Sorry ARIMA, but I’m Going Bayesian

I was going through this blog and wondered whether someone has tried to build an equivalent analysis in pymc3.

Any leads are appreciated.

2 Likes

We dont have a Bayesian Structural time series module. You can try tensorflow_probability: https://github.com/tensorflow/probability/tree/master/tensorflow_probability/python/sts

You can certainly build models like this pretty easily in PyMC3. I use them frequently in my applied work.

junpenglao, what would it take to port that functionality over to PyMC4? If PyMC is almost ready to be stably based off of Tensorflow Probability, then it shouldn’t be too hard to get bsts from Tensorflow Probability to work, would it? At any rate, what discrete things would need to be done?

I’d be glad to start looking into this and potentially working on it, if some guidance were given. Bayesian Structural time series do not have a good way of being done anywhere, in either R or Python (by my last check), and they personally interest me.

I will be happy to contribute.

1 Like

Could you share any examples?

A little late but have you looked at timeseers ?

https://github.com/MBrouns/timeseers

https://discourse.pymc.io/t/hierarchical-time-series-with-prophet-and-pymc3-by-matthijs-brouns/5988

1 Like