentation is the hidden Markov model (HMM). ¹ At the core of the HMM are two sub-models: a model for latent regimes and a predictive model for observations (conditioned on a regime).

In this post, we consider the problem of learning a parametric HMM when neither the regimes nor the model parameters are known. In this case, estimating the observation model parameters requires knowing which datapoints belong to which regime. However, knowing which datapoint belong to a given regime requires knowledge of the parameters for the model, which are unknown. This, at first glance, looks like a circular problem.

Sequential on Gerardo Duran-Martin

Streaming hidden Markov models

Introduction