I have a binomial outcome that I am trying to predict using a gbm in h2o.
I have set quite a low min_rows value for each node and it appears to be overfitting. See plots below.
When I group the results into percentiles and plot Observed vs Predicted I get good results for Validation but Training shows a higher proportion of observed outcomes in the upper percentiles and lower observed outcomes in the lower percentiles. See images below.
I would have thought training would have predicted more accurately and the Validation and Hold Out plots would have been more inaccurate. Is this the expected result and why does it occur?