Develop better win probability model

[guidopetri]

Develop better win probability model

[guidopetri]

I've started adding the clock times and increment as features in the WP model, specifically in 77e86e9. I need to reorganize a lot of this code so that I can develop a better model I think, and there's a few features/different transformations that I still want to explore before calling this good for now.

Kyle Burris of Guardians fame suggested adding an "expected pace of play" feature by training a model to predict clock times based on move number, but using only the data from the winning side of a match. The idea here being that if your clock time is below the expected pace, then you'll have a harder time coming back from that and winning the game (unless, of course, you're winning OTB by a lot, but that's nuance that will be captured by the model).

[guidopetri]

Done in befca67 . The pace of play doesn't seem to affect the models much, but it seems to make the statsmodels logit model worse in Brier and better in AUC, and the opposite way around for the sklearn LR model, both for 2800+ blitz players and for the entire dataset.

For now I think I'm closing this issue. I've added in clock times in v2, so I'm happy with this version of the model.