Probability calibration for sklearn models

[gngdb]

Probability calibration was recently given it's own module in Scikit-learn. Our model could probably benefit from this (or just using a logistic regression model) to minimise log loss. Should really have thought about this already.

[scottclowe]

This calibration results in a lower log-loss. Note that an alternative would have been to increase the number of base estimators which would have resulted in a similar decrease in log-loss.

Currently I am fixing the issue by using lots of trees.