GVARResult.generalized_fevd#
- GVARResult.generalized_fevd(horizon=20)[source]#
Variance shares from the generalized responses, normalized to sum to one.
The counterpart of
fevd()for systems with no defensible recursive ordering. It carries one caveat that the orthogonalized version does not, and the caveat is structural rather than cosmetic: generalized shocks are not orthogonal to each other, so the raw contributions do not add up to the forecast error variance. Dividing by their sum forces rows to one, which makes the table readable and makes the numbers a relative ranking rather than a decomposition.Read a cell as “how much of this variable’s forecast error is associated with that shock, relative to the others”, not as “how much is caused by it”. When the shocks are close to orthogonal the two readings coincide; when they are strongly correlated the normalization is doing real work and the shares should be treated as indicative.