MerQuR+Statistics Glossary

Statistics Glossary → Impulse response function (IRF) and Cholesky decomposition

Impulse response function (IRF) and Cholesky decompositionconcept

In a vector autoregression (VAR), the function showing how a one-time shock to one variable affects all variables in the system in later periods, and the method used to separate the shocks from each other.

Once a VAR is estimated, its individual coefficients are hard to interpret, so the system's dynamics are summarised with impulse responses. The IRF at horizon h is the change in variable j h periods after a shock to variable i. With a unit shock, each variable is shocked by 1 unit. But VAR residuals are usually correlated: when one variable is shocked, the others are shocked at the same time too. The shocks are therefore orthogonalised. The Cholesky decomposition factors the residual covariance matrix into a lower-triangular matrix and defines mutually independent shocks of 1 standard deviation.

The Cholesky decomposition depends on the ordering of the variables. The first variable can affect the others within the same period but is not affected by them within that period; later variables respond to the same-period shocks of earlier ones. This ordering is a theoretical assumption, and different orderings can give different IRFs, with larger differences when the residuals are highly correlated. The uncertainty of IRFs is usually shown with bootstrap confidence bands; if the band includes zero, the response is statistically uncertain. In a stationary VAR the responses die out over time.

An everyday example

A central bank studies the interest rate, inflation and growth with a VAR. After a 1-standard-deviation rise in the interest rate, the impulse response shows inflation starting to fall after 2-3 quarters, reaching its lowest point in quarter 6 and then slowly returning to its previous level. When the variable ordering is changed the size of the response differs somewhat, so the result is reported together with its sensitivity to ordering.

In MerQur

In VAR analysis, when the 'Impulse-response type: Orthogonal (Cholesky)' checkbox is ticked a Cholesky-orthogonalised 1 SD shock is used, otherwise a unit shock. 'IRF bootstrap replications and Random seed' produce a percentile band by residual bootstrap; 0 means no band.

Analyses where this concept appears

VAR (Vector Autoregression)

Related concepts

Forecast error variance decomposition (FEVD)