A Bayesian vector autoregression is a multivariate time-series model that estimates the joint dynamics of several macro variables while imposing prior beliefs to discipline the large number of parameters. The priors shrink coefficients toward parsimonious benchmarks, taming overfitting and producing more stable forecasts and impulse responses than unrestricted VARs.
How it works
A standard VAR regresses each variable on lags of itself and every other variable; with many variables and lags the parameter count explodes and estimates become noisy. A BVAR adds a prior distribution — classically the Minnesota prior, which centres own first lags near one and shrinks distant and cross-variable lags toward zero — and combines it with the data likelihood to form a posterior. The result is regularised coefficients, coherent forecast fan charts, and identifiable structural shocks.
Why it matters now
BVARs are the workhorse behind many central-bank and sell-side conditional forecasts in 2025-2026, used to map policy-rate, energy-price, and fiscal scenarios into GDP and inflation paths when post-pandemic structural breaks make naive extrapolation unreliable.
Example
A euro-area BVAR conditioned on stabilising energy prices and a gradual ECB easing path might imply quarterly GDP growth sustaining at least 0.7%, with the posterior fan chart quantifying the probability the outturn clears that threshold — a directly bettable, model-anchored claim rather than a point guess.
Frequently asked
- What is a Bayesian VAR?
- A Bayesian VAR is a multivariate time-series model that estimates the joint dynamics of several macro variables while imposing prior beliefs to discipline a large parameter count. The priors shrink coefficients toward parsimonious benchmarks, taming overfitting and producing more stable forecasts and impulse responses than unrestricted VARs. Central banks like the ECB and the New York Fed use them for conditional forecasting.
- How does a Bayesian VAR differ from a standard VAR?
- A Bayesian VAR adds a prior distribution that regularises coefficients, while a standard VAR estimates every parameter freely by OLS. With many variables and lags an unrestricted VAR's parameter count explodes and estimates become noisy. The BVAR's prior — classically the Minnesota prior — shrinks distant and cross-variable lags toward zero, allowing larger, more stable systems that would otherwise overfit.
- What is the Minnesota prior in a Bayesian VAR?
- The Minnesota prior is the classic shrinkage scheme that centres each variable's own first lag near one and pushes distant and cross-variable lags toward zero. It encodes the belief that macro series behave roughly like random walks and that recent own-lags carry
Glossary · carry
Carry is the income earned (or paid) for holding a position over time, independent of price change — the coupon, yield, or interest-rate differential captured while waiting. Positive carry pays you to hold; negative carry costs you. It is the return that accrues if nothing moves.
Read more →
- Why do central banks use Bayesian VARs for forecasting?
- Central banks use Bayesian VARs because the priors stabilise large multivariate systems and yield coherent conditional forecasts and probabilistic fan charts. The ECB and the New York Fed both document large BVARs in their research, mapping policy-rate, energy, and fiscal scenarios into GDP and inflation paths. Post-pandemic structural breaks make naive extrapolation unreliable, and shrinkage delivers more robust out-of-sample accuracy.
- Can a Bayesian VAR produce probabilistic forecasts?
- Yes, a Bayesian VAR produces full posterior predictive distributions, not just point forecasts, so it naturally yields probabilistic fan charts. The posterior quantifies the likelihood an outturn clears a threshold — for example, the probability quarterly GDP growth exceeds 0.7%. This makes BVAR output directly comparable to prediction-market pricing and bettable claims rather than single-number guesses.