🏈 NFL Analytics

A personal NFL analytics platform that runs entirely on your machine: a local warehouse of every NFL play since 2007, a visual dashboard, auto-updating data and news, a prediction-market price tracker, and an AI analyst you can ask questions in plain English.

No API keys. No subscriptions for the core experience. One ~2 GB download.

What's inside

Piece	What it does
Warehouse (DuckDB)	909k+ plays (2007–2025), player/team stats, rosters, injuries, officials, draft history, and the full 2026 schedule with betting lines — all queryable in milliseconds
Jarvis UI (React + FastAPI)	A dark, glowing "command center": division constellation with all 32 team logos → per-team HUDs in team colors (stat rings, efficiency charts, roster, coach lineage) → live prediction-market board → built-in streaming AI chat (Ctrl-K)
Dashboard (Streamlit)	The simpler original UI: division standings → team pages → players, league leaders, schedules & lines
News engine	Auto-polls ESPN plus all 32 official team websites every 6 hours; headlines are tagged to players/teams in the warehouse
Kalshi tracker	Records prediction-market prices (game winners, spreads, totals, win totals, Super Bowl futures) every 6 hours, building line-movement history
Prediction model	Opponent-adjusted EPA ratings → win probabilities, honestly backtested against 18 years of closing lines (spoiler: Vegas wins — the model's value is calibration, and the report shows exactly by how much)
AI analyst	A chat page (and MCP server for Claude Code/Desktop) that writes and runs real SQL against your warehouse to answer questions like "which QBs perform best traveling east?"

Quick start

See SETUP.md for the full guide. The short version:

git clone https://github.com/parthakker/nfl-analytics.git
cd nfl-analytics
pip install -e .
python scripts/refresh_data.py --bootstrap   # ~2 GB from nflverse, one time
python -m streamlit run dashboard.py

Architecture

nflverse releases ─┐  (nightly-updated public data)
ESPN + team sites ─┼─► scripts/refresh_data.py / poll_news.py / snapshot_kalshi.py
Kalshi API ────────┘            │ (scheduled: weekly / 6h / 6h)
                                ▼
              nfl.duckdb + news.duckdb + kalshi.duckdb
                                │
              ┌─────────────────┼──────────────────┐
              ▼                 ▼                  ▼
        dashboard.py     MCP server (15 tools)   model/
        (Streamlit UI)   (Claude Code/Desktop)   (ratings, backtest)

Design principles: compute, don't retrieve (questions are answered by SQL over plays, not by searching documents); verified semantic layer (every data gotcha — and NFL data has many — is documented in docs/dictionary/ and enforced in CLAUDE.md); honest modeling (walk-forward backtests with an untouched holdout, reported even when the answer is "the market is better").

Data credits

All stats data from the outstanding nflverse project. News from ESPN and official team site feeds. Market data from Kalshi's public API. This is a personal, non-commercial project; all data remains property of its respective owners.

License

MIT — see LICENSE.