other SVARs to be considered

[donotdespair]

Papers we should implement

SVARs

• [x] Antolín-Díaz & Rubio-Ramírez (2018, AER) - narrative sign restrictions
• [x] Arias, Rubio-Ramírez & Waggoner (2018, ECTA) - zero and sign restrictions (zero restrictions only works for IRF at horizon 0 now)
• [ ] Arias, Rubio-Ramírez & Waggoner (2021, JOE) - Proxy-VARs share the same algos
• [ ] Baumeister & Hamilton (2015, ECTA) - sign restrictions via calibrated prior
• [ ] Giacomini & Kitagawa (2021, ECTA) - Robust Bayesian Inference for Set-Identified Models (won a paper of the year or even better in ECTA!)
• [x] Rubio-Ramirez, Waggoner & Zha (2010, RESTUD) - flexible sign restrictions
• [ ] Rubio-Ramirez, Waggoner & Zha (2010, RESTUD) - flexible exclusion restrictions

BVARs

• [x] Giannone, Lenza & Primiceri (2015, RESTAT) - hyper parameters
• [ ] Lenza & Primiceri (2022) How to estimate a vector autoregression after March 2020

Papers that are useful

• [ ] Arias, Rubio-Ramírez, Shin & Waggoner (2024, WP) - TVP with sign restrictions

Parallel computing

• [x] parallelize with OpenMP
• [x] progress bar with parallelization

[adamwang15]

Development progress of Giannone, Lenza & Primiceri (2015, RESTAT) in branch develop-glp

• [x] log prior of hyper parameters
• [x] log marginal likelihood with dummy observations
• [x] log posterior of hyper parameters
• [ ] mcmc algorithm
• [x] reproduce figure 1

[donotdespair]

prepare a separate specify_* function and estimate method for the particular models by #14

• [ ] RRWZ2010 with sign restrictions
• [ ] ARRW2018 with zero and sign restrictions
• [ ] ADRR2018 with sign and narrative restrictions
• [ ] GLP2015 with estimated hyper-parameters

[donotdespair]

Hey @adamwang15

I have finished work with specify_prior_bsvarSIGN. It is working well and passes all the relevant tests in inst/tinytest/test_specify.R. But it crashes other parts of the algo and I haven't yet introduced these changes. That's the upcoming task for the both of us!

Cheers, T

[adamwang15]

Great! (I think) I fixed the crash and now the code runs as usual, I will keep experimenting the GLP code!

[donotdespair]

On Robust Adaptive Metropolis (RAM) by Vihola (2012)

Hey @adamwang15 Amongst the Adaptive Metropolis algorithms we should try for the GLP model hyper-parameters sampler we should try the one by Vihola (2012). It is robust even for large dimensions of the candidate sampling density. There is an R package ramcmc implementing the RAM and it's written using RcppArmadillo. It is nice, small, and without other dependencies than Rcpp and RcppArmadillo. So, we could use it directly!

There's just one point we need to investigate: The RAM is designed for symmetric candidate generating densities. I am not sure how it'd work if one needs to use truncation (to avoid sampling negative variance coefficients). I'm investigating this.

• The answer might be here (but I'm not implementing this): Tang & Yang (2024, JMLR)
• software in R and C for the Adaptive Metropolis by Rosenthal is given on his website amcmc.

BTW, you were right about the meaning of $\alpha_i$. Nice! And also, that's much simpler than what I was coding!

Cheers, T

[adamwang15]

Hey @donotdespair thanks! Okay I missed this one! In the meantime, I coded an adaptive Metropolis algorithm and it seems to be working. This answer suggests log-transformation to sample positive parameters. I will try bringing in the ramcmc code, looks like it can be adjusted to include log-transformation as well.

• [x] use ramcmc for proposal variance in the Metropolis algorithm