Problems running main.R

[matdehaven]

"horserace_ecy" robustness

Gives the following error:

Error in lhout[idraw] <- gout$lhout : replacement has length zero

This is coming from "1_vol_bvar_estimation.R" script, in the gdraw() function on line 78.

"horserace_nfci" robustness

Gives the following error:

Error in MASS::mvrnorm(n = 1, mu = rep(0, length(optout$x)), Sigma = optout$opt$H) : 
  'Sigma' is not positive definite

This is coming from "1_vol_bvar_estimation.R" script on line 45.

"1M" robustness

runs out of memory on my computer (16 GB)

[fernando-duarte]

Here are my results.

"horserace_ecy" robustness

Runs to completion without error in my computer.

"horserace_nfci" robustness

Runs to completion without error in my computer.

"1M" robustness

Runs to completion without error in my computer. I have 32GB.

"horserace_no_vfci_yes_gz" robustness

Gives the same error that Matt is getting in "horserace_ecy":

Error in lhout[idraw] <- gout$lhout : replacement has length zero

I ran it 10 times, and each time it gave the same error, so it does not seem to be run-dependent.

"horserace_no_vfci_yes_gz" robustness

Gives the same error that Matt is getting in "horserace_nfci" robustness:

Error in MASS::mvrnorm(n = 1, mu = rep(0, length(optout$x)), Sigma = optout$opt$H) : 
  'Sigma' is not positive definite

I also ran this one 10 times, and each time it gave the same error, so it does not seem to be run-dependent either.

[matdehaven]

It seems if I adjust the seed in 1_vol_bvar_estimation.R I am able to get those robustness checks to run. I will confirm that this is true and see if there is a seed that works for me.

I think that the Bayesian VAR is approximating a matrix inverse operation numerically, which relies upon the seed. But it must be that the matrix is close to not positive definite/invertible, so we sometimes get these errors. It is odd that it depends not just on the seed, but also which computer we are running it on.

[matdehaven]

If I change the below line

macro_dynamics/1_vol_bvar_estimation.R

Line 12 in cba1833

set.seed(1234)

To a seed of 83622 (randomly generate seed) all of the robustness checks work for me.

[fernando-duarte]

I tried setting the seed set.seed(8632) got the same error as before when running horserace_nfci:

Error in MASS::mvrnorm(n = 1, mu = rep(0, length(optout$x)), Sigma = optout$opt$H) : 
  'Sigma' is not positive definite

[fernando-duarte]

I tried setting the seed set.seed(968) got the same error as before

[matdehaven]

• Try editing the tightness of the bayesian priors at the following line:

macro_dynamics/1_vol_bvar_calibration.R

Lines 168 to 172 in cba1833

	mn_tight_calib <- 3
	mn_decay_calib <- 0.5
	if (type == "horserace_gz" | type == "horserace_tedr" | type =="horserace_ecy" | type == "horserace_gsfci" | type == "horserace_nfci" | type == "horserace_no_vfci_yes_gz" | type == "horserace_gz_tedr" | type == "horserace_no_vfci_yes_tedr"| type == "horserace_no_vfci_yes_gz_tedr") {
	mn_tight_calib <- 0.5 # with second financial variable
	}

[fernando-duarte]

Setting:

mn_tight_calib <- 3

made the entire code run without errors. On the other hand, some of the plots came out funky, e.g.:

mcmc_out_horserace_no_vfci_yes_tedr_irplot_all_shocks_TEDR_shock.pdf

mcmc_out_horserace_ecy_irplot_all_shocks_VFCI_shock.pdf

I also tried a value for mn_tight_calib of 1, but that gave the same error as when I set it to the original value of 0.5.

Any suggestions on what value to pick?

[matdehaven]

I was only able to get the code to run when I increased the mn_tight_calib to 100. I got the same errors as before trying 3, 5, and 10.

For those figures, however, I think the problem is just that the y-axis scale is hard-coded here:

macro_dynamics/1_vol_bvar_calibration.R

Lines 55 to 61 in dd24f4b

	} else if (type == "horserace_tedr") {
	vars_in_system <- c('lgdp', 'lpce', 'vfci', 'tedr', 'fedfunds')
	var_names <- c("Log Real GDP", "Log Core PCE", "VFCI", "TEDR", "Fed Funds")
	shock_names <- c("Real GDP shock", "Core PCE shock", "VFCI shock", "TEDR shock", "Fed Funds shock")
	yaxis_vfci_shock <- list(c(-0.015,0.002),c(-0.005,0.005),c(-0.04,0.3),c(-0.04,0.4),c(-0.005,0.001))
	yaxis_vfci_response <- list(c(-0.1,0.1),c(-0.05,0.05),c(-0.04,0.3),c(-0.04,0.1),c(-0.015,0.15))

[matdehaven]

I think I found the solution.

I have added the following code:

macro_dynamics/1_vol_bvar_calibration.R

Lines 188 to 192 in 15962e8

	disperse = TRUE ## if TRUE, disperse the starting point using inverse hessian as covariance matrix
	## For some of the robustness checks, the inverse hessian matrix doesn't work, so we set disperse to FALSE
	if (type == "horserace_gz" | type == "horserace_tedr" | type =="horserace_ecy" | type == "horserace_gsfci" | type == "horserace_nfci" | type == "horserace_no_vfci_yes_gz" | type == "horserace_gz_tedr" | type == "horserace_no_vfci_yes_tedr"| type == "horserace_no_vfci_yes_gz_tedr") {
	disperse <- FALSE
	}

The "disperse" option uses the inverse hessian to adjust the start of the bayesian draws. This inverse hessian matrix is the one that doesn't work for some of us, on some seeds, for some of the horserace robustnesses. If I turn it off, then the code runs for me. I have set the mn_tight_calib value to 3 for all robustness runs, which is what we have been using for the baseline.

@fernando-duarte @ti719 Can someone confirm the code runs for them in this new state?

[matdehaven]

Also, I have updated the y-axes for the IRF figures to try to fix the problem Fernando was noting earlier. There might still be a few that need tweaking.

[ti719]

The y axis would need to be adjusted for some calibrations based on the magnitude of the IRFs. In my computer, the baseline calibration of 0.5 runs but some of the others tried here do not work. We should probably not go to a calibration above 5

…

On Tuesday, March 4, 2025, Matthew DeHaven ***@***.***> wrote: I was only able to get the code to run when I increased the mn_tight_calib to 100. I got the same errors as before trying 3, 5, and 10. For those figures, however, I think the problem is just that the y-axis scale is hard-coded here: https://github.com/VFCI/macro_dynamics/blob/dd24f4b420d6cda343dea66ca38955 98453b42d5/1_vol_bvar_calibration.R#L55-L61 — Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADAMISLDBC64FNPDTL6J4C32SXE7XAVCNFSM6AAAAABWWHGDZSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMOJYGE2DONBQGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> [image: matdehaven]*matdehaven* left a comment (VFCI/macro_dynamics#11) <#11 (comment)> I was only able to get the code to run when I increased the mn_tight_calib to 100. I got the same errors as before trying 3, 5, and 10. For those figures, however, I think the problem is just that the y-axis scale is hard-coded here: https://github.com/VFCI/macro_dynamics/blob/dd24f4b420d6cda343dea66ca38955 98453b42d5/1_vol_bvar_calibration.R#L55-L61 — Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADAMISLDBC64FNPDTL6J4C32SXE7XAVCNFSM6AAAAABWWHGDZSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMOJYGE2DONBQGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[matdehaven]

Closing this issue as it seems resolved by setting disperse=false for the financial robustness checks.