Should I normalize my input data before feeding into PyMC3?

I have a doubt about data normalization

I am estimating a fairly complex non-linear model with many parameters. Some of those have an interpretation as physical quantities. However, when using the raw data as input the sampler (NUTS) does not seem to converge but works better when rescaling the input data to be between 0,1.

Unfortunately, however, when the input is rescaled the estimated parameter loose interpretation as well defined physical quantity and it’s hard to set their priors to a reasonable range.

Any solutions to this?

What is the practice around data scaling, should it be performed or not ?

1 Like

The problem is that the correlations can take on pretty high values. There will soon be improvements to the sampler tuning that should help here. However, if the model works better if you rescale that’s a pretty compelling reason to do so. Can’t you add a Detereministic that scales the parameter to your original space?

1 Like