likelihood in different datapoints

Topics: Kalman Filtering
Mar 20, 2014 at 11:34 AM
Dear Musketeers,

A colleague is coding SMC algorithm to estimate DSGEs. I have recommended to use Iris and he is, indeed, happy with that. Naturally, he need an efficient routine to calculate the likelihood. What would be the most efficient way to do it in Iris when one wants calculate the predictive density in each data point (not a sum of those like loglik returns).

Cheers,

Antti
Coordinator
Mar 20, 2014 at 12:13 PM
Just to make sure I understand correctly: you simply want to have access to all the individual contributions that make up the overall likelihood in the Kalman filter? If so, submit a feature request, and I'll add this. It's relatively easy.

Maybe also describe to me a little bit the context in which you work (what sequence of functions you call, etc) so that I can design it the most efficient way.

Jaromir
Mar 20, 2014 at 12:41 PM
Thanks Jaromir! This is indeed what he wants, ie each component of likelihood (when loglikelihood is the sum of conditional densities). I will make a feature request.

Supposing that I understand the Sequential Monte Carlo (see, eg, http://economics.sas.upenn.edu/~schorf/papers/smc_paper.pdf) correctly, the idea is that we start with subsample (ie let the data speak initially little) and increase the sample step-by-step. In each step we actually need the predictive density (likelihood) just in the last (T^th) observation. I think this makes no difference in Kalman filtering, one always need to calculate the predictive density in whole sample to obtain the contribution of the last observation.

The SMC has the benefit that it is easy to parallellize massively. More importantly, it works with nasty, eg multi-modal, posterior distribution and is, therefore, fool-proof.

Cheers,

Antti
Coordinator
Mar 26, 2014 at 10:45 AM
Hmmm. Interesting. Thanks, Antti. Maybe we should just implement the SMC algorithm in IRIS directly. Seems like it's up my alley.... :-)
Coordinator
Mar 26, 2014 at 1:13 PM
Indeed. But don't split up with anyone just because you want to work on SMC :))
Coordinator
Mar 26, 2014 at 1:16 PM
Hahaha :)))) I should finish the neural network stuff first, as well... but I think this is probably worth putting on the list of "wants" :)
Coordinator
May 8, 2014 at 4:26 PM
An issue has been created for this discussion thread, and a new feature added.
Marked as answer by jaromirbenes on 5/8/2014 at 9:26 AM