=======================
Question
So what about my last question? Are IRF confidence intervals based on
bootstrapping valid even if the VECM has non-normal residuals?

Answer

The bootstrap was developed to avoid Monte Carlo simulations using random draws from a preset distribution (say Normal). The bootstrap will proxy the empirical distribution of your residuals given that the residuals are "sufficiently" noisy (it doesn't matter if they test non-normal!)

By noisy, I mean no heteroscedasticity or autocorrelation issues. There is a huge literature for heteroscedasticity issues (Wild bootstrap) and I guess (i am not an expert) block bootstrap that divides your data into blocks to preserve the original time series properties within a block (when dependence is an issue).

 

Back to your question, it seems it has to do with what you mean "non-normal" residuals? My guess is a few outliers or asymmetry? So bootstrap is valid.

 

=======================

Question d.o.f correction

 

Regarding d.o.f corrections when calculating IRFs in gretl

 

Using Allin's example

 

open data9-7
var 4 UNEMP PRIME

 

For the UNEMP equation the Sum squared resid=3.613493 so that sqrt(3.613493/60)=0.24540759

In other commercial software this is baptized : no df correction.

If you prefer df correction multiply IRFs with 60/(60-1-8) where 1+8 corresponds to the parameters number in one VAR equation sqrt(3.613493/(60-9))=0.2661819