Codebase for A statistical view of expected points models in American football

Get play-by-play data

• download data4.csv from https://upenn.box.com/s/kp43egvarlv6bsgbxuchnbc9fcj9iave and put in data/ folder
• you can also generate the dataset from code:
  • enter data folder
  • run d1_data_acquisition.R -> output data_nflFastR_pbp_1999_2022.csv
  • run d2_data_acquisition.R -> output data2.csv
  • run d3a_data_TeamQualityMetrics_epa0HyperParamTuning.R to tune the hyperparameters for the 8 hand-crafted team quality metrics
  • clear the environment workspace and run d3b_data_TeamQualityMetrics_epa0.R -> output data3.csv, which contains one initial train/test split column and 8 hand-crafted team quality metrics built from EPA0 fit from this initial training set.
  • run d4_data_drives.R -> output data4.csv

Model comparison

• enter model_comparison folder
• tune XGB (XGBoost) params: run param_tuning.R parallelized on a cluster via run_param_tuning.sh, then transfer the outputted .yaml files (which store the tuned params) from the folder param_tuning_results into the folder param_tuning_results_FINAL
  • the saved .yaml files that store the tuned XGB hyperparameters should already be in param_tuning_results_FINAL
• evaluate EP models (prediction accuracy): run eval_EP_models.R (on a cluster via run_eval_driveEP_models.sh) -> output FIXME
• train and save models on the full dataset: FIXME

Plots/visualizations

• enter plotting folder
• run A_plot_EP.R to visualize EP models
  • Before visualizing XGB models, need to train and save full XGBoost models via model_comparison/train_full_models.R; some of these models should already be saved in the Github
• run A_plot_team_quality.R to visualize our hand-crafted team quality metrics
• run A_plot_selection_bias.R to visualize selection bias induced by not adjusting for team quality
• run A_plot_summary_stats.R to visualize some data summary statistics