Hi,
I have a question that might sound philosophical. It concerns the treatment of scenario data in the estimation of a BVAR (Bayesian vector autoregressive model) and scenario analysis.
In econometrics, the standard approach (following Waggoner and Zha conditiional forecasting in VARs, 1999, https://tzha.net/wp-content/uploads/2025/07/RESTAT_WZ_1999.pdf”) is that estimation of the parameters of the model and construction of the posterior of the scenario MUST BE DONE JOINTLY.
My understanding is that in this way we are treating the scenario as additional information (data) for the estimation of the parameters of the model. The scenario tilts the posterior of the parameters of the model.
While I can see why this approach has merits, I struggle to see why this is the only correct way to construct scenarios.
What I find more intuitive is to separate the estimation from the “intervention” analysis. As if I want to explore particular paths of the variables taking as fixed the model that generates those paths.
Is the second approach (the latter) invalid? Does it violate some conditions necessary to describe the posterior of the scenario? Does the invalidity apply only to a multivariate autoregressive model?
For example predict(), in rstanarm, would not re-estimate, and apparently that step would be redundant for non-recursive systems.
The posterior of the scenario in the two-step algorithm would consist of the parameter and error distributions generated in the estimation phase. I would draw from the same set independently of which hypothetical scenario I am entertaining.
Any view on this? Any useful reading, non necessarily from econometrics?
Thanks
Gianni