ScholarGate
Get started
Regression modelEconometricsEconometrics / time seriesModel

Bayesian ARMA Model

Also known as: Bayesian ARMA, B-ARMA, Bayesian autoregressive moving average, ARMA with Bayesian inference

OriginatorBox & Jenkins (classical ARMA); Bayesian treatment developed through work of Zellner, Geweke, and others in 1970s–1980sYear1970s–1980sSources2Related methods8

The Bayesian ARMA model applies Bayesian inference to the classical autoregressive moving average framework for stationary univariate time series. Rather than producing single point estimates for the AR and MA parameters, it yields full posterior distributions, naturally incorporating prior knowledge and providing coherent uncertainty quantification over forecasts and impulse responses.

Key highlights

  • Provides full posterior distributions for all parameters, giving credible intervals rather than asymptotic confidence intervals.
  • Naturally propagates parameter uncertainty into multi-step forecasts, producing honest probabilistic prediction bands.
  • Allows incorporation of prior economic knowledge (e.g. stationarity-inducing priors on AR coefficients).
  • Handles small samples better than classical MLE by regularising estimates through the prior.
  • Facilitates formal Bayesian model comparison across different ARMA orders via marginal likelihoods or WAIC.

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 Bayesian ARMA when you have a stationary univariate time series and need honest uncertainty quantification beyond point estimates, when the sample is small and you can encode economically motivated priors, or when you want to propagate parameter uncertainty into forecasts and scenario analysis. It is also valuable when comparing competing ARMA orders via Bayes factors. Do not use it as a substitute for pre-testing stationarity: if the series contains a unit root, Bayesian ARIMA or Bayesian VECM is appropriate instead. For large datasets with many series, the computational cost of MCMC may favour classical ARMA or Bayesian VAR.

Strengths & limitations

Strengths
  • Provides full posterior distributions for all parameters, giving credible intervals rather than asymptotic confidence intervals.
  • Naturally propagates parameter uncertainty into multi-step forecasts, producing honest probabilistic prediction bands.
  • Allows incorporation of prior economic knowledge (e.g. stationarity-inducing priors on AR coefficients).
  • Handles small samples better than classical MLE by regularising estimates through the prior.
  • Facilitates formal Bayesian model comparison across different ARMA orders via marginal likelihoods or WAIC.
Limitations
  • Computationally intensive: MCMC sampling can be slow for long series or high-order models.
  • Posterior inference is sensitive to the choice of prior, especially in small samples where the prior dominates.
  • Requires stationarity; applying a Bayesian ARMA to a unit-root series yields misleading results.
  • Order selection within the Bayesian framework (model averaging or reversible-jump MCMC) adds complexity compared to classical information-criteria selection.

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

How is a Bayesian ARMA different from classical ARMA?

Classical ARMA uses maximum likelihood to obtain point estimates; Bayesian ARMA places priors on the parameters and updates them with the data to produce full posterior distributions. This yields credible intervals and probabilistic forecasts that account for parameter uncertainty, not just sampling variability.

What priors should I choose for the AR and MA coefficients?

A common choice is independent Normal(0, σ_prior²) priors centred at zero, which shrinks coefficients toward zero and implicitly favours parsimony. For AR parameters, priors can be truncated to the stationarity region. Inverse-gamma priors are standard for the error variance.

Do I need to difference the series before fitting a Bayesian ARMA?

Yes. Bayesian ARMA assumes stationarity. If unit-root tests indicate non-stationarity, difference the series (or use a Bayesian ARIMA that includes the differencing order d) before estimation.

How do I choose the AR order p and MA order q in a Bayesian framework?

You can use information criteria on the classical model as a starting guide, or perform formal Bayesian model comparison via marginal likelihoods (e.g., computed by bridge sampling) or WAIC across candidate orders. Bayesian model averaging integrates over order uncertainty entirely.

How many MCMC iterations are typically needed?

A common practice is 4 000–10 000 warm-up (burn-in) iterations followed by 10 000–50 000 sampling iterations, with convergence assessed via R-hat < 1.01 and effective sample size > 400 for each parameter.

Sources

  1. 1.
    Geweke, J., & Meese, R. (1981). Estimating regression models of finite but unknown order. International Economic Review, 22(1), 55–70.
    https://scholar.google.com/scholar?q=Estimating+regression+models+of+finite+but+unknown+order+Geweke+Meese+1981
  2. 2.
    Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time Series Analysis: Forecasting and Control (5th ed.). Wiley.
    ISBN 978-1118675021

You have read it. What now?

Ask about this methodSimilar methods

Cite this page

ScholarGate. (2026, June 3). Bayesian ARMA model. ScholarGate. https://scholargate.app/econometrics/bayesian-arma-model

Bayesian ARMA Model | ScholarGate