Saltar al contenido principal

Cookies en Macro by Mark.Usamos cookies esenciales para inicio de sesión, idioma y seguridad. Con tu permiso, también usamos cookies opcionales de preferencias, para analítica y para medición publicitaria con el píxel de Meta (Facebook). Lee nuestra política de cookies.

Macro by MarkMacro by Mark
  • Inicio
Cargando…
Cargando…
← Modelos
MacroConceptosPolíticaModelos

Structural VAR (SVAR) Model

Identified vector autoregression that turns reduced-form residuals into economically named structural shocks.

ResumenGráficoMatemáticasReferencias

How do you recover causal structural shocks from a reduced-form VAR when the covariance matrix alone can't tell them apart?

Background

A reduced-form VAR estimates a set of forecast-error covariance matrices and moving-average coefficients, but those objects are silent about causation. The residual vector utu_tut​ bundles every contemporaneous structural disturbance into a single composite innovation for each variable. Sims (1980) proposed the VAR framework to let the data speak, but extracting economic stories - a monetary policy shock, an oil supply disruption, a technology impulse - requires additional restrictions that map the reduced-form residuals utu_tut​ into orthogonal structural shocks εt\varepsilon_tεt​. That mapping is the identification problem. This model extends the reduced-form VAR, already documented in the VAR page of this platform.

The core issue is algebraic. A VAR with nnn variables has a reduced-form covariance matrix Σu\Sigma_uΣu​ with n(n+1)/2n(n+1)/2n(n+1)/2 unique elements. The structural impact matrix B0−1B_0^{-1}B0−1​ that maps εt\varepsilon_tεt​ into utu_tut​ via ut=B0−1εtu_t = B_0^{-1} \varepsilon_tut​=B0−1​εt​ has n2n^2n2 free parameters. The orthogonality assumption E[εtεt′]=InE[\varepsilon_t \varepsilon_t'] = I_nE[εt​εt​′]=In​ delivers Σu=B0−1B0−1′\Sigma_u = B_0^{-1} B_0^{-1\prime}Σu​=B0−1​B0−1′​, which supplies n(n+1)/2n(n+1)/2n(n+1)/2 equations for n2n^2n2 unknowns. The gap is n(n−1)/2n(n-1)/2n(n−1)/2 restrictions that must come from economic theory, institutional knowledge, or external data. Every identification scheme - recursive ordering, sign restrictions, narrative instruments, long-run neutrality, heteroskedasticity-based approaches - fills exactly this gap.

Recursive identification (Cholesky decomposition of Σu\Sigma_uΣu​) imposes a causal ordering: variables ordered earlier cannot respond contemporaneously to shocks hitting variables ordered later. Sims (1980) and Christiano, Eichenbaum, and Evans (1999) used recursive structures to study monetary policy. Sign restrictions (Uhlig 2005; Faust 1998) replace zero restrictions with inequality constraints on impulse-response signs over specified horizons, accepting set identification rather than point identification. External instruments / proxy SVAR (Stock and Watson 2012; Mertens and Ravn 2013) use an outside variable correlated with one structural shock but uncorrelated with the others, recovering the column of B0−1B_0^{-1}B0−1​ associated with that shock via instrumental variables on the reduced-form residuals.

Central banks and academic researchers lean on SVAR identification daily. Kilian (2009, American Economic Review) used a recursive SVAR in global crude oil production, global real economic activity, and the real price of oil to separate oil supply shocks from aggregate-demand and oil-specific-demand shocks. ECB staff research uses sign restrictions to study euro-area monetary transmission (for example, Jarociński and Karadi 2020, who combine them with high-frequency surprises to separate policy shocks from central-bank information shocks). Academic work on fiscal multipliers (Blanchard and Perotti 2002) uses institutional timing to identify government spending shocks. Proxy SVARs have become a standard way to identify monetary policy shocks after Gertler and Karadi (2015) demonstrated the approach using high-frequency futures surprises as instruments.

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

B0−1B_0^{-1}B0−1​

Structural impact matrix

Maps the vector of orthogonal structural shocks εt\varepsilon_tεt​ into the reduced-form residuals utu_tut​ via ut=B0−1εtu_t = B_0^{-1} \varepsilon_tut​=B0−1​εt​. Each column is the instantaneous impulse vector of one structural shock.

The starting point is an estimated reduced-form VAR: yt=c+A1yt−1+⋯+Apyt−p+uty_t = c + A_1 y_{t-1} + \cdots + A_p y_{t-p} + u_tyt​=c+A1​yt−1​+⋯+Ap​yt−p​+ut​ with E[utut′]=ΣuE[u_t u_t'] = \Sigma_uE[ut​ut′​]=Σu​. Everything about SVAR identification concerns the relationship ut=B0−1εtu_t = B_0^{-1} \varepsilon_tut​=B0−1​εt​, where εt\varepsilon_tεt​ is a vector of mutually uncorrelated structural shocks with unit variance. The structural impact matrix B0−1B_0^{-1}B0−1​ (sometimes written A0−1A_0^{-1}A0−1​ or SSS) maps structural shocks into the observable residuals. Each column of B0−1B_0^{-1}B0−1​ gives the instantaneous response of every variable to one structural shock.

The identification scheme determines which B0−1B_0^{-1}B0−1​ (or which set of matrices) is admissible. Recursive identification sets B0−1B_0^{-1}B0−1​ equal to the lower-triangular Cholesky factor of Σu\Sigma_uΣu​, imposing n(n−1)/2n(n-1)/2n(n−1)/2 zero restrictions on the upper triangle. Sign restrictions constrain the signs of selected entries of the impulse-response function Θh=ΦhB0−1\Theta_h = \Phi_h B_0^{-1}Θh​=Φh​B0−1​ at horizons h=0,1,…,Hh = 0, 1, \ldots, Hh=0,1,…,H, where Φh\Phi_hΦh​ is the reduced-form MA coefficient at horizon hhh. Because sign restrictions typically do not pin down a unique B0−1B_0^{-1}B0−1​, the identified set is characterized by rotating through all orthogonal matrices QQQ such that B0−1=P~QB_0^{-1} = \tilde{P} QB0−1​=P~Q satisfies the sign constraints, where P~\tilde{P}P~ is the Cholesky factor of Σu\Sigma_uΣu​. Proxy SVAR uses an external instrument ztz_tzt​ satisfying relevance E[ztε1t]≠0E[z_t \varepsilon_{1t}] \neq 0E[zt​ε1t​]=0 and exogeneity E[ztεjt]=0E[z_t \varepsilon_{jt}] = 0E[zt​εjt​]=0 for j≠1j \neq 1j=1, and recovers the first column of B0−1B_0^{-1}B0−1​ up to scale via two-stage least squares on the reduced-form residuals.

Structural impulse-response functions are the primary output. The response of variable iii to structural shock jjj at horizon hhh is θij,h=ei′ΦhB0−1ej\theta_{ij,h} = e_i' \Phi_h B_0^{-1} e_jθij,h​=ei′​Φh​B0−1​ej​. Forecast error variance decompositions attribute the hhh-step-ahead forecast error of each variable to each structural shock. Historical decompositions allocate the realized path of each variable to the cumulative contribution of each structural shock over the sample period. Confidence intervals come from bootstrap methods - the recursive-design bootstrap and the wild bootstrap are standard for recursive SVAR; the Bayesian approach of drawing from the posterior over QQQ is standard for sign restrictions.

Assumptions

Maintained

Invertibility of B0B_0B0​

The structural impact matrix B0B_0B0​ is non-singular, so the mapping from structural shocks to reduced-form residuals is bijective.

If violated:A singular B0B_0B0​ means some structural shocks are linearly dependent and cannot be separately identified.

Maintained

Orthogonality of structural shocks

E[εtεt′]=InE[\varepsilon_t \varepsilon_t'] = I_nE[εt​εt​′]=In​. Structural shocks are mutually uncorrelated with unit variance.

If violated:If structural shocks are correlated, the decomposition Σu=B0−1B0−1′\Sigma_u = B_0^{-1} B_0^{-1\prime}Σu​=B0−1​B0−1′​ does not hold and impulse responses conflate multiple shocks.

Testable

Correct reduced-form specification

The underlying VAR is correctly specified in lag length, deterministic components, and included variables. Residuals utu_tut​ are serially uncorrelated.

If violated:Omitted variables or wrong lag length contaminate the reduced-form residuals, and structural identification inherits the misspecification.

Maintained

Identification restrictions are valid

The exclusion, sign, or instrument restrictions used to pin down B0−1B_0^{-1}B0−1​ are economically justified and correctly imposed.

If violated:Wrong restrictions recover shocks that do not correspond to any economically meaningful disturbance. Impulse responses and variance decompositions are uninterpretable.

Testable

Instrument relevance (proxy SVAR)

The external instrument ztz_tzt​ has non-zero correlation with the target structural shock: E[ztε1t]≠0E[z_t \varepsilon_{1t}] \neq 0E[zt​ε1t​]=0.

If violated:Weak instruments produce imprecise and biased estimates of the structural impact vector, analogous to the weak-instrument problem in standard IV.

Maintained

Instrument exogeneity (proxy SVAR)

The external instrument is uncorrelated with all non-target structural shocks: E[ztεjt]=0E[z_t \varepsilon_{jt}] = 0E[zt​εjt​]=0 for j≠1j \neq 1j=1.

If violated:Contaminated instruments mix the target shock with other shocks, and the recovered impulse responses reflect a linear combination of multiple structural disturbances.

Applications

The Federal Reserve system uses SVAR identification extensively. Kilian (2009) decomposed oil-price fluctuations into supply, aggregate-demand, and oil-specific-demand shocks using a recursive SVAR with production-ordering restrictions, producing the most cited oil-shock decomposition in the literature. Gertler and Karadi (2015) pioneered the proxy SVAR for monetary policy using high-frequency fed-funds-futures surprises around FOMC announcements as external instruments, and this design has become the benchmark for monetary policy identification at the Fed, ECB, and Bank of England. Blanchard and Perotti (2002) exploited institutional features of the US tax and transfer system to identify fiscal shocks, establishing the timing-restriction approach that underpins most fiscal-multiplier estimates used in policy analysis.

Sign-restricted SVARs are standard at institutions that need to remain agnostic about specific causal orderings. The ECB's structural analysis uses Bayesian sign-restricted VARs to decompose euro-area GDP into demand, supply, and monetary-policy contributions without committing to a particular Cholesky ordering. The Bank of Canada and Reserve Bank of Australia publish scenario analyses built on sign-identified technology and demand shocks. Researchers use sign restrictions when the economic theory provides directional predictions (an expansionary monetary shock raises output and lowers the interest rate) but not zero-impact restrictions.

SVAR breaks down when the number of variables is large relative to the sample, because the reduced-form VAR itself becomes unreliable. For systems with more than 8-10 variables, factor-augmented VARs (FAVAR) or Bayesian VARs with Minnesota-type priors are preferred. The method also struggles when structural breaks or time-varying parameters are present - the fixed-coefficient assumption means the identified shocks blend pre- and post-break behavior. If the researcher suspects regime-dependent transmission, a Markov-switching SVAR or TVP-VAR is more appropriate.

The choice between identification schemes depends on what the researcher is willing to assume. Recursive ordering is clean and delivers point identification but requires a defensible causal ordering - a strong demand if the timing of contemporaneous interactions is genuinely ambiguous. Sign restrictions avoid ordering assumptions but deliver set identification, meaning the impulse responses are intervals rather than point estimates, and these intervals can be wide. Proxy SVAR identifies a single shock without restricting the rest of the system but requires an external variable that satisfies both relevance and exogeneity, which is hard to find outside of financial-market-surprise settings.

Extensions and open questions

Variants

  • Cholesky / recursive SVAR (Sims 1980)
  • Long-run restricted SVAR (Blanchard and Quah 1989)
  • Sign-restricted SVAR (Uhlig 2005)
  • Proxy SVAR / IV-SVAR (Stock and Watson 2012; Mertens and Ravn 2013)
  • Heteroskedasticity-identified SVAR (Rigobon 2003; Lewis 2021)
  • Narrative sign restrictions (Antolin-Diaz and Rubio-Ramirez 2018)

Open questions

  • How to combine sign restrictions with proxy instruments in a coherent framework (Arias, Rubio-Ramirez, and Waggoner 2021)
  • Whether uniform priors over Q are inadvertently informative about structural parameters (Baumeister and Hamilton 2015)
  • Weak-instrument diagnostics tailored to the proxy SVAR setting (Montiel Olea, Stock, and Watson 2021)

Keep going

Read next

Continue with a related concept, model, or example.

DSGE model

Financial Accelerator (BGG)

a contrasting take on Structural VAR

Contrast it →
Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
Ayuda
Política de privacidadTérminos de usoÉtica y cumplimiento

© 2026 Mark Jayson Nation LLC