Uses of Simulation and High Performance Computing in Econometrics
Michael Creel
UAB
Outline
- What is econometrics?
- Uses of simulation in econometrics
- Benefits of HPC to econometricians
- ParallelKnoppix
- PelicanHPC
Disclaimer
This presentation is heavily biased toward my own work. I'm only attempting to give examples
of things that work pretty well. I'm not attempting to give a broad picture of all the possibilities.
What is econometrics?
- An economic model is a mathematical representation of how economic agents behave.
- Econometrics is the application of statistical methods to estimate the parameters of economic models.
- Economic data provides information, which is used to choose parameters to make the model resemble economic
reality.
- Wikipedia entry
- my notes
Example of econometric model
Uses of simulation in econometrics
- Monte Carlo
- Bootstrapping
- Solving a model: PEA
- Simulation-based estimation
Monte Carlo
- Econometric theory usually gives analytic results for behavior of estimators as the
sample size goes to infinity (asymptotic theory). This may be a poor approximation to behavior in
finite samples.
- Wikipedia entry: notice the bit about simulating pi
- In econometrics, typically, an estimation method is proposed, and Monte Carlo is used
to evaluate how well it works.
- go to PK cluster, and look over mc_example1
Bootstrapping
- Wikipedia entry
- Most analytical results for the behavior of econometric estimators are based on asymptotics.
These results may not be reliable in finite samples.
- Bootstrapping is a means of learning about finite-sample performance by resampling.
Bootstrapping
- The basic idea is to compute a statistic using artificial samples,
and to then use the empirical distribution of replications of the statistic
- Resampling may be done using a number of techniques:
- resample from data (single obsns or blocks).
- resample from residuals after fitting model
- etc...
Bootstrapping
go to PK cluster, and look over bootstrap_example1
Solving a model: PEA
- Sometimes, simulation must be used simply to solve a model. Solving the model is required before data
can be simulated from the model.
- This is not always easy to do.
Solving a model: PEA
- For example, an agent's beliefs about future events can influence current decisions.
- If beliefs are rational, they should be correct, on average. Such beliefs are known as
"rational expectations".
- The functional form of rational expectations is in general not known when the model is nonlinear. Without
knowing the form of expectations, the model can't be solved.
Solving a model: PEA
- Simulation methods can be used to fit a model of expectations to data generated by the model. When the data generated by the model,
conditional on expectations, is consistent with the model of expectations, a solution has been found.
- go to PK cluster, and look over pea_example
Simulation-based estimation
Simulation-based estimation
Simulation-based estimation
Simulation-based estimation
go to PK cluster, and look over SNM
How well does this work?
The "ParallelKnoppix" cluster
- Running ParallelKnoppix
- 2 servers, each has 2 Xeon 3.6 GHz CPUs and 8GB RAM
- Donated by Intel Software for development of ParallelKnoppix - thanks!
PelicanHPC
- PelicanHPC is the successor to PK - its foundation is Debian Live, so much easier to
maintain and customize
- Pelican homepage
- switch to VMware to show PelicanHPC
[any material that should appear in print but not on the slide]