Smoothed Variables , SMC estimation

Estimation ML/Bayesian estimation
1

Dear Prof @jpfeifer ,

After an SMC estimation, I see a variable matrix called smoother variables and updated variables in the results matrix.

I see this even when I run a fresh mod file and after deleting the previous workspace.

1)Are these variables smoothed by Kalman Smoother?

Does SMC automatically run the Kalman Smoother without any other option? Smoother option in the estimation command with SMC returns SMC doesn’t support smoother option.

2)If these are not indeed smoother objects, is there any way to run Kalman Smoother with SMC estimation?

3)Is calib_smoother reliable to produce series of unobservables/theoretical variables?

4)Also, is there a way to check the health of SMC estimation?

I am using 10000 particles, 30 steps, lambda=2, 0.25 target.

Thank you.

2

If you use order=1, Dynare will run the Kalman smoother. But from your description I don’t understand what Dynare version which which options you are using to generate these issues.

3

estimation(datafile=trs1,

posterior_sampling_method = ‘hssmc’,

posterior_sampler_options = (‘particles’, 10000, ‘steps’, 30,

‘lambda’, 2, ‘target’, 0.25),

mode_compute=5, mh_replic=0, prefilter=0) ;

I am using dynare 6.4.
Is order =1 a default option?

Is not mode_compute redundant with SMC?

4

Can your privately provide me with the full files necessary to recreate the issue?

5

Sorry Professor @jpfeifer , the question regarding smoothed variables was a mistake with my understanding, there were no issues with the estimation per se. Of, course the estimation procedure produces the path of all variables including unobserved variables, which can be accessed from the results.

Could you please answer a few of these important questions I am struggling with?:

  1. Is the number of steps option in dynare while using hsssmc indeed the number of stages mentioned in DSGE papers using the the technique as well as Herbst & Schorfheide SMC paper?
    ****I am using 10000 particles with 30 steps
  2. How does one choose the apropriate number of steps/stages for smc estimation?
    If the number of stages is indeed the number of steps, I see researchers mentioning 100, 200 even 600 stages, which will take at least 36 to 48 hours in my system. Are these many stages required for a robust, acceptable estimation?
    What I find is that the results are consistent across many runs(more than 20).
    The model is a small scale NK DSGE model.
  3. For a sub sample estimation I find the posterior for a few(3-4) parameters are not informed well from the data(rugged posteriors, uncertainty, bimodal etc). Is it acceptable practice to give these posterior means( a few of them where mode finders where struggling earlier) as priors for RWMCMC with 4 blocks and 1000000 draws(0.5 burn in) to get daignostics of convergence? The posterior moves a little (not drastically) away from the SMC estimates this way and the posterior plots also look well behaved and well informed(posteriors are taller and narrower than priors).

I am attaching the output window

smc2.log (5.4 KB)

mcmc2.log (61.3 KB)

SMC estimation and the prior posterior plots. Please let me know if these are acceptable or should I make changes to the estimation settings

smc1.log (5.4 KB)

smc1 is the full sample result which I want to validate if healthy or not.
smc2 is the sub sample result, whose posterior means I give as prior means for a few parameters and do mcmc, the mcmc results are in mcmc2.
The most important parameter is pibar, whose prior value I do not change in both estimation methods.

6

Also attaching the prior-posterior plots here

smc2(2).pdf (46.8 KB)

smc2(1).pdf (36.6 KB)

smc1(2).pdf (43.5 KB)

smc1(1).pdf (35.2 KB)