Not sure how I missed that…as I said, I’m missing something obvious, thanks 
Maybe it is because I expected Pymc to return the initial values that caused the problem, not the log-probabilities, but I should have read the message more carefully.
I don’t think mu is negative because I take the exponent of beta * np.log(traindf.days_since_last_order+0.01) and then the outer product with lambda0 which is also positive?
I hear you though about maybe keeping the transformations to the end, and staying in the latent space as far as possible.