跳到主要内容

Macro by Mark 的 Cookie.我们使用必要 Cookie 处理登录、语言和安全。经您同意,我们还会将可选 Cookie 用于保存偏好、分析,以及通过 Meta(Facebook)Pixel 衡量广告效果。请阅读我们的Cookie 政策。

Macro by MarkMacro by Mark
  • 首页
加载中…
加载中…
← 模型
Macro概念政策模型

Bayesian VAR (BVAR) Model

Bayesian shrinkage on a vector autoregression - Minnesota-style priors stabilize forecasts for medium-sized macro systems.

概述图表数学参考资料

How do you estimate a high-dimensional VAR without drowning in parameter proliferation?

Background

The VAR's strength - treating every variable as endogenous - is also its weakness. A VAR(p)\text{VAR}(p)VAR(p) with nnn variables requires n2pn^2 pn2p slope coefficients. With n=7n = 7n=7 and p=4p = 4p=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 Business & Economic Statistics 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)(j,k)(j,k) at lag lll shrinks at rate 1/l21/l^21/l2, penalizing distant lags, and cross-variable coefficients receive an additional shrinkage factor λcross<1\lambda_{\text{cross}} < 1λcross​<1. Two hyperparameters control the overall tightness (λ)(\lambda)(λ) and the relative weight on cross-variable versus own-variable lags. When λ\lambdaλ approaches ∞\infty∞ the posterior collapses to the OLS VAR. When λ\lambdaλ 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 λ\lambdaλ 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.

Extensions since the original Minnesota prior include stochastic volatility (Cogley and Sargent, 2005), time-varying parameters (Primiceri, 2005), large BVARs whose Minnesota-type shrinkage tightens as the system grows (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.

Connections to explore

Math ↗The structure and its conditions, step by step.References ↗Academic sources and guidance for using the model.

How the Parts Fit Together

β\boldsymbol{\beta}β

Stacked coefficient vector

The vec⁡(B′)\operatorname{vec}(B')vec(B′) of all n2p+nn^2 p + nn2p+n parameters in the reduced-form VAR, treated as a random vector with a prior distribution.

Inputs are identical to a standard VAR: an n×Tn \times Tn×T matrix of endogenous variables yty_tyt​ observed at uniform frequency, a chosen lag order ppp, and a deterministic component (intercept, trend, or neither). The additional ingredient is the prior specification: a prior mean vector b0b_0b0​ for the stacked coefficient vector β\betaβ, and a prior precision matrix (or its parameterization through hyperparameters). Under the Minnesota prior, the user sets overall tightness λ\lambdaλ, cross-variable shrinkage λcross\lambda_{\text{cross}}λcross​, lag decay ddd (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 BBB conditional on the innovation covariance Σ\SigmaΣ is matrix-normal; the marginal posterior for Σ\SigmaΣ is inverse-Wishart. No MCMC is needed. The posterior mean for BBB 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 one-step-ahead predictive distribution is a multivariate ttt with known location, scale, and degrees of freedom - no simulation needed at h=1h = 1h=1. At horizons h≥2h \geq 2h≥2 the iterated forecast is a nonlinear function of BBB, so no closed form exists even under conjugacy, and forecasts are simulated in practice: iterate the VAR forward for each posterior draw and collect the empirical distribution across draws. The same simulation approach is required for non-conjugate BVARs at every horizon. 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.

Assumptions

Testable

Covariance stationarity (or known nonstationarity treatment)

Eigenvalues of the companion matrix lie inside the unit circle, or unit-root behavior is absorbed by the prior (sum-of-coefficients dummy, dummy-initial-observation prior). Stationarity is not strictly required for Bayesian estimation, but the prior must be coherent with the data-generating process.

If violated:If variables are I(1) and the prior assumes stationarity, the posterior can concentrate on explosive parameter regions. The sum-of-coefficients prior of Doan, Litterman, and Sims (1984) handles this by shrinking the sum of lag coefficients toward 1.

Maintained

Correct prior family

The prior distribution is a reasonable representation of beliefs about the parameter space. For the Minnesota prior: each variable is close to a random walk, cross-variable effects are smaller than own effects, and distant lags matter less than recent ones.

If violated:A badly misspecified prior can dominate the posterior in small samples. Symptoms: marginal likelihood that is flat or multimodal in the hyperparameter space, or posterior predictive checks that fail systematically for a subset of variables.

Testable

Linearity

The conditional mean of yty_tyt​ given past information is linear in the lagged values. No regime switches, threshold effects, or asymmetric responses.

If violated:Linear BVARs miss recession-era dynamics where responses are asymmetric. Threshold BVAR or Markov-switching BVAR extends the framework at the cost of much heavier computation.

Testable

Gaussian innovations (for conjugate inference)

ut∼N(0,Σ)u_t \sim N(0, \Sigma)ut​∼N(0,Σ). Required for the normal-inverse-Wishart conjugate posterior to be exact. Not required for MCMC-based BVARs with flexible error distributions.

If violated:Heavy tails or skewness in the innovations invalidate conjugate credible intervals. Point forecasts remain reasonable but interval coverage deteriorates. Stochastic volatility or t-distributed errors are the standard fix.

Testable

No structural breaks

The coefficient matrix BBB and the covariance matrix Σ\SigmaΣ are constant over the sample period.

If violated:Time-varying parameter BVARs (Primiceri 2005) allow BtB_tBt​ and Σt\Sigma_tΣt​ to drift as random walks. Constant-parameter BVARs estimated over samples spanning the Great Moderation and post-2008 period systematically overstate volatility in the later subsample.

Testable

Correct lag order

The chosen lag order p is sufficient to capture the relevant dynamics. Underfitting omits important information; overfitting increases posterior uncertainty without improving forecasts.

If violated:Marginal likelihood comparison across p values is the Bayesian approach to lag selection. Selecting p by AIC/BIC and then conditioning on it ignores model uncertainty - Bayesian model averaging over p is the principled solution.

Applications

Central banks such as the European Central Bank use BVARs as part of their forecasting toolkit, and the Giannone-Lenza-Primiceri (2015) specification with marginal-likelihood-optimized hyperparameters is a common baseline. Variables typically 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.

Large BVARs (20 to 130 variables) became feasible once the prior was allowed to tighten as the system grows. Banbura, Giannone, and Reichlin (2010) showed that a BVAR with about 130 variables, under Minnesota shrinkage set in relation to the number of variables, matches or beats factor models for U.S. macroeconomic forecasting. Global-local priors (horseshoe, Dirichlet-Laplace) are a later route to the same scale. These large systems offer a direct alternative to dynamic factor models - the architecture behind the New York Fed's Staff Nowcast (Bok et al. 2018) - 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 about 150 variables, even efficient shrinkage priors struggle with the O(n2)O(n^2)O(n2) scaling of the covariance matrix. Factor-augmented VARs or penalized reduced-rank regression handle ultra-high-dimensional panels more gracefully.

Extensions and open questions

Variants

  • Minnesota prior BVAR: the original Litterman specification with diagonal prior covariance, random-walk prior mean, and quadratic lag decay.
  • TVP-BVAR: time-varying parameter BVAR (Primiceri 2005) where coefficients follow random walks and the covariance matrix has stochastic volatility. Estimated by MCMC.
  • Large BVAR: systems with 20-130+ variables using tighter shrinkage (Banbura et al. 2010) or global-local priors (horseshoe, Dirichlet-Laplace).
  • BVAR with stochastic volatility: constant coefficients but time-varying Σt\Sigma_tΣt​. Lighter than full TVP-BVAR; captures the Great Moderation without changing the propagation mechanism.
  • Hierarchical BVAR: hyperparameters (λ,λcross)(\lambda, \lambda_{\text{cross}})(λ,λcross​) are given their own priors and sampled jointly with the model parameters, fully integrating over prior uncertainty.

Open questions

  • How should BVARs handle the zero lower bound? When the policy rate is constrained at zero, the linear VAR dynamics misrepresent the monetary transmission mechanism. Shadow rate BVARs and censored-normal approaches exist but lack consensus.
  • Optimal shrinkage in ultra-large systems (n>200)(n > 200)(n>200): global-local priors, factor structure priors, and reduced-rank approaches compete without a clear winner. Computational scaling remains the binding constraint.
  • Prior specification for emerging economies: the Minnesota random-walk prior reflects U.S. macro dynamics. For countries with structural breaks, high inflation regimes, or short samples, the appropriate prior family is an open research question.

Keep going

Read next

Continue with a related concept, model, or example.

Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
帮助
隐私政策使用条款伦理与合规

© 2026 Mark Jayson Nation LLC