Training issue with zeros in spend data

[adavoli91]

Hi,

I'm facing some problems with model training: when the spend data contains a non-negligible number of 0's (while the total spend is always positive for each channel), the walker "doesn't walk"—i.e., model.trace is characterized by constant values equal to the initial ones. The training doesn't proceed, since the walker doesn’t move around the parameter space.

I noticed that the problem arises particularly when multiple channels show this feature (a non-negligible number of 0's in the time series).

I tried to extract the MCMC diagnostics, where divergences are stored, and I found that all the samples have diverging = True.

One solution I found is to clip spend data to very small values (~1e-10). However, in this case the model predicts a non-vanishing revenue for the clipped spend data. This effect, accumulated over time, becomes non-negligible. A possible workaround to this problem would be to use custom priors to restrict the parameter space. However, I’d like to avoid this approximate approach, since it may exclude potentially valid solutions from regions of the parameter space.

I'd like to know if it’s possible to deal with 0's in the time series without having to manually clip them or restrict the parameter space.

Thanks!