Getting wrong results while trying to replicate 2017 pycon talk

The code was taken from Which holds the jupyter notebooks for the pycon 2017 talk. As one can check he was getting a much better fit, while the exact same code on my system is giving this wrong fit.


This notebook contains the list of packages and the code.

If anything I would trust the new fit much more, the old advi initialization tend to underestimate the covariance which makes the samples also return underestimated variance sometimes.

Thanks for the quick response.
As you might have guessed I am new to all this, and that’s why when HPD interval size went from 66% to 25% I got confused. To be honest it’s still still hard for me digest that such a drastic change can occur even after we have around 100 data points. But, I guess I need to develop more intuition about all this.

Hmmm maybe I misunderstood you, which part of the mismatch you are referring to?