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。
Bayesian shrinkage on a vector autoregression - Minnesota-style priors stabilize forecasts for medium-sized macro systems.
Graph workspace
The defining feature of BVAR is the bias-variance tradeoff controlled by the shrinkage hyperparameter . Showing how IRFs and forecast intervals change as moves from 0 (pure prior) to (pure OLS) is the most direct way to build intuition for what shrinkage actually does.
Posterior own-lag
The shrinkage-weighted compromise between the prior belief (a random walk, coefficient 1) and the OLS point estimate. Closer to 1 under tight shrinkage, closer to OLS under loose.
0.9000
coefficient
Posterior cross-lag
The shrinkage-weighted cross-variable coefficient. Prior mean is 0, so this is pulled toward zero more aggressively than the own lag.
0.0375
coefficient
IRF half-life
Number of quarters until the impulse response to a unit shock in variable 1 decays below 50% of its impact value, counted on the posterior system even when that lies beyond the 24 plotted quarters; blank when the response never halves (an explosive or unit-root posterior). Longer under tight shrinkage: the Minnesota prior centers the own lag on a random walk (coefficient 1), so tightening it moves the posterior toward a unit root and slows the decay.
7.00
quarters
4Q forecast interval width
Width of the 68% forecast band (plus or minus one standard deviation) for variable 1 at the one-year horizon, from variable 1's own shocks carried through the posterior-mean coefficients; parameter uncertainty is not included. Because the Minnesota prior pulls the own lag toward a random walk, tighter shrinkage makes shocks persist and WIDENS this band. In a full BVAR the predictive band also carries parameter uncertainty, which a tighter prior reduces, so there the net effect can go either way.
3.47 pp
percentage points
Overall shrinkage: 0.2
Controls how strongly the estimates are pulled toward the model's starting beliefs.
Low : the IRF flattens toward the random-walk prior, its half-life lengthens and the one-year forecast band widens, because near-random-walk dynamics let shocks accumulate. High : the IRF approaches the OLS estimate, decays faster and the band narrows.
Cross-variable shrinkage: 0.5
How much weight other variables' past values get compared with a variable's own past.
Low values flatten cross-variable IRF responses toward zero. High values (near 1) treat own and cross lags symmetrically.
OLS own-lag (var 1): 0.8
What the raw data says about how much variable 1 repeats itself, before shrinkage.
Determines the OLS IRF decay rate (dashed gray). The posterior IRF is pulled between this and the prior mean of 1.
OLS cross-lag (2 → 1): 0.15
What the raw data says about how much variable 2 affects variable 1.
Affects the shape of the IRF at intermediate horizons. Shrinkage pulls this toward the prior mean of 0.
How large the random shocks to variable 1 are, which scales forecast uncertainty.
Wider widens the one-year forecast band in proportion. The impulse responses are to a unit shock, so their shape does not change.