Time Series MCMC
Also known as: MCMC time series, Bayesian time series MCMC, time series posterior sampling, sequential Bayesian MCMC
Time series MCMC applies Markov chain Monte Carlo methods to Bayesian inference over time-ordered data. Rather than optimising a single parameter estimate, it draws samples from the full joint posterior of parameters and latent states, yielding probability distributions that honestly reflect uncertainty about dynamics, trends, and seasonal patterns across every time point.
Key highlights
- Produces full posterior distributions over latent states and parameters, enabling calibrated credible intervals at every time point.
- Naturally handles missing observations, irregular sampling, and non-Gaussian measurement noise within the probabilistic framework.
- Prior information — known dynamics, mean-reversion rates, seasonality — can be formally encoded and updated with data.
- Works with short series where maximum-likelihood asymptotics break down.
- Model comparison via marginal likelihoods or WAIC is straightforward from the MCMC output.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use time series MCMC when you need full posterior uncertainty over latent states and parameters of a dynamic model — for example, in state space, hidden Markov, or dynamic linear models. It excels with short or irregularly-spaced series where asymptotic maximum-likelihood guarantees do not hold, and when genuine prior information (e.g., known seasonal structure) should be incorporated. Do not use it as a drop-in replacement for classical ARIMA when a simple frequentist fit suffices, and avoid it when computational cost is prohibitive and the series is long enough that MLE converges reliably.
Strengths & limitations
- Produces full posterior distributions over latent states and parameters, enabling calibrated credible intervals at every time point.
- Naturally handles missing observations, irregular sampling, and non-Gaussian measurement noise within the probabilistic framework.
- Prior information — known dynamics, mean-reversion rates, seasonality — can be formally encoded and updated with data.
- Works with short series where maximum-likelihood asymptotics break down.
- Model comparison via marginal likelihoods or WAIC is straightforward from the MCMC output.
- Computationally intensive: state space posteriors are high-dimensional and MCMC can be slow for long series.
- Convergence is harder to guarantee in non-linear models, requiring particle MCMC or HMC rather than simple Gibbs steps.
- Posterior inference is sensitive to prior choices for variance parameters, especially with short series.
- Requires careful model specification; misspecified state evolution equations lead to poor posterior predictions.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is the Forward Filtering Backward Sampling algorithm?
FFBS is a blocked Gibbs sampler for state space models. It first runs a Kalman filter forward through time to compute filtered marginal distributions, then samples an entire state trajectory in one pass backwards. This avoids sampling each state one-at-a-time, which would mix poorly, and is the standard workhorse for linear Gaussian time series MCMC.
When should I use particle MCMC instead of FFBS Gibbs sampling?
FFBS requires the observation and evolution equations to be linear and Gaussian. When your model is non-linear (e.g., logistic growth dynamics) or has non-Gaussian noise (e.g., count data, heavy tails), the Kalman filter steps are not exact. Particle MCMC — which embeds a particle filter inside an MCMC kernel — handles these cases at the cost of higher computational load.
How many MCMC iterations do I need for a time series model?
There is no universal answer, but state space models with many latent states can mix slowly. A common starting point is 2 000 warm-up iterations and 2 000 sampling iterations per chain, with at least 4 chains. Check R-hat (target below 1.01) and bulk/tail effective sample sizes before trusting summaries.
Can time series MCMC handle missing observations?
Yes. Within the state space framework, missing observations are simply time points where no likelihood contribution is added. The Kalman filter propagates uncertainty through the gap and FFBS samples the missing states from their prior-predictive, effectively imputing them as part of the posterior.
How does this differ from classical ARIMA?
ARIMA is a frequentist model estimated by maximum likelihood, returning point estimates and asymptotic standard errors. Time series MCMC is a Bayesian posterior sampling approach that returns full distributions over states and parameters, supports non-stationary or structurally changing dynamics, incorporates priors, and handles missing data naturally — at considerably higher computational cost.
Sources
- 1.Carter, C. K. & Kohn, R. (1994). On Gibbs sampling for state space models. Biometrika, 81(3), 541–553.
- 2.West, M. & Harrison, J. (1997). Bayesian Forecasting and Dynamic Models (2nd ed.). Springer.ISBN 978-0387947259
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Time series MCMC. ScholarGate. https://scholargate.app/bayesian/time-series-mcmc