GARCH and related models in Python

[Kevin Sheppard]

I have finished some initial work on ARCH models and released it as a stand alone package. It is univariate only, but can jointly estimate a variety of ARCH models (GARCH, TARCH, EGARCH, HARCH) with a variety of models for the conditional mean (AR, HAR, LS) and some alternative distributions (Normal, Student's t)

http://github.com/bashtage/arch

https://pypi.python.org/pypi/arch

Docs at 

http://arch.readthedocs.org/en/latest/

It depends on statsmodels (and all of statsmodels dependencies, although expects more modern SciPy than SM 0.5 requires).  I've tried to stick with statsmodels names were appropriate.

Comments and contributions are welcome.

[Javier Burroni]

Thank you!

--
" To be is to do " ( Socrates )
" To be or not to be " ( Shakespeare )
" To do is to be " ( Sartre )
" Do be do be do " ( Sinatra )

[Da Jiao]

Currently still no Anaconda build for OS X?

Best,

Chris

[Kevin Sheppard]

If you have numba installed there is little reason to use a compiled version.  You can just install with the --no-binary flag. 

[josef...@gmail.com]

On Tue, Oct 13, 2015 at 1:43 PM, Kevin Sheppard <kevin.k....@gmail.com> wrote:

If you have numba installed there is little reason to use a compiled version.  You can just install with the --no-binary flag. 

Hi Kevin,

What's your experience now with numba? Should we consider it soon as an optional speedup for statsmodels when we don't have a cython developer available?

Josef

[Kevin Sheppard]

It has matured lot but is also still progressing quickly.  It is really good for most scalar algos but can call BLAS or most numpy matrix functions without losing any speed gains.  I probably wouldn't take a mandatory dep on it, but since the main way to use it is via a decorator, it is simple to make it optional by using a wrapper decorator that just noops if not installed (and possibly warns about performance).

[josef...@gmail.com]

Thanks for the info

On Tue, Oct 13, 2015 at 2:39 PM, Kevin Sheppard <kevin.k....@gmail.com> wrote:

It has matured lot but is also still progressing quickly.  It is really good for most scalar algos but can call BLAS

to clarify: it **cannot** call BLAS without loosing the speed gains

which means it will not help for kernel regression, which is a main candidate for speedups.

But there are or will be other case where we have nobs or nobs**2 loops that cannot be vectorized (at least not without blowing up memory) and that might be less demanding.

Josef

[tom yitav]

Hi Kevin, can you please show an example for creating multivariate GARCH from the univariate GARCH? I am trying to estimate the covariance matrix for several assets returns. Thanks