This github repo contains the code for "Analytics, have some humility: a statistical view of fourth-down decision making"

0. Get the data

• run d1_data_acquisition.R
• then run d2_data_acquisition.R
• then run d3_data_TeamQualityMetrics_epa0.R
• output data7b.csv

1. Simulation study: random walk football in 1_simulation/sim_v2

• run sim2.R parallelized on a cluster via run_sim_2_1_AJ.sh and run_sim_2_2_AJ.sh
• then run sim_2_aggregate_results.R (optionally on a cluster via run_sim_2_aggregate_results.sh)

2. Fourth-down decision making in 2_Decision_Making

• models
  • tune the XGBoost first-down win probability models in T2_param_tuning_xgb.R on a cluster via T2_run_param_tuning_xgb_WP_AJ.sh
  • test the accuracy of the various first-down WP models in T3_test_wp.R
  • tune the baseline coach XGBoost model in D2_coach_decision_model_tune.R
  • bootstrap stability analysis to select $B=100$ in D4_stability_analysis.R and D4b_stability_analysis_results.R
  • fit $B=100$ bootstrapped first-down win probability models and the FG, Go, and Punt models in D5_fit_BootWPModels.R
• decision making
  • make the fourth-down decision plots (exs. 1 thru 5 in the paper) in D7_decision_making_makePlots.R
  • compare traditional decision making to ours and evaluate coaches in D8_humility.R

3. Shiny app in 3_shiny

• run app.R for an interactive fourth-down decision making Shiny app!