phdntom / ncaa

NCAA

The goal of this project is to develop a collection of scripts/routines to help predict the odds of the NCAA tournament. To clarify, these programs do NOT build a "perfect bracket", they predict the odds! (Of course if you wanted to just go with the odds, you could use this to make a bracket, purely from machine predictions)

This project started in 2014 with a Kaggle competition but the great thing about the tourney is, it happens every year.

The input data may grow/change depending on what seems to be working, and what information is available/reliable.

CURRENT Modules

Two current prediction attempts exist:

1. Chessrank metric; "chessrank"

2. Matrix Factorization / Collaborate Filtering; "matrix_factorization"

split_seasons.py

Minimal script to quickly split the seasons.csv into multiple seasonX.csv in a directory structure.

season_accumulator.py

Gathers stats for each seasonX.csv and builds a Rating Matrix in the collaborative filtering sense.

Using the netflix movie rating paradigm as an example, each user has rated a collection of movies. The matrix is N_users x N_movies and is in general not square and is missing many entries (i.e. all users have not rated all movies).

By analogy, for sports contents, there are N_teams and N_opponents (in this case N_teams==N_opponents) where each N_team rates a subset in N_opponents. In this case, the matrix is square but still has many missing entries, that is to say, all teams have not played all opponents.

Metric: game_score and game_rating Uses a chessrank style system where ratings are awarded based on point differential.

Outputs: YmatrixX.npy Each "rating" matrix is written as a file in the directory structure used by split_seaons.py.

matrix_factorization.py

Input: Rating Matrix Y Ouput: P, Q matrix which optimally factorize Y given K latent features.

main.py

Calls each module

Loads all YmatrxX.npy matrices in the directory structure and procudes a completed rating matrix by finding an optimal factorization using N_latent features as the inner product dimension.

Outputs: eYmatrix.npy