NFL 4th Down Coach Tendency Model (fourth-down-optimizer)

This project predicts what an NFL coach is likely to call on 4th down — punt, field goal, or go for it — based on game situation context (distance, field position, score, quarter, game clock, season) plus the offensive coach. It includes an interactive Streamlit app that outputs P(go) and a “coach effect” vs a baseline.

---

Demo (Streamlit)

The app takes:

• Yards to go (ydstogo)
• Field position (yardline_100)
• Quarter
• Score differential (Offense − Defense; + means offense leads)
• Game clock (minutes:seconds remaining)
• Season
• Offensive coach

It outputs:

• Coach likelihood: P(go), P(punt), P(field_goal)
• Coach effect vs baseline: coach P(go) − baseline P(go)
• Optional expanders:
  • League tendency (situation-only model)
  • Value benchmark (expected WPA by action; exploratory)

---

Results (temporal holdout)

Trained on 2016–2022 and evaluated on 2023–2024 (neutral situations):

• Accuracy: 0.8796
• Macro F1: 0.8339
• Log loss: 0.3217

---

Data

NFL play-by-play data (2016–2024) is cached locally as a parquet file: data/pbp_2016_2024.parquet.

Neutral situation filter (used for modeling & scorecards):

• Quarter: Q2–Q4
• Time remaining: > 2:00
• Score differential: within ±10

The app can still be used outside of neutral situations, but it shows a warning because predictions may be less reliable.

---

Quickstart (Windows PowerShell)

1) Create and activate a virtual environment

py -m venv .venv
.\.venv\Scripts\Activate.ps1

2) Install dependencies

pip install -r requirements.txt

3) Run the full pipeline

.\run_all.ps1

4) Launch the app

.\run_app.ps1