particle filter dynare

[jpfeifer]

When downloading and installing the unstable version, you can find the up-to-date manual in the doc-folder. It's called dynare.pdf

[shengan28]

By including mode_check, the mode check plots are created. How can we get the parameter trajectories and parameter densities plots to be created? I am able to create them with other estimation programs, but not this one.

[jpfeifer]

Which estimation command exactly did you use?

[shengan28]

I have been playing around with the linear_model.mod program.

See https://github.com/DynareTeam/dynare/blob/master/tests/particle/linear_model.mod

Are the option-options the new, or the old approaches to programming SMC.
For example,

Code: Select all

options_.particle.number_of_particles = 20000;
options_.particle.resampling.status = 'systematic';
options_.particle.resampling.neff_threshold = .1;


can also be programmed within the estimation command:

Code: Select all

estimation(order=2,nobs=100,mode_compute=8, mode_check,datafile=linear_model_data,mh_replic=0, filter_algorithm=sis, resampling=systematic, resampling_threshold=.1,  mode_file=linear_model_mode);

or more specifically, my question is which approach is preferable?

Also, is there documentation regarding these options?>

Code: Select all

options_.gstep(1) = 1e-4;
options_.gstep(2) = .1;

options_.mode_check_neighbourhood_size = 0.05;

set_dynare_threads('local_state_space_iteration_2',3);

[jpfeifer]

In general, you can set any Dynare option manually the way you see it above. If you use the preprocessor interface, all it does is translate the commands to statements like the one above. The preprocessor interface is less prone to mistakes and therefore usually more preferable.

The gstep part is size of the gradient step used in numerical solvers.

Code: Select all

mode_check_neighbourhood_size

is in the manual.

set_dynare_threads is a command to

set the number of threads used by some MEX files when compiled
with OpenMP support (i.e with --enable-openmp is given to configure) or any
other parallel library.

[shengan28]

That is good insight. Thanks.

My next question has to do with programming Particle filters.
The command

Code: Select all

filter_algorithm=sis

is the sequential importance sampling. However, by adding

Code: Select all

resampling=systematic or generic


Doe this make the particle filter the sequential importance resampling filter? It seems obvious, but the manual does not explicitly say this.

Does this option work with any of the other filters? It must not do anything to the APF

Lastly, the

Code: Select all

mode_compute=11

runs the on-line estimation. Does this option over ride the particle filters in the

Code: Select all

filter_algorithm=

command?

[jpfeifer]

Concerning the first question on

Code: Select all

filter_algorithm=sis

and

Code: Select all

resampling=systematic

the answer is yes. And it is the default.

Code: Select all

resampling_threshold

only works with

Code: Select all

generic

and is useless with

Code: Select all

systematic

or

Code: Select all

none

. With

Code: Select all

apf

it induces a second resampling step on current particles which is not necessary. The literature then recommands

Code: Select all

resampling=none

for apf.

Keep in mind that maximum likelihood will not function in presence of resampling since it generates a discontinuity of the criterion with respect to the structural parameters. In that case, avoid gradient-based methods.

Concerning

Code: Select all

mode_compute=11

it is an auxiliary particle filter and it overrides the other options that control other particle filters. It is useful to obtain fast estimation on parameters by including them as extra states but suffers from a severe lack of theoretical foundations.

[shengan28]

Thank you for the detailed reply. That makes things clear.
Can someone check the NKLF option with the linear_model.mod . It runs for several minutes before crashing.

[jpfeifer]

Could you please provide the mod-file

[shengan28]

The NLKF filter starts out strong, but then crashes at Simplex iteration number: 4964-4-3, maybe after 20 minutes.

linear_model_nlkf.mod

(1.44 KiB) Downloaded 52 times

Notice that this file expects the linear_model.mode.mat file, but This board won't let me upload .mat files. Be sure to eliminate the

Code: Select all

mode_file=linear_model_mode

in the estimation command.

Thanks for looking into this bug.

~Shawn

[jpfeifer]

You may try it again with the proposed change at https://github.com/DynareTeam/particles/pull/8