Bayesian Estimation Questions

General DSGE Modeling
1

Hi everyone,

I’ve shortened my earlier questions for clarity and would really appreciate your input:

  • I have 5 shocks and 5 observables, and I’ve introduced a correlation between two shocks. Does this correlation count as an additional shock for estimation, allowing me to add a 6th observable?
  • Historical decomposition depends on the chosen observables. If the number of shocks limits the observables I can include, how can I ensure the results are robust? I’ve noticed that while the decomposition looks strong for included observables, it often appears weaker for excluded ones.
  • My dataset spans 1980Q1–2024Q2, with a major shock (e.g., COVID-19) assumed in 2019Q4. I’m considering using the heteroskedastic_shocks block to account for time-varying shock variances. Does this approach make the model deterministic (like perfect foresight), and how does it differ from standard stochastic models? Also, how are periods defined under heteroskedastic_shocks—do they align with the presample, estimation, or another timeline? If I mix heteroskedastic and non-heteroskedastic shocks, how does the heteroskedastic_filter option affect estimation?
  • Finally, I plan to treat 1980–2019Q3 as presample data and 2019Q4–2024Q2 for likelihood calculation, assuming the economic structure differs during the shock period. Is this approach reasonable? Do DSGE models typically include data from shock periods in estimation, or are there cases where excluding such periods provides better results? Or is this assumption subjective?

Thanks in advance for your help!

Best,
Yamoi

2
  1. No, correlation between shocks does not mean there is effectively a 6th shock.
  2. See e.g.

3. The heteroskedastic_shocks allows for unforeseen structural breaks in the shock variances. It does not make the model deterministic. It simply allows you to have periods where bigger shocks are considered to be more likely.
4. That assumption is always subjective. After all, we don’t know which process generated the data.