What It Is
A small “bronze/silver/gold” lakehouse demo using public NFL play-by-play data (nflverse), paired with a betting-style
dashboard that overlays live lines and model projections to surface potential +EV bets (simulation-first).
What It Is Not
Not betting advice and not a guaranteed winning system. The model is intentionally lightweight and misses many real-world
factors (injuries, weather, matchup specifics). Use as a learning/analysis tool.
Model (NFL)
Team form: last N weeks (default 8), recency-weighted.
Uses team EPA/play and points for/against from gold tables.
Projected score: combines (a) points-based form and (b) rating-based margin (from EPA-to-spread regression).
Probabilities: converts projected margin/total into win/cover/over probabilities using historical σ for margin/total.
Market comparison: converts odds to implied probabilities (vig-adjusted) and computes EV/unit.
Market blend: to reduce overconfidence, recommendations use a blended probability for EV:
p_used = α·p_model + (1−α)·p_market(no‑vig) (α≈0.35 when market probabilities are available).
Data Pipeline (Bronze → Silver → Gold)
Bronze: raw play-by-play Parquet files (nflverse releases)
Silver: cleaned subset of plays + game finals
Gold: compact Parquet tables for the site:
- gold_team_weekly, gold_qb_weekly, gold_game_summary
- optional gold_player_weekly and gold_team_allowed_weekly for props/matchups
The pipeline also emits Quality checks and Lineage so you can see transformations and schemas.
Bet Sizing + Logging
Guided Picks ranks markets by EV/unit and suggests conservative stakes (fractional Kelly or flat %) for paper betting.
Use the local Bet Log to track the exact odds/line/time you took and mark results (stored in your browser only).