Fantasy Premier League Machine Learning

Repository for preparation, analysis, model fitting and prediction for fantasy premier league data.

Credit to https://github.com/vaastav/Fantasy-Premier-League for building a fantastic data resource.

Project aims

First stage of the project is to train models and perform model selection, predicting player scores for each game week of a premier league season. We are only using features that are available on https://fantasy.premierleague.com. We take a model agnostic approach to the models we use, this repo will support scikit-learn and pytorch models.

Second stage of the project is to implement heuristic algorithms that build teams within the game restrictions to maximise the expeceted points of a team for either short-term or long-term gains.

We aim to provide a platform for fast experimental iteration and model comparison using a combination of pytorch-lightning for defining standardised datamodules and dataloaders, hydra-zen for config management and boiler-plate free hyperparameter optimization with Optuna or Hyperopt. All experiments are logged in a standardised fashion using MLflow

Quick start

1. Clone repository https://github.com/vaastav/Fantasy-Premier-League which contains the raw data
2. Edit fpl_ml/user_config.py setting the directories for the repo set in (1.). Output directory where we will write processed data, and the mlruns directory where experiments will be logged to.
3. Run python run_prepare_data.py, this takes the raw data from vaastav/Fantasy-Premier-League and processes it into tabular data. In the prepared data, columns of features prefixed with X_ and the target variable is total_points
4. Run python run_train.py. Currently runs a hyperparameter grid search for RandomForestRegressor and GradientBoostedRegressor.
5. Above your mlruns directory run mlflow ui. This will launch a local server for visualising the experiments that have been run.

Example output

Example below shows performance on validation set for predicting player scores for the 2019-2020 season. Each datapoint is the number of points a player got in a given game week. The purple line is the identity (x == y). Datapoints can come from any season. Future efforts will hold out a season for the validation set and a season for the test set.

Current features

• Data preparation
• Train sklearn models
• Logging with mlflow
• Hydra-zen configs

TODO:

• Unittesting
• Pytorch support
• Heueristic simulations for building teams
• Containerisation