Model Simulator

Pre-launch state. The right-most column will fill with graded hit rates once Sept 4 picks start landing. The backtest column shows the v2 weights' HC hit rate against 916 real bookmaker lines - that's the ground we're standing on, not a marketing claim.

Pre-launch. No NFL picks have published yet. TE receiving HC hit 63.6% across 916 real bookmaker lines in 2024 walk-forward validation - we don't market live numbers until at least 50 picks are graded.

Vegas spreads and totals tell you the implied script before kickoff. A 9-point favourite runs out the clock; a 50+ total lifts every receiver; a 38 total buries volume on both sides. We classify every game into one of five buckets: Blowout Favourite, Blowout Underdog, Shootout, Grind, or Competitive.

Each bucket carries a position-specific modifier. Blowout favourite RBs get a +15% touch boost (clock-eating); blowout underdog WRs get +15% on receiving yards (garbage-time pass volume). The modifier surfaces on every pick card on this week's picks and tints rows where the script favours the projection.

Thresholds: |spread| ≥ 7 splits blowout from competitive; total ≥ 50 promotes a competitive game to shootout; total ≤ 41 demotes it to grind. Source of truth: classify_game_script() in NFL/nfl_pick_selector.py.

Anytime touchdown is a Poisson process - the probability of scoring at least one TD is P(TD ≥ 1) = 1 − e−λwhere λ is the expected TD count for the game. We shrink each player's observed TD rate toward a position-and-role-aware prior (k = 8), then multiply by a game-total adjustment and an opponent TD-defence factor.

Role priors are calibrated from 2020-24 actual hit rates: WR1 (target share ≥ 0.22) sits at 28%; WR2 (≥ 0.15) at 20%; RB1 (feature back) at 40%; TE1 (target share ≥ 0.18) at 22%.

A.1 Honest calibration (2026-06-10). Per-role isotonic regression calibration now fit on 2023 only; 2024 is the permanent untouched holdout. Fallback chain: role → position → ALL. Calibration version stamped on every pick. Official 2024 holdout numbers (real bookmaker lines, n=261): HC 19.5% hit rate, +28.7% EV at avg +559 odds. Positive EV roles on 2024: RB (+86% EV), TE (+65% EV), WR3 (+30% EV). WR1/WR2 demoted to BET until out-of-sample WR calibration stabilises. These are the published launch figures - not the pooled in-sample numbers. Full data in NFL/data/nfl_td_real_line_brier.json.

For every (team, position, market) we compute the average yards (or TDs) allowed per game. The league average across all 32 teams is the comparison anchor. Each cell's defense_factor = team_avg / league_avg: above 1.0 means the defence allows more than average (favourable for the offence); below 1.0 means tougher than average.

The factor multiplies into every projection that involves that defence. A WR facing a defence with a 1.20 receiving_yards factor gets a 20% baseline lift; facing a 0.85 factor takes a 15% haircut. The full table renders on /nfl/defencewith heat-map cells; clicking any team opens this week's picks targeting that slice - the same data, two surfaces.

Source: nfl_team_defense table, rebuilt daily by nfl_daily_refresh.py from nfl_game_logs. Read live at /api/nfl/dvp.

DFS and same-game-parlay players don't bet props in isolation - they bet correlated combinations. When a QB throws for 300 yards, their WR1 almost certainly cleared their receiving line too (historical r ≈ 0.82). Stack Builder surfaces the model's highest-correlation pre-built stacks each week, with a combined edge and EV score that accounts for the joint distribution rather than treating the props as independent.

Pre-built stack types: QB+WR1 (correlation ~0.82), QB+WR2 (~0.61), RB+opp-WR (~0.55, the "blowout flip" stack), and game-stack (both teams' pass-catchers, ~0.48 in shootout environments).

Backing analysis lives in NFL/nfl_triple_stacks.py. The Stack Builder is available on the All-Access plan.