学术与研究资料
Sims (1980) -- Macroeconomics and Reality (Econometrica). The founding paper. Argued for unrestricted VARs as an alternative to large structural models.
Econometrica · 1980
Litterman (1986) -- Forecasting with Bayesian Vector Autoregressions (Journal of Business and Economic Statistics). Introduced the Minnesota prior for VAR shrinkage.
Journal of Business & Economic Statistics · 1986
Christiano, Eichenbaum, and Evans (1999) -- Monetary policy shocks: what have we learned and to what end? (Handbook of Macroeconomics). Canonical Cholesky-identified monetary policy VAR.
Handbook of Macroeconomics · 1999
Stock and Watson (2001) -- Vector Autoregressions (Journal of Economic Perspectives). The standard survey article. Clear exposition of the methodology and its limits.
Journal of Economic Perspectives · 2001
Kilian and Lutkepohl (2017) -- Structural Vector Autoregressive Analysis (Cambridge University Press). The modern reference on structural identification.
Cambridge University Press · 2017
New Introduction to Multiple Time Series Analysis
Lutkepohl's VAR reference text.
Springer · 2005
研究依据
证据与数据
Use stationarity checks, lag selection, residual diagnostics, and forecast evaluation before interpreting impulse responses.
校准或测量
Lag length, variable ordering, transformations, and sample window shape the result.
适用边界
- Reduced-form VARs do not identify causal shocks.
- Small samples can overfit.
- Regime changes can break stability.
使用指南
何时足够
Forecasting and conditional projections of a small set of jointly determined macro variables when the data-generating process is reasonably stable and the question does not require causal identification. Out-of-sample evaluation of GDP, inflation, and interest-rate paths is the classic use case (Sims 1980 Econometrica). The reduced form is also a valid first-pass check of whether the joint dynamics among a variable set are internally consistent with a maintained theory.
何时仅作示意
Do not use the reduced-form VAR as evidence about structural causal effects. The model is atheoretic by design; the ordering of variables in a Cholesky factorization imposes an identification assumption that must be defended separately. Impulse-response functions from an unidentified VAR describe correlations, not the response to a well-defined shock.
何时更换模型
Switch to an SVAR (empirical:svar) when causal identification is required. Switch to a BVAR (empirical:bvar) when the variable count is large enough that OLS estimation overfits and shrinkage priors are needed to stabilize the forecast. Switch to a state-space model when the panel has missing data or unobserved components that require the Kalman filter.
证伪信号
Out-of-sample forecast errors that systematically exceed a naive random-walk or AR(1) benchmark across multiple episodes and variable choices indicate the VAR's lag structure is overfit or the parameters have shifted. A rolling-window evaluation showing stable in-sample fit but deteriorating out-of-sample accuracy is a concrete warning that the model is not capturing the underlying dynamics.