Statistics Glossary → Forecast error variance decomposition (FEVD)
Forecast error variance decomposition (FEVD)concept
A decomposition in a VAR model showing what share of the error variance of a variable's h-step-ahead forecast comes from its own shocks and what share from shocks to the other variables.
In a VAR model each variable depends on its own lags and the lags of the others, and the model's shocks spread through the whole system over time. The error of an h-step-ahead forecast is the sum of the shocks that occur in the intervening periods. FEVD splits the variance of this error by source: for example, 60% of the 12-month-ahead forecast error of inflation might come from its own shocks, 30% from interest-rate shocks and 10% from exchange-rate shocks. For each variable and horizon the shares add up to 100%.
Because shocks are usually correlated, they are orthogonalized before decomposition; the most common way is the Cholesky decomposition, and results obtained this way depend on the order of the variables in the model: a shock to a variable earlier in the order can affect the others in the same period, but not the reverse. The ordering should be chosen on theoretical grounds, or the order-invariant generalized FEVD used. FEVD is derived from the same information as the impulse response functions and is not evidence of causation; it only summarizes the dynamic relationships within the model.
An everyday example
A café chain tracks coffee sales, advertising spend and temperature together in a VAR model. Of the uncertainty in next month's sales forecast, 70% comes from unexpected swings in sales themselves, 20% from advertising shocks and 10% from temperature shocks; six months ahead, advertising's share rises to 35%. This suggests the effect of advertising builds up over time.
In MerQur
Part of the VAR analysis output, given as the variance decomposition (FEVD) alongside the impulse response functions.