Continuous Hidden Markov Models

• Continuous Hidden Markov Models are statistical models that extend classic HMMs to handle continuous-valued latent states and observations with interpretable Markov dynamics.
• They employ inference techniques like expectation-maximization, particle filtering, and matrix exponential computation to manage irregular, missing, or incomplete data.
• CHMMs are applied across diverse fields such as finance, ecology, healthcare, and machine learning to achieve scalable likelihood-based inference and reliable state decoding.

Continuous Hidden Markov Models (CHMMs) generalize the classical hidden Markov model framework by allowing the observed data—or, in some formulations, the underlying latent state space—to be continuous-valued. CHMMs are central to modern approaches in time series, longitudinal and survival analysis, finance, movement ecology, signal processing, disease modeling, and machine learning. They unify clustering, temporal dependence, and flexible handling of real-valued, possibly irregular and incomplete data, while enabling interpretable Markovian state dynamics and scalable likelihood-based inference.

1. Model Classes and Formal Definitions

Several mathematical instantiations of CHMMs exist, each oriented toward a particular structure of the latent process and the observation model.

Discrete-latent, continuous-observation CHMMs:

For sequence data $x_1,\ldots,x_T \in \mathbb{R}^d$, CHMMs posit a finite-state, time-homogeneous Markov chain $z_t \in \{1, \ldots, K\}$, with initial distribution $\pi$, transition matrix $A$, and state-dependent emission density $f_k(x; \theta_k)$. The joint sequence density is

$$p(x_{1:T}, z_{1:T}) = \pi_{z_1} f_{z_1}(x_1) \prod_{t=2}^T [A_{z_{t-1},z_t} f_{z_t}(x_t)]$$

This structure supports Gaussian, Student-t, Laplace, and Generalized Error emissions as detailed in (Alswaidan et al., 22 Jun 2026), as well as vector-valued observations and missing data patterns with tractable likelihoods (Pandolfi et al., 2021).

Continuous-time state process CHMMs:

In continuous-time models, the latent process $X_t$ is a time-homogeneous Markov jump process or diffusion on a discrete or continuous state space. For the jump-process case with generator matrix $Q$,

$$P(t) = \exp(Q t),\qquad P_{ij}(t)=\mathbb{P}[X_{t+\Delta t}=j \mid X_t=i]$$

Observations can be aligned to arbitrary (even irregular) time schedules, with emissions given by conditional densities such as Gaussian, categorical/multinomial, or more general exponential family forms (Verma et al., 2018, Wang et al., 2021, Liu et al., 2021). Model variants admit semi-Markov sojourn times or control for absorbing states, such as for dropout in longitudinal biomedical data (Pandolfi et al., 2021).

Continuous-state latent CHMMs:

A further generalization introduces a continuous latent state, so that $x_0\sim \mu$, $z_t \in \{1, \ldots, K\}$0, $z_t \in \{1, \ldots, K\}$1, with transition and emission densities parameterized analytically or with neural networks (Jarboui et al., 2021). The Kalman filter/smoother is a classical instance, subsumed by this formulation when $z_t \in \{1, \ldots, K\}$2 and $z_t \in \{1, \ldots, K\}$3 are linear-Gaussian.

2. Inference and Learning Algorithms

Inference in CHMMs, whether for discrete-time, continuous-time, or continuous-latent-state models, typically leverages expectation-maximization (EM) or Bayesian sampling with tractable recursion or approximation.

Discrete-latent, continuous-observation (Baum-Welch/ECM EM):

The E-step employs forward–backward recursions to compute state-marginal posteriors $z_t \in \{1, \ldots, K\}$4 and joint posteriors $z_t \in \{1, \ldots, K\}$5: $z_t \in \{1, \ldots, K\}$6

$z_t \in \{1, \ldots, K\}$7

The M-step maximizes soft-counts in closed form for parametric emissions (e.g., weighted mean and variance for Gaussian), with ECM or ECME subroutines for Student-t, Laplace, or GED emissions (Alswaidan et al., 22 Jun 2026). Algorithms accommodate missing and partly observed vectors (Pandolfi et al., 2021).

Continuous-time EM (finite-state jump processes):

Between observation times $z_t \in \{1, \ldots, K\}$8 and $z_t \in \{1, \ldots, K\}$9, the probability of state transitions is

$\pi$0

The forward–backward algorithm uses $\pi$1 to form $\pi$2 (probability of states $\pi$3 across $\pi$4), enabling efficient expected sufficient-statistic computation. Integral formulas for expected jump counts and sojourns are (see (Liu et al., 2021)): $\pi$5 Numerically, these are computed via eigendecomposition (when $\pi$6 is diagonalizable), block-matrix exponentiation, or uniformization (series expansion in $\pi$7).

Particle-based and Neural EM:

For continuous-latent CHMMs, inference makes no use of exact recursions. Instead, Sequential Monte Carlo (particle filtering) approximates the smoothing distribution $\pi$8, and a Monte Carlo EM updates parameters by gradient ascent on a sample-based lower bound (Jarboui et al., 2021, Wang et al., 2021).

Complexity:

Vectorized eigendecomposition yields $\pi$9 scaling, with fast implementations for large state spaces or high-frequency datasets (Liu et al., 2021).

3. Handling Missing Data, Irregular Spacing, and Dropout

CHMM frameworks provide structured solutions for incomplete, intermittent, or irregularly observed data.

Missing At Random (MAR):

When $A$0 is only partially observed at some $A$1, Gaussian emission models allow the likelihood to marginalize analytically over missing entries, relying only on observed-subvector means and covariance blocks. Conditional expectation and covariance formulas impute unobserved elements, preserving MAR validity (Pandolfi et al., 2021).

Dropout and Absorbing States:

To capture monotone missingness (subject dropout), an absorbing state $A$2 is added to the latent chain, with $A$3, and no emission is defined in this state. Forward–backward posteriors are set to degenerate support on the dropout state after the dropout time, and $A$4 captures drop-out intensities (Pandolfi et al., 2021).

Irregular or event-driven observation times:

Continuous-time HMMs support arbitrary observation schedules; all in-interval expectations and recursions are parameterized by time duration $A$5. This supports analysis of EHRs, movement paths, or any event-driven process (Verma et al., 2018, Blackwell, 2018).

4. Theoretical Properties and Identifiability

Equivalence and model reduction:

For CHMMs with continuous observations, the equivalence problem—deciding whether two CHMMs specify the same law on observation sequences—admits a polynomial-time algorithm. This leverages the observation-density matrix's functional decomposition and a reduction to finite-symbol HMM equivalence (Darwin et al., 2020). Key insight: emission-density matrices $A$6 can be linearly decomposed; equivalence then becomes an algebraic orthogonality problem.

Spectral properties and autocorrelation structure:

The spectral identity

$A$7

with $A$8 as nontrivial eigenvalues of $A$9, bounds the number of temporal autocorrelation decay modes by the rank of $f_k(x; \theta_k)$0 ($f_k(x; \theta_k)$1). For $f_k(x; \theta_k)$2, a CHMM can generate the empirically observed slow decay of absolute-value autocorrelations in financial returns, without needing semi-Markov generalizations (Alswaidan et al., 22 Jun 2026).

Extensions—semi-Markov, covariates, and infinite/unknown state models:

CHMMs are generalizable to hidden semi-Markov models (arbitrary sojourn distributions), latent diffusions, regime-switching SDEs, and allow incorporation of covariate effects on transition rates or emission laws (Engelmann et al., 2022, Wang et al., 2021, Krishnamurthy et al., 2016). Bayesian models with an unknown number of latent states use reversible-jump MCMC for model selection and clustering (Luo et al., 2021), enabling simultaneous inference of state- and cluster-number in large-scale heterogeneous panels.

5. Applications in Biomedical, Financial, and Ecological Domains

Longitudinal data with missingness and dropout:

In clinical longitudinal panels, CHMMs support clustering of latent health status, account for MAR and dropout, and allow covariates to affect transitions via initial/transition logits. Monte Carlo studies validate low-bias and robustness to moderate dropout rates (Pandolfi et al., 2021).

Disease progression and healthcare utilization:

For progression in chronic disease cohorts (e.g., COPD, glaucoma, Alzheimer's), continuous-time HMMs support irregular visit times, multinomial or Gaussian emissions, and large numbers of states; fitted models recover interpretable latent structure, robust transition matrices, and accurate counterfactual/predictive inference (Verma et al., 2018, Liu et al., 2021).

Financial and econometric time series:

CHMMs with heavy-tailed emission families (Student-t, GED, Laplace) reproduce clustering, excess kurtosis, and slow ACF decay in daily equity returns, closing known gaps of Gaussian-state HMMs. Copula-composed multi-asset CHMMs preserve marginal and cross-asset properties relevant for risk management and simulation (Alswaidan et al., 22 Jun 2026). Filter-based continuous-time HMMs provide explicit formulas for regime-filtered stochastic volatility and leverage phenomena (Krishnamurthy et al., 2016).

Animal movement and ecology:

Integrated CHMMs with path-dependent emission models support inference of latent movement behavioral states over continuous time, enhancing Monte Carlo and MCMC efficiency for large tracking datasets (Blackwell, 2018).

Machine learning and signal processing:

Neuralized CHMMs, with arbitrary parameterizations of transition and emission densities, enable flexibility beyond exponential families and match or outperform deep RNNs (LSTM/GRU) in segmenting and predicting structured temporal data with interpretable latent embeddings (Jarboui et al., 2021).

6. Computational and Algorithmic Considerations

Efficient implementation depends on the form of the latent chain, the emission models, data regularity, and missingness.

• Finite-state, discrete-time CHMMs: Utilize vectorized Baum-Welch/ECM for forward-backward and closed-form updates per emission family. Complexity $f_k(x; \theta_k)$3 per sequence.
• Continuous-time, finite-state: Leverage matrix exponentials ($f_k(x; \theta_k)$4 per interval), or uniformization ($f_k(x; \theta_k)$5) for large-scale/ill-conditioned problems. Special routines for large $f_k(x; \theta_k)$6 or complex state spaces (e.g., biophysical grids or high-resolution progression models) (Liu et al., 2021).
• Continuous-latent, neural models: Require high-throughput Sequential Monte Carlo with GPU/mini-batch support; inference and learning are sample-based, with gradient ascent in parameter space.
• Semi-Markov and path-dependent models: Require solution of integro-differential forward-backward equations, best implemented with Volterra solvers and adaptive quadrature (Engelmann et al., 2022).
• Handling equivalence and reduction: Polynomial-time algorithms based on functional decompositions manage model specification and comparison for both analytic and composite emission densities (Darwin et al., 2020).

7. Significance, Limitations, and Extensions

CHMMs unify the modeling of sequential, irregular, missing, and heavy-tailed real-valued data with interpretable latent structures, offering scalable, likelihood-based inference and predictive state decoding across science and engineering. Limitations include potential identifiability or label-switching for large state spaces, memorylessness of exponential sojourns (addressed by semi-Markov extensions), non-convexity of neural parameterizations, and computational scaling for high-frequency or high-dimensional state constructs.

Current frontiers include unsupervised learning of both state number and latent cluster structure in large, heterogeneous populations, efficient scan methods for massive event-driven data, and neural–Bayesian hybrids that simultaneously capture nonparametric emission structure and interpretable latent states (Luo et al., 2021, Alswaidan et al., 22 Jun 2026, Jarboui et al., 2021). The flexibility and extensibility of the CHMM paradigm, supported by unifying algorithmic advances, continue to drive its adoption and methodological innovation.