Skip to content

Files

Latest commit

 

History

History
164 lines (117 loc) · 4.99 KB

README.md

File metadata and controls

164 lines (117 loc) · 4.99 KB

This is a project to choose NCAA brackets for me, using historical data fed through a TensorFlow classifier.

All of the Python scripts in this repo have help info accessible with --help.

See also the TODO list.

Setup

To setup, install Python 3 and then run:

./setup.sh

This will setup a virtualenv, install the requirements from pip, and activate the direnv .envrc (if applicable).

If you don't have direnv, you will need to source bin/activate.

If you have a CUDA GPU, you might want to run pip install tensorflow-gpu at this point to get the GPU-enabled version instead.

You might also want to consider building TensorFlow from source, since the default package doesn't have SSE enabled.

Getting data

This repo comes with data for 2002-2017 in the csv folder. If you need more data:

./fetch_csvs.py get_schools
./fetch_csvs.py get_games -y $YEAR
./fetch_csvs.py get_players -y $YEAR

The loaders try to cache as much as possible since the NCAA site is really slow, so if you want to update a CSV, you need to delete it, then run ./fetch_csvs.py with the relevant command and year.

See the README in the csv folder for details.

Training a model

./train.py -y $TRAINING_YEARS -p $VALIDATION_YEAR -o $MODEL_OUT

Example:

./train.py -y 2002,2003,2004 -p 2005 -o model_2002-2004

See other training options with ./train.py --help. You may also want to change the hidden layers, but this currently isn't a command-line option and you will need to edit the code.

This trains a Keras classifier using player stats to predict which team will win in a matchup.

Evaluating a model

./evaluate.py can be used to evaluate a trained model against a given year's data. You should also see the training set accuracy during training at the end of each epoch.

Predicting games

Once you have a trained model, open predict.py and edit the bracket at the top. TODO: Make this a config file or something.

The format is a tree of tuples where (a, b) means that a will play against b. If a or b are tuples, then we will first determine the winner of that match up, and then use the winner for the next one.

Example:

(
  ("Team A", "Team B"),
  ("Team C", "Team D"),
)

This means that the first round is Team A vs Team B and Team C vs Team D, and the second round will be the winners of the first round.

Each string needs to be the name of a team from the "Schools" dropdown in the NCAA stats page. An easy way to find these is to look through the ncaa_schools.csv file.

Once you've added your bracket:

./predict.py -m $YOUR_MODEL -y $YEAR

And you should get output like:

Villanova vs Mt. St. Mary's: Villanova wins
Wisconsin vs Virginia Tech: Wisconsin wins
Villanova vs Wisconsin: Villanova wins
Virginia vs UNCW: Virginia wins
Florida vs ETSU: Florida wins
Virginia vs Florida: Virginia wins
Villanova vs Virginia: Villanova wins

For filling in a bracket, add --wait to the command for it to stop after each line and wait for you to hit enter.

Predicting scores

The neural network version of this doesn't really work right now, so ./predict_score.py uses a simpler algorithm based on the model's historical scores.

To run:

./predict_score.py -y $YEAR "Team a" "Team b"

Where $YEAR is the year to use for the prediction (most likely this is the year before the year you want to predict, since the year you want to predict presumably doesn't have win/loss data).

Each team parameter needs to be the name from the "Schools" dropdown in the NCAA stats page. An easy way to find these is to look through the ncaa_schools.csv file.

The algorithm uses:

  • The average score for every team in the given year. This gives us a base prediction for how many points that team will score in a game.

  • For the two teams we're considering, the mean of the difference between their opponent's mean score that season and their opponent's mean score against them.

    This tells us how the opponent team will effect the score (i.e. if the opponent is not very good, our team will score more points than usual, vice-versa).

To predict each team's score, take its mean score and add the other team's mean effect on the score.

For example, in 2016, Villanova's average score was 76.5 points, and the average team scored 8 points fewer against them. Kansas's average was 71.1 points and the average team scored 3.9 points fewer against them.

Because of this we predict that in Villanova vs Kansas, Villanova will score 72.6 points (76.5 - 3.9) and Kansas will score 63.1 points (71.1 - 8).

$ ./predict_score.py -y 2016 Villanova Kansas
Historical prediction: Villanova 72.6 to Kansas 63.2 (total: 135.8)

This algorithm could use work but it comes up with plausible numbers.