Deploy win probability model

[guidopetri]

Deploy win probability model

[guidopetri]

Pivot time: I was going to use the statsmodels logit model (that had slightly better metrics) but it turns out that the pickled file is 175mb in size, because it stores the training data within the model itself. There's some options to reduce the size footprint but it's not enough. The sklearn LR model is... 855 bytes. So I will instead use that.

[guidopetri]

I think this is ready to deploy as of 1ea33d5 . Obviously something will go wrong :) but I have tested it a lot (and even found a bug in the evals code) so I hope it will work fine.

[guidopetri]

I found a few issues:

• forgot to install sklearn
• forgot to create the new win probabilities table
• a couple of positions without evals bleed through (5rk1/p4p2/2p3pp/5q2/3R1NnP/4P1P1/PPQ2PK1/8 w - - 13 for example)
• a couple of positions don't have a clock time even though they should (https://lichess.org/SM7RGAXR#69 for example)

[guidopetri]

Oh god... I found an off-by-one error in the positions vs evals/clocks. For positions, I was counting the initial board state - for evals/clocks coming from lichess, I don't count the initial state. I think I have to re-run everything :(

[guidopetri]

Well... this looks like it's deployed and running fine now. I'll probably have to re-run a bunch of stuff at some point because of this off-by-one error though.