Prof @Pfeiffer, I’m running a small open economy with households, heterogeneous firms, monetary and fiscal policy to study the pass-through transmissions on EME countries.
The Bayesian estimation is ok computing with option 9 and based on quarterly data (DATABfiltro2), I got an acceptable convergence, but when I check the IRFs, for example, the shock ‘wbar’ that measures the UIRP effect (uncovered interest rate parity deviations) on the nominal exchange variations (variable “piis”) i get the appended figure “CodEconModel_Bayesian_IRF_sigma_wbar_1” (check piis_obs), where the shock is -0.2% (revaluated exchange) in the first period and then jump to 0.2% in the 5th quarterly (depreciated exchange), but this look unusual from an economic point of view or from the usual IRF behavior. This kind of jump for 1st period to 5th period happened in other variables too.
Then, i got the IRF from the stochastic simulation and i see a different dynamic (check “CodEconModel_IRF_sigma_wbar”), some more realistic. The original data has been prefilter by HP before passed to the estimation.
I’d like to have a smoother IRF behavior, but i don’t know if this is happening by a seasonal behavior latent in the observable variables or something else or another thing I haven’t seen. I’d appreciate some recommendation to check. I append the code and data, it was run with dynare 6.2 on matlab 2024b.
Thank you so much for any suggestion,
jcsc,
CodEconModel.mod (14.9 KB)
CodEconModel_Bayesian_IRF_sigma_wbar_1.eps (106.0 KB)
CodEconModel_IRF_sigma_wbar.eps (29.8 KB)
DATABfiltro2.xlsx (40.5 KB)
If there is such a discrepancy between the IRFs in stoch_simul and after estimation, it is most likely the aggregation over the parameter draws that leads to this. I am nevertheless puzzled by the tight bands around piis_obs. Do you have an explanation for such a weird behavior?
Prof. Pfeiffer, Thanks for your response. As you said, effectively those IRF behaviours are so weird but i think the source of error could be on the large variations generated in some observable variables used in the Bayesian estimation, and the significant cross correlations that i observed among some variables until lag 5, for example. Considering that, I did a new review of the model but now taking the quarter over quarter variation of the observables to get a smoother version of the data and passed the filtered observables to the bayesian estimation (DSGE, DSGE-VAR1). I get a better result than before and the IRF (check figure) look better than previous, but the bayesian estimation required me more observables and more simulations (more than 500 thousand) to get better log kernel, but still some posteriors aren´t smoother and some parameter densities are quase-bi-modal (using option 6, i append some figures) and the convergence is more difficult (i append figure mdiag). Considering the data works better in this case, I appreciate any suggest in this point to get a better convergence, thank u so much!.
I append some figures after running the model with the observables in quater-over-quarter variation.
dsge_pass_through_bayes_op17_Bayesian_IRF_sigma_wbar_1.eps (134.8 KB)
dsge_pass_through_bayes_op17_mdiag.eps (35.9 KB)
dsge_pass_through_bayes_op17_PriorsAndPosteriors1.eps (249.7 KB)
dsge_pass_through_bayes_op17_PriorsAndPosteriors2.eps (227.1 KB)
dsge_pass_through_bayes_op17_PriorsAndPosteriors3.eps (201.7 KB)
dsge_pass_through_bayes_op17_PriorsAndPosteriors4.eps (213.4 KB)
dsge_pass_through_bayes_op17_PriorsAndPosteriors5.eps (142.5 KB)
I would suggest to rerun a different mode-finder like 5 starting from your current mode. You may need many more draws in any case.