Bayesian VAR (BVAR) Model

The VAR's strength - treating every variable as endogenous - is also its weakness. A VAR(p) with n variables requires $n^2*p$ slope coefficients. With n = 7 and p = 4 the system has 196 slopes plus 7 intercepts. Quarterly macro datasets rarely exceed 200 observations, so the parameter-to-observation ratio approaches the danger zone where OLS overfits noise. By the mid-1980s practitioners at the Federal Reserve Bank of Minneapolis had hit this wall repeatedly. Robert Litterman's 1986 Journal of Forecasting paper proposed the solution that became the standard: treat the VAR coefficients as random variables and impose a prior distribution that shrinks most of them toward zero. The resulting Bayesian VAR (BVAR) traded a small amount of bias for a large reduction in forecast variance.

The Minnesota prior - named for the Minneapolis Fed where Litterman, Thomas Doan, and Christopher Sims developed it - encodes a specific belief: each variable behaves roughly like a random walk with drift, and cross-variable lags are less important than own lags. Formally, the prior mean for a variable's own first lag is 1 (unit root belief); all other lag coefficients have prior mean 0. The prior variance on coefficient (j,k) at lag l shrinks at rate 1/l^2, penalizing distant lags, and cross-variable coefficients receive an additional shrinkage factor lambda_cross < 1. Two hyperparameters control the overall tightness (lambda) and the relative weight on cross-variable versus own-variable lags. When lambda approaches infinity the posterior collapses to the OLS VAR. When lambda approaches zero every variable becomes an independent random walk.

Giannone, Lenza, and Primiceri (2015, Review of Economics and Statistics) showed how to choose the Minnesota hyperparameters by marginal likelihood maximization, removing the last subjective element. Their approach treats lambda as a continuous hyperparameter and optimizes the marginal data density over a grid. The result: BVAR forecasts that are competitive with or better than large factor models, mixed-frequency models, and professional consensus forecasts for GDP, inflation, and unemployment at horizons of 1-8 quarters. The ECB's suite of BVARs follows this approach almost exactly.

Extensions since the original Minnesota prior include stochastic volatility (Cogley and Sargent, 2005), time-varying parameters (Primiceri, 2005), large BVARs with global-local shrinkage (Banbura, Giannone, and Reichlin, 2010), and hierarchical priors that learn the prior tightness from the data jointly with the coefficients. The BVAR has become the default forecasting tool at central banks worldwide - not because it always wins point-forecast accuracy contests, but because it consistently avoids the catastrophic forecast failures that plague unrestricted VARs in moderate-to-large systems.

Inputs are identical to a standard VAR: an n x T matrix of endogenous variables $y_t$ observed at uniform frequency, a chosen lag order p, and a deterministic component (intercept, trend, or neither). The additional ingredient is the prior specification: a prior mean vector $b_0$ for the stacked coefficient vector beta, and a prior precision matrix (or its parameterization through hyperparameters). Under the Minnesota prior, the user sets overall tightness lambda, cross-variable shrinkage lambda_cross, lag decay d (usually 2), and optionally a sum-of-coefficients or dummy-initial-observation prior to encode beliefs about unit roots and cointegration.

Estimation under the natural conjugate prior (normal-inverse-Wishart) is analytic. The posterior for the coefficient matrix B conditional on the innovation covariance Sigma is matrix-normal; the marginal posterior for Sigma is inverse-Wishart. No MCMC is needed. The posterior mean for B is a matrix-weighted average of the OLS estimate and the prior mean, where the weight on the prior increases with the prior precision relative to the data precision. With a non-conjugate prior (asymmetric shrinkage, stochastic volatility, or heavy-tailed innovations), Gibbs sampling or Hamiltonian Monte Carlo replaces the closed-form solution.

Forecasting uses the posterior predictive distribution. For conjugate BVARs the h-step-ahead forecast distribution is a matrix-t with known location, scale, and degrees of freedom - no simulation needed for point forecasts and intervals. For non-conjugate BVARs, forecasts are generated by iterating the VAR forward for each posterior draw and collecting the empirical distribution across draws. Structural identification (Cholesky, sign restrictions, proxy instruments) applies to BVARs exactly as it does to frequentist VARs, but posterior uncertainty about the reduced-form parameters propagates naturally into the identified impulse responses.

The European Central Bank maintains a suite of BVARs as part of its forecasting infrastructure. The Giannone-Lenza-Primiceri (2015) specification with marginal-likelihood-optimized hyperparameters is the baseline. Variables include real GDP growth, HICP inflation, the short-term interest rate, unemployment, credit growth, and the exchange rate. Staff produce density forecasts at horizons of 1-8 quarters and compare BVAR output against the DSGE model (NAWM II) and the broad macroeconomic projection exercise. BVARs consistently rank first or second in real-time forecast accuracy evaluations for euro area GDP and inflation.

Large BVARs (n = 20-130 variables) have become feasible through global-local shrinkage priors. Banbura, Giannone, and Reichlin (2010) showed that a BVAR with 130 variables and appropriate Minnesota shrinkage matches or beats factor models for U.S. macroeconomic forecasting. The Federal Reserve Bank of New York's nowcasting model incorporates BVAR components. The Bank of Canada's LENS model (Large Empirical and Semi-structural) uses a BVAR core with 40+ variables. These large systems offer a direct alternative to dynamic factor models when the user wants structural interpretability alongside forecast accuracy.

Structural identification in BVARs follows the same logic as in frequentist VARs - Cholesky, sign restrictions, narrative identification, proxy instruments - but with a key advantage: parameter uncertainty propagates into the identified impulse responses through the posterior. In a frequentist SVAR, confidence bands around IRFs are computed by delta method or bootstrap, both of which have known coverage problems in small samples. In a BVAR, credible bands from the posterior are typically better calibrated. Rubio-Ramirez, Waggoner, and Zha (2010) provide the algorithmic machinery for sign-restricted identification in BVARs.

BVARs break down when the assumption of linearity and time-invariant parameters is seriously violated. Recessions involve nonlinear dynamics - credit crunches, zero lower bound constraints, financial accelerator effects - that a linear BVAR cannot represent. TVP-BVARs partially address this but at enormous computational cost and with new identification challenges. For systems with more than ~150 variables, even efficient shrinkage priors struggle with the $O(n^2)$ scaling of the covariance matrix. Factor-augmented VARs or penalized reduced-rank regression handle ultra-high-dimensional panels more gracefully.