This is my submission for march ml mania 2015.

It is built on top of Lasagne, using a very convenient wrapper nolearn.

The model only uses player data, collect.py is the script that has all the routines to collect that data from the web.

I then use 5 players per team, choosing those that have played the most games.

The model itself (implemented in model.py) is a neural network with two hidden layers.

• First layer (implemented in custom_layers.py) is a partially connected layer, that ensures that features for every player have equal weights going forward to the second layer. First team players are connected to the first half of neurons in the next layer, and second team players are connected to the second half.
• Second layer is just a dense layer.

For the bracket challenge I prepared two brackets (bracket.py), one uses naive approach, where a team with higher probability always advances to the next round. Another uses smarter dynamic programming approach, maximizing the expected score of the bracket. Both brackets are in the root of the repo.