Michael Creel michael.creel @ uab.es

Contact information, working papers,
are here

Information about published papers is here

Computer code is at https://github.com/mcreel

IDEA Econometrics II

- On Selection of Statistics for
Approximate Bayesian Computing (or the Method of Simulated
Moments) (with Dennis Kristensen) (2015, Computational
Statistics & Data Analysis)

- code using written in the Julia language for replication of results or using the method with for your own problem is here: https://github.com/mcreel/SelectStatisticsABC
- contact me if you would like the data file for the jump
diffusion example

- I highly recommend getting to know the Julia language - it
has a friendly Matlab-like syntax, but speed like C.

- Indirect
Likelihood Inference

- this is a revised version (June, 2013). The link also
provides the paper as well as OLD software to replicate the
DSGE and auction examples as in that version. Newer and
*considerably improved*code is in the bullet items that follow. The results using the new code will appear in yet another revision, expected later this year.

- most recent code for the auction model: https://github.com/mcreel/ABCAuction

- this code uses adaptive importance sampling, as described
in algorithms 2 and 3 of the CS&DA paper, above. This
gives very good results using many fewer simulations than in
the 2013 version of the Indirect Likelihood Inference paper.

- the code also uses local linear nonparametric fitting, as
suggested by Beaumont et al. 2002 and Gao and Hong, 2015.
This gives results with about 20% smaller RMSE than local
constant nonparametric fitting.

- most recent code for the DSGE model: https://github.com/mcreel/ABCDGSE

- this code uses the selection of statistics from the
CS&DA paper, importance sampling, and local linear
regression. The code is now follows theory completely, it is
pretty fast, and it seems to work very well

- here are results for the DSGE model described in the
paper, n=160, using a single set of statistics selected
using the method of the CS&DA paper. pmean, etc., psdev,
etc., refer to the prior, without a p refers to the SBIL
estimator:

`true mean pmean sdev. psdev bias pbias rmse prmse`

alpha 0.33000 0.32340 0.30000 0.01385 0.05774 -0.00660 -0.03000 0.01533 0.06506

beta 0.99000 0.98891 0.97250 0.00335 0.01299 -0.00109 -0.01750 0.00352 0.02179

delta 0.02500 0.02394 0.05500 0.00321 0.02598 -0.00106 0.03000 0.00337 0.03969

gam 2.00000 2.02897 2.50000 0.16843 1.44338 0.02897 0.50000 0.17074 1.52753

rho1 0.90000 0.89794 0.49500 0.01492 0.28579 -0.00206 -0.40500 0.01505 0.49568

sigma1 0.02000 0.02154 0.05000 0.00362 0.02887 0.00154 0.03000 0.00393 0.04163

rho2 0.70000 0.72020 0.49500 0.08187 0.28579 0.02020 -0.20500 0.08424 0.35171

sigma2 0.01000 0.01082 0.05000 0.00283 0.02887 0.00082 0.04000 0.00294 0.04933

nss 0.33333 0.33537 0.31250 0.01071 0.03608 0.00203 -0.02083 0.01089 0.04167

- The older (2011) version with additional examples is here
- An even older working paper was the origin of this work. It
might still have some elements of interest. The VAR example is
based on flawed code, though, ignore it: "A
Data Mining Approach to Indirect Inference"

- Econometrics
notes Graduate level notes for econometrics, with numerous
examples embedded as links. If you use the notes, you'll find
the examples ready to run using virtualization software on the accompanying
live image.
The live image has GNU Octave, Julia and R installed. Most
examples use Octave, but I plan to add more examples using
Julia. R is included mostly to run other people's code.

- "Econometrics on GPUs" (with S. Mandal and M. Zubair). Examples of econometric-relevant code running on GPUs, using CUDA and Thrust. Has examples of speedups >50X compared to a single CPU core. Using multiple GPU devices, speedups of >200X can be optained. Working paper code archive (note: the code archive was missing the pg.cu file, it has been corrected as of 02 Dec. 2012).
- Octave
code for econometrics: bfgs, simulated annealing, MLE,
GMM, kernel estimation, etc.

- (no longer in development, but works perfectly well). The amazing PelicanHPC bootable CD to create a HPC cluster for parallel computing in minutes!