Bayesian Autoregressive (AR) Model
Bayesian Autoregressive Model · Also known as: Bayesian autoregressive model, BAR model, Bayesian AR, Bayesian time-series autoregression
The Bayesian AR model estimates an autoregressive time-series process by combining a likelihood derived from the AR structure with prior distributions over the lag coefficients and error variance. Rather than producing single point estimates, it yields full posterior distributions, enabling principled uncertainty quantification and probabilistic forecasting.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use the Bayesian AR model when you have a univariate time series and want to forecast with full uncertainty quantification, when the sample is short (fewer than 60 observations) and classical estimates are unstable, or when you have genuine prior information — for example, that seasonally adjusted macro series are likely to be near-stationary. It is also appropriate when you want to formally compare models by marginal likelihood (Bayes factors). Avoid it when the series is clearly multivariate and cross-variable spillovers matter (use a Bayesian VAR instead), when the lag structure is strongly nonlinear, or when computational resources are very constrained and a simple classical AR suffices.
Strengths & limitations
- Provides full posterior distributions over parameters, yielding honest credible intervals rather than asymptotic confidence intervals.
- Performs well in small samples by borrowing strength from informative priors, unlike classical MLE which can produce unstable estimates.
- Propagates parameter uncertainty into forecasts, producing calibrated predictive intervals.
- Allows formal model comparison via marginal likelihood and Bayes factors.
- Priors can regularise coefficients toward stationarity, mitigating explosive AR estimates.
- Naturally accommodates missing observations within the posterior framework.
- Posterior inference via MCMC is computationally more demanding than closed-form OLS, especially for long series or high lag orders.
- Results depend on prior specification; poorly chosen priors can dominate the posterior with limited data.
- Assessing MCMC convergence requires additional diagnostics (trace plots, R-hat, effective sample size).
- Does not capture cross-variable dynamics; a multivariate extension (Bayesian VAR) is needed for systems of series.
- Model selection for the lag order p is more involved than classical AIC/BIC criteria.
Frequently asked
How is the Bayesian AR model different from classical AR?
Classical AR estimates coefficients by maximising the likelihood (OLS or MLE), yielding point estimates with asymptotic standard errors. The Bayesian AR treats parameters as random variables, combines the likelihood with a prior, and produces a full posterior distribution. This yields credible intervals that are valid in small samples and forecasts that account for parameter uncertainty.
What is the Minnesota prior and when should I use it?
The Minnesota prior sets the mean of each AR lag coefficient to zero (or one for the first own-lag), with variance shrinking as the lag grows. It encodes the belief that more distant lags matter less. It is a practical default for macroeconomic AR and VAR models, especially when the number of parameters is large relative to the sample.
Do I need MCMC for a Bayesian AR model?
Not always. With a Normal-Inverse-Gamma conjugate prior the posterior is available analytically and sampling is fast. MCMC (Gibbs or HMC) is needed when priors are non-conjugate — for example, priors that enforce stationarity exactly — or when the model is hierarchical.
How do I choose the lag order p in a Bayesian AR model?
You can use Bayesian information criteria computed on the posterior, compare models by marginal likelihood (Bayes factor), or place a prior directly over p and do Bayesian model averaging. In practice, marginal likelihood comparison is the most principled approach.
Can the Bayesian AR model handle non-stationary series?
A standard Bayesian AR assumes stationarity. For non-stationary series you should difference the data first (checking with a unit root test), or use a Bayesian AR model with a unit-root prior that explicitly allows for near-integrated behavior. The choice of prior near the unit circle matters significantly for small samples.
Sources
- Zellner, A. (1971). An Introduction to Bayesian Inference in Econometrics. Wiley. ISBN: 978-0471169376
- West, M., & Harrison, J. (1997). Bayesian Forecasting and Dynamic Models (2nd ed.). Springer. ISBN: 978-0387947259
How to cite this page
ScholarGate. (2026, June 3). Bayesian Autoregressive Model. ScholarGate. https://scholargate.app/en/econometrics/bayesian-ar-model
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- ARMA modelEconometrics↔ compare
- Autoregressive modelEconometrics↔ compare
- Bayesian ARIMA modelEconometrics↔ compare
- Bayesian ARMA modelEconometrics↔ compare
- Bayesian VAR modelEconometrics↔ compare
- Vector AutoregressionEconometrics↔ compare