EconSieve - Transposed-Ensemble Kalman Filter (TEnKF) and Nonlinear Path-Adjusting Smoother (NPAS)


Apart from the smoother, I literally stole most of the code from these two projects:

* https://github.com/rlabbe/filterpy
* https://github.com/pykalman/pykalman

They deserve most of the merits. I just made everything look way more complicated. Sometimes filterpy was more efficient, sometimes pykalman. Unfortunately the pykalman project is orphaned. I tweaked something here and there:

  • treating numerical errors in the UKF covariance matrix by looking for the nearest positive semi-definite matrix
  • eliminating identical sigma points (yields speedup assuming that evaluation of each point is costly)
  • extracting functions from classes and compile them using the @njit flag (speedup)
  • major cleanup

NPAS is built from scratch. I barely did any testing as a standalone filter and just always used it in combination with the ‘pydsge’, where it works very well.

Some very rudimentary documentation can be found here.

Installation with pip (elegant via git)

First install git. Linux users just use their respective repos.

Windows users probably use anaconda and can do

conda install -c anaconda git

in the conda shell as they kindly tell us here. Otherwise you can probably get it here.

Then you can simply do

pip install git+https://github.com/gboehl/econsieve

If you run it and it complains about missing packages, please let me know so that I can update the setup.py!

Installation with pip (simple)

First, be sure that you are on Python 3.x. Then:

The simplest way is to clone the repository and then from within the cloned folder run (Windows user from the Anaconda Prompt):

pip install .

Updating

The package is updated very frequently (find the history of latest commits here). I hence recommend pulling and reinstalling whenever something is not working right. Run:

pip install --upgrade git+https://github.com/gboehl/econsieve

Citation

pydsge is developed by Gregor Boehl to simulate, filter, and estimate DSGE models with the zero lower bound on nominal interest rates in various applications (see [here](https://gregorboehl.com) for research papers using the package). Please cite it with:

We appreciate citations for pydsge because it helps us to find out how people have been using the package and it motivates further work.