Can we have some discussion about how to implement this? Because "training" the encoding happens here and imo this is not straightforward to swap out for a resampled/cross-validated version.

In principle what we'd basically do:

• get_state_dt stays as it is. This is used only during predict on new data. We should perhaps factor out the switch logic into a separate function.
• During train, we estimate the state as-is, but instead of transforming using the state, we transform data using the CV strategy.
  Two comments:
• We should perhaps consider factoring out those CV methods if we see more of them popping up.
• Before we do anything we should perhaps see whether we might not want to simply pull in vtreat as a dependency.

this thread alternatively advocates adding a small noise to avoid overfitting and contains some additional interesting info