On Tue, 16 Jun 2020, Riccardo (Jack) Lucchetti wrote:
On Tue, 16 Jun 2020, Artur Tarassow wrote:
> Hey all,
>
> during the last couple of days I had to use Python's popular
'statsmodels'
> package for estimating ARIMA type of models. Here is a comparison with
> gretl's built-in apparatus for those who are interested:
>
>
https://github.com/atecon/benchmark_arima
>
> To summarize: 'In terms of speed, statsmodels is no competitor at all.' ;-)
>
> Let me add one thing: The computational heavy task is done in some
> kalman-function which calls cython code which is written in C. However, tt
> seems that there optimization algorithm is not well-programmed compared to
> what Jack and Allin (maybe Sven?) have implemented. Well done, guys! :-)
Much of the credit goes to Oleh Komashko, who some time ago did a great job
in comparing our ARIMA code to other software's and led us to the present mix
of algorithms. Unfortunately, Oleh has disappeared from the community for
some time now but I hope he'll be back again, because he's a great hansl
coder and an asset for the while gretl community.
Yes, I was thinking of Oleh the other day, and I hope he comes back.
The big speed-up in our arma came from switching from Kalman to Guy
Melard's algorithm AS197.
Allin