Bayesian Vector Autoregression (BVAR)
Also known as: BVAR, Bayesian vector autoregression, Minnesota prior VAR, Bayesian VAR (BVAR)
Bayesian VAR adds Minnesota or other prior distributions to a vector autoregressive model to control over-parameterisation. Introduced by Litterman (1986) and extended to high dimensions by Bańbura, Giannone and Reichlin (2010), it outperforms classical VAR on short series and high-dimensional macroeconomic forecasts.
Key highlights
- Controls over-parameterisation, so it stays stable where a classical VAR breaks down on short or high-dimensional series.
- Outperforms classical VAR for macroeconomic forecasting when many variables are involved.
- Priors encode economically sensible structure (e.g. random-walk behaviour, lag decay) and produce full posterior distributions for forecasts.
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 BVAR when you have a multivariate continuous time series and want to forecast or study dynamic relationships, especially when the data are short or the number of variables is large relative to the sample (at least about 40 observations). A prior structure such as the Minnesota or Litterman prior must be chosen, and the series should satisfy stationarity. It is most valuable in high-dimensional macroeconomic settings where a classical VAR would be too heavily parameterised.
Strengths & limitations
- Controls over-parameterisation, so it stays stable where a classical VAR breaks down on short or high-dimensional series.
- Outperforms classical VAR for macroeconomic forecasting when many variables are involved.
- Priors encode economically sensible structure (e.g. random-walk behaviour, lag decay) and produce full posterior distributions for forecasts.
- Requires choosing and tuning a prior (Minnesota/Litterman) structure, and results can depend on those hyperparameters.
- Needs the series to satisfy stationarity and assumes a multivariate time series is available.
- Needs a reasonable sample (about 40 observations or more) and is more involved to specify and estimate than a plain VAR.
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 BVAR different from a classical VAR?
A classical VAR estimates every coefficient freely by least squares, which over-parameterises the model when there are many variables and lags. BVAR adds a prior that shrinks the coefficients toward a parsimonious benchmark, trading a little bias for much greater stability and better forecasts, especially on short or high-dimensional series.
What is the Minnesota prior?
The Minnesota (Litterman) prior is the standard prior for BVAR. It centres each variable's own first lag near a random walk, shrinks other coefficients toward zero, and shrinks more aggressively at longer lags. This encodes the economically sensible belief that recent own-history matters most.
How much data do I need?
BVAR needs a multivariate continuous time series with at least about 40 observations. Much of its advantage shows up precisely when data are short or the system is large, because the prior compensates for the limited information in the sample.
When should I prefer a classical VAR?
When the system is small, the series are long, and the data are informative, a classical VAR may be adequate and simpler, since the benefit of shrinkage shrinks as the sample grows relative to the number of parameters.
Sources
- 1.Litterman, R. B. (1986). Forecasting with Bayesian Vector Autoregressions—Five Years of Experience. Journal of Business & Economic Statistics, 4(1), 25-38.
- 2.Bańbura, M., Giannone, D., & Reichlin, L. (2010). Large Bayesian Vector Auto Regressions. Journal of Applied Econometrics, 25(1), 71-92.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Bayesian VAR. ScholarGate. https://scholargate.app/econometrics/bvar