Negative initial value for Gamma prior

As the message says, this is reporting the log probability (logp) of the initial value for beta under your specified prior

My guess is that np.log(traindf.days_since_last_order+0.01)) is allowed to be negative (indeed you add 0.01, so i guess there are zeros and log(0.01) < 0. So mu is negative, and the Poisson is undefined.

You should use pm.math.exp(mu) instead of just mu in the Possion. Reading a bit more carefully, I see you exponentiation inside the outer product, but it’s probably safer to work entirely in the latent space until the last moment