fantasy

Predict player points per game (half ppr) using stacked ensemble machine learning, leveraging Sports Reference player statistics from the previous 20 years.

Reproducibility

To reproduce this work, clone the repository, navigate to the root directory, and download/activate the environment:

git clone https://github.com/bscod27/fantasy.git
cd fantasy
conda env create -f environment.yml
conda activate fantasy 

Results can be reproduced by executing scripts/scraper.py >> scripts/wrangle.R >> scripts/train_baslearner.py >> scripts/train_metalearner.py >> scripts/generate_predictions.py:

1. scripts/scraper.py - web scrapes Sports Reference player statistics from the last 20 years and writes it as data/data_[t-20]_[t].csv; requires a command-line argument specifying the current year, t
2. scripts/wrangle.R - engineers lag 1 features by player and partitions the data into data/traindat.csv (training data with labels) and data/newdat.csv (most recent year statistics to use as inputs for predictions)
3. scripts/train_baselearner.py- trains several lower-level base learner using 5-fold groupwise (by year) cross-validation, holding out the most recent 4 years as a test set; requires a command-line argument specifying a model type within [enet, knn, svr, rf, xgb]; writes the following files:
   • ensemble/train_base_[model].csv - transformed train set (stacked out-of-fold predictions)
   • ensemble/test_base_[model].csv - transformed test set (averaged across models trained in each fold)
   • ensemble/models_base_[model].pkl - pickled models trained in each fold
4. scripts/train_metalearner.py - trains an Elastic Net model on the transformed feature space and prints the model performance; saves the pickled meta learner as ensemble/meta_learner.pkl
5. scripts/generate_predictions.py- leverages the saved models to generate predictions on data/newdat.csv; writes these predictions as preds/predictions.csv