Python version

[rpanai]

Are you planning and/or interested to have a python version?

[alhauser]

There are at least two Python ports:

https://github.com/jamalsenouci/causalimpact https://github.com/tcassou/causal_impact

They were developed completely independently from this original R version of CausalImpact, so there's no "official" Python version; we haven't compared the ports to the R version so far (would be interesting to do, though).

[rpanai]

Hi, I'm aware of these two ports but I was wondering if anything is going to come from you.

[alhauser]

No, we don't have plans to provide an own Python port.

[WillianFuks]

If anyone finds this open thread, I've just ported this library for Python using TensorFlow Probability, you can check it here:

tfcausalimpact

All features have been ported, the code is 100% tested, results have been compared to R's and the project is fully documented.

Hope it's useful :)

[rpanai]

@WillianFuks it looks promising. The problems with the others two ports rely on statsmodels 0.8.0 which is kind of a pain. Are you planning to build a conda package of it? If not I can do as I did for jamalsenuci port https://anaconda.org/teamcore/causalimpact

[WillianFuks]

@rpanai good idea :) If you could create the package that would be really cool :+1:

Yeah, I agree about statsmodels version 0.8. Newer version 0.9 was a great improvement and in fact the new features made it possible to compute all the inferences and p-value in this new package (also the team is awesome, they helped me a lot to complete this project).

[WillianFuks]

@WillianFuks it looks promising. The problems with the others two ports rely on statsmodels 0.8.0 which is kind of a pain. Are you planning to build a conda package of it? If not I can do as I did for jamalsenuci port https://anaconda.org/teamcore/causalimpact

I've just published a new package. This time results are equivalent to the original package and does not rely on statsmodels anymore (now it uses TensorFlow Probability).

[rpanai]

Nice. Do you still need a conda package?

[WillianFuks]

That would be awesome!

[rpanai]

@WillianFuks is the new package https://github.com/WillianFuks/tfcausalimpact?

[WillianFuks]

Yeap!

[Cherishzhang]

@WillianFuks we did some A/A test experiments to measure the quality of control set. We just found python package performs much better than R package when selecting more items as control set. wondering is it expected or not?

[WillianFuks]

Hi @Cherishzhang ,

For Python we have pycausalimpact and tfcausalimpact, I'm supposing you are using the latter (the former doesn't use Bayesian inference and therefore results are expected to be different).

It's hard to say, I'm not sure if you should observe much different results. Both packages (R and Python) are built using similar principles and formulas with some implementation differences. One of those differences, for instance, is that the linear regression for tfcausalimpact is based upon a horseshoe prior whereas R's uses spike-and-slab.

Both are equivalent but maybe the way horseshoe models sparsity for the linear regression was more helpful to the dataset you are working with.

It could just as well be the case that on another dataset R might outperform as well. But in general I'd expect both packages being in accordance with conclusions.

In a nutshell, maybe the assumptions in the Python package were helpful to the dataset you're working with but for the most part I'd expect that both will give similar results.

[aazz7777]

I compare R and python both model.

I'm wonderi

---

if you look at the bottom left, it is R, second python(VI), python(hmc) and prob respectively.

How can we tell which one is performed better and which one I should use for the recommendation. I got some discrepancy of the output for my data.

[WillianFuks]

This is quite weird indeed. If you plot the forecasted data do both packages plot same line? In your data, is there an expectancy if results are bigger or lower than the counterfactual?

[aazz7777]

yes this is the plot from both model R One is Python

No expectancy with this data.