Dear all, I have a problem with IRF’s generated by stock_simul command after Bayesian estimation. Bayesian IRFs are all right but the simulation IRFs using posterior modes are explosive. I tried pruning bu they become wiggly. What might be the problem?

Is it safe to use bayesian irfs? and what is the difference between the two?

wiggly irfs.eps (35.7 KB)

bayesian irfs.pdf (10.1 KB) modecheck2.pdf (8.0 KB)

modecheck1.pdf (15.5 KB) posterior2.pdf (66.4 KB)

posterior1.pdf (68.8 KB)

Why do you need pruning? Your model is estimated at first order?

Thank you very much Prof. Pfeifer. After estimation I tried to build regular IRFs with posterior mean. In forum I found an entry about trying pruning. That’s why I tried that comment. I managed to solve the problem. I just worked the stoch_simul at order 1 and the IRFs were clear. But is there a reason why it was wiggly at stock_simu order 2?

Another question I would like to ask is why some papers report Bayesian IRFs where others report regular IRFs with posterior means? How should I decide which one to use?

At `order>1`

, IRFs are simulation-based GIRFs. For them to look smooth you need a lot of replications.

Regarding the IRFs, it is a matter of taste. Bayesian IRFs are pointwise IRFs. As such, they are a consistent summary of the posterior distribution but harder to interpret, because they do not correspond to a particular parameter vector. In contrast, IRFs at the posterior mean correspond to one particular parameter vector, but do not take parameter uncertainty into account.

1 Like