In this case I ended up with two very simple trees which I train for each ATM (with a prior trained on the larger fleet). You could try a mix though. I didn't have time to try too many things.
The following step which I didn't mention, was to use the trees for generating predictions into the future. Little monte carlo simulations using one of the tree (selected using that relative entropy methods, you could even use a weighted mix of both trees) that I run a few thousand times for each ATM and to generate a forecasted range of when the ATM is going to run out of cash.
The following step which I didn't mention, was to use the trees for generating predictions into the future. Little monte carlo simulations using one of the tree (selected using that relative entropy methods, you could even use a weighted mix of both trees) that I run a few thousand times for each ATM and to generate a forecasted range of when the ATM is going to run out of cash.