Cannot replicate fit

[cosmin-novac]

Hi. I'm trying to replicate your notebook results, but running the code on my machine with no changes leads to these unflattering results:

Any ideas?

Edit: For the sake of completeness:

[cosmin-novac]

Okay, I managed to fix this by using adam instead of adadelta as optimizer. The point prediction is good now, but then the quantiles fall out of place:

[cosmin-novac]

I managed to get somewhat accetable quantiles using RMSprop as optimizer. But it's scary how results can differ...