Bayesian SARIMA Model
Bayesian Seasonal Autoregressive Integrated Moving Average Model · Also known as: Bayesian SARIMA, Bayesian seasonal ARIMA, BSARIMA, Bayesian seasonal time-series model
The Bayesian SARIMA model combines the classical Box-Jenkins Seasonal ARIMA framework with Bayesian inference to handle seasonal time-series data. Rather than producing a single point estimate, it yields a full posterior distribution over model parameters, propagating parameter uncertainty directly into forecasts and enabling principled incorporation of prior knowledge.
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 Bayesian SARIMA when you have a univariate time series with clear seasonal periodicity (e.g., monthly retail sales, quarterly GDP, daily electricity demand) and want full predictive distributions rather than point forecasts. It is especially valuable with short series — where classical maximum-likelihood SARIMA overestimates precision — or when domain knowledge can be encoded in priors to stabilise estimates. It is also preferred when comparing competing SARIMA orders via marginal likelihoods (Bayes factors). Avoid it when the series is non-seasonal (plain Bayesian ARIMA suffices), when computational resources are severely limited (MCMC is slower than OLS-style fitting), or when the practitioner cannot commit to a reasonable prior and the sample is large enough that priors become irrelevant anyway.
Strengths & limitations
- Full posterior predictive intervals that correctly propagate both parameter uncertainty and observation noise, avoiding overconfident forecasts.
- Principled incorporation of prior knowledge about seasonal strength, trend direction, or parameter ranges.
- Model comparison via marginal likelihoods or WAIC, making it straightforward to choose among competing SARIMA orders.
- Works well with short series where maximum-likelihood estimates are unreliable.
- Naturally handles missing observations through data augmentation within the MCMC scheme.
- Computationally intensive: MCMC sampling can take minutes to hours for long series or complex seasonal structures.
- Results are sensitive to prior choices, especially with short series; poorly chosen priors can dominate the posterior.
- Requires stationarity checks and appropriate differencing before modelling, just as classical SARIMA does.
- Model order selection (p, d, q, P, D, Q) still requires diagnostic effort and cannot be fully automated.
- Interpretation of the posterior is more involved than reading a single coefficient table.
Frequently asked
How do I choose the prior for the AR and MA coefficients?
A common default is a normal prior centred at zero with moderate variance (e.g., N(0, 0.5²)) combined with stationarity and invertibility constraints enforced by rejection sampling or parameter transformations. If domain knowledge suggests persistent seasonality, a prior slightly above zero for seasonal AR coefficients is reasonable. Sensitivity analysis — rerunning with tighter and looser priors — should be reported.
How is the seasonal order s determined?
The period s is fixed by the data-collection frequency: s = 12 for monthly data, s = 4 for quarterly, s = 7 for daily data with weekly seasonality. It is not estimated from the data but set based on the known periodicity. Confirm it visually with a seasonal subseries plot or by inspecting ACF spikes at multiples of s.
Can I use Bayesian SARIMA for multi-step forecasting?
Yes. Each posterior draw of the parameters generates a full trajectory of future values. The collection of trajectories constitutes the joint predictive distribution, from which any horizon's marginal distribution can be read off. This is more honest than classical plug-in forecasting because it reflects parameter uncertainty in every step.
What software can fit a Bayesian SARIMA model?
Stan (via RStan or PyStan) and JAGS are the most flexible options for custom priors. The bsts package in R implements Bayesian structural time-series models that subsume SARIMA structures. Python's PyMC and NumPyro also support custom ARIMA-like specifications within a Bayesian framework.
When should I prefer Bayesian SARIMA over classical SARIMA?
Prefer the Bayesian version when the series is short (fewer than ~60 observations), when you have genuine prior knowledge to incorporate, when you need full predictive distributions rather than symmetric normal approximations, or when comparing model orders formally via Bayes factors. For very long series with no informative priors, classical maximum-likelihood SARIMA is faster and produces nearly identical point forecasts.
Sources
- 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
- Geweke, J., & Whiteman, C. (2006). Bayesian forecasting. In G. Elliott, C. W. J. Granger, & A. Timmermann (Eds.), Handbook of Economic Forecasting (Vol. 1, pp. 3–80). Elsevier. link ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Seasonal Autoregressive Integrated Moving Average Model. ScholarGate. https://scholargate.app/en/econometrics/bayesian-sarima-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.
- ARIMA modelEconometrics↔ compare
- Bayesian VAR modelEconometrics↔ compare
- SARIMA modelEconometrics↔ compare
- State Space ModelEconometrics↔ compare