Bayesian ARIMA Model
Also known as: Bayesian ARIMA, BARIMA, Bayesian Box-Jenkins model, Bayesian integrated time series model
The Bayesian ARIMA model combines the classical Box-Jenkins ARIMA framework with Bayesian inference. Instead of obtaining single point estimates for autoregressive and moving average parameters, it places prior distributions over them and uses observed data to update beliefs into a full posterior distribution, enabling coherent uncertainty quantification and probabilistic forecasting.
Key highlights
- Produces full posterior and predictive distributions, enabling honest uncertainty quantification beyond classical confidence intervals.
- Naturally incorporates informative prior knowledge about parameter magnitudes or stationarity constraints.
- Principled model comparison via Bayes factors or DIC without ad hoc information criteria.
- Better calibrated forecast intervals in small samples where MLE is imprecise.
- Posterior predictive checks allow transparent model adequacy assessment.
- Extendable to hierarchical or time-varying structures within a unified probabilistic framework.
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 ARIMA when you have a univariate stationary or trend-differenced time series and wish to incorporate prior knowledge about the dynamics, obtain full predictive distributions rather than point forecasts, or honestly propagate parameter uncertainty into forecast intervals. It is especially valuable with short series where classical maximum-likelihood estimates are unreliable, or when you want principled model comparison via Bayes factors. Do not use it as a simple drop-in for classical ARIMA if computational resources are limited and the series is long with no informative priors — classical ARIMA or frequentist alternatives are faster and nearly as accurate in large samples. Also avoid it when the assumption of Gaussian innovations is severely violated without a suitable heavy-tailed likelihood extension.
Strengths & limitations
- Produces full posterior and predictive distributions, enabling honest uncertainty quantification beyond classical confidence intervals.
- Naturally incorporates informative prior knowledge about parameter magnitudes or stationarity constraints.
- Principled model comparison via Bayes factors or DIC without ad hoc information criteria.
- Better calibrated forecast intervals in small samples where MLE is imprecise.
- Posterior predictive checks allow transparent model adequacy assessment.
- Extendable to hierarchical or time-varying structures within a unified probabilistic framework.
- Computationally expensive: MCMC sampling is orders of magnitude slower than closed-form OLS or MLE for classical ARIMA.
- Results are sensitive to prior specification, especially in small samples; poorly chosen priors can dominate the posterior.
- Order selection (p, d, q) still requires care; joint Bayesian search over orders increases model complexity significantly.
- Convergence of MCMC chains must be diagnosed, adding an extra validation step compared to classical estimation.
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 do I choose priors for the AR and MA coefficients?
A common approach is to use normal priors centred at zero with moderate variance, truncated or constrained to the stationarity and invertibility regions of the parameter space. If past studies provide information about plausible coefficient magnitudes, these can be encoded as informative priors. Sensitivity analysis — re-running inference with alternative priors — is recommended to assess robustness.
Is Bayesian ARIMA better than classical ARIMA?
Not universally. In large samples, Bayesian and classical ARIMA produce very similar point forecasts and the computational overhead of MCMC is rarely justified. The Bayesian approach adds genuine value when sample sizes are small, informative priors are available, or when full predictive distributions (not just point forecasts) are needed for decision-making under uncertainty.
How is the model order (p, d, q) chosen in the Bayesian framework?
Orders can be compared using the Deviance Information Criterion (DIC), WAIC, or Bayes factors. Some implementations place a prior over the model order itself and perform Bayesian model averaging across candidate ARIMA specifications, which propagates model uncertainty into forecasts.
Can Bayesian ARIMA handle seasonal data?
Yes, by extending to a Bayesian SARIMA model, which adds seasonal AR and MA polynomials and a seasonal differencing operator. This is the Bayesian analogue of the classical SARIMA(p,d,q)(P,D,Q)_s specification.
What software supports Bayesian ARIMA estimation?
Stan (via RStan or CmdStan), JAGS, PyMC, and BUGS all support custom Bayesian ARIMA models. Some R packages such as 'forecast' with Bayesian wrappers and the 'bsts' package (Bayesian structural time series) provide related functionality.
Sources
- 1.Pole, A., West, M., & Harrison, J. (1994). Applied Bayesian Forecasting and Time Series Analysis. Chapman & Hall.ISBN 978-0412416903
- 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?
Cite this page
ScholarGate. (2026, June 3). Bayesian ARIMA model. ScholarGate. https://scholargate.app/econometrics/bayesian-arima-model