The Predictive Playbook – Week 4 WR Projections: Model, Calibration & Slate

Posted by:

Nathan

|

On:

|

, , ,

🔗 GitHub Repo: https://github.com/ndbryant21-eng/fantasy-football-ml
📓 Notebooks folder (main): https://github.com/ndbryant21-eng/fantasy-football-ml/tree/main/notebooks
📒 This week’s WR notebook (branch): https://github.com/ndbryant21-eng/fantasy-football-ml/blob/wr-week04/notebooks/WR_weekly_V1.ipynb

🚀 TL;DR

  • Built the WR sibling of the QB/RB pipeline: train → calibrate → project → tier.
  • Training set: 15,383 rows (WR-only, 2018–2024).
  • Backtest (2024): Linear calibration helped across the board:
    • MAE: 4.03 → 3.82 (−5.1%)
    • RMSE: 5.34 → 5.18 (−3.1%)
    • R²: 0.533 → 0.561 (+0.028)
    • Calibrator: ŷ ≈ 1.273·raw − 2.382 (lifts elite outcomes, reins in low/mid).
  • Reliability: Raw under-shot the top decile by ~−4.00 PPR; after calibration it’s ~−1.48.
  • Week 4 slate (pregame): 96 WRs (3 per team). PPR mean 10.91, 90th 17.63; tiers ≈ 22 Smash / 32 Start / 25 Stream / 17 Sit.
  • Macro vs Week 2: projections essentially flat (median Δ −0.08 PPR).

🧠 Model (WR)

Regressor: tree-based (XGBoost if available, RF fallback) + post-hoc calibration.
Features (shifted/rolled):

  • Player form: targets (L3/L6), receptions (L3), yards (L3), TDs (L6); plus air-yards/routes/snap when present.
  • Opponent context: WR fantasy points allowed (L8) by defense.
  • Environment: home/away and Vegas implied team points (from total + spread) when lines exist.

Why this mix? WR scoring is dominated by usage + leverage (targets/air) and game environment. Short rolling windows catch role shifts; opponent allowance and vegas stabilize TD/volume expectations.

🧪 Data flow & guardrails

  • Primary tables: nfl_data_py weekly + schedules/lines.
  • Early-season fallback: when current-season weekly isn’t posted, build the live pool from prior-season usage (top-3 WRs by targets per team) and derive rolling features from 2024.
  • Consistent export schema:
    player_name, team, opp_team, season, week, is_home, vegas_implied_pts, proj_fp, tier.

📏 Calibration & Reliability (2024 validation)

Before calibration: classic S-curve miscalibration—slightly high in the low/mid bins, too low at the top.
After linear calibration (fit on prior seasons):

  • Overall: MAE 4.03 → 3.82, RMSE 5.34 → 5.18, R² 0.533 → 0.561.
  • Top decile bias: −4.00 → −1.48 PPR; MAE improved from 5.70 → 4.90.
    Read: linear mapping compresses extremes helpfully, but the residual reliability curve is still a little S-shaped. Next up: isotonic calibration to flatten remaining decile bias without losing the top-end gains.

🗓️ Week 4 Outlook (pregame)

  • Population: 96 WRs (top-3 per team).
  • Distribution: mean 10.91, median 10.77, 90th 17.63, max 22.78 PPR.
  • Tiers: 22 Smash / 32 Start / 25 Stream / 17 Sit.
  • Week 4 vs Week 2: macro steady (avg Δ −0.02, median Δ −0.08 PPR); the movement you’ll see is largely player/team-specific (usage/lines/injuries), not a global calibration swing.

How to use it

  • Season-long: Prefer Start/Smash with positive vegas_implied_pts and softer WR-allowed (L8) opponents.
  • DFS: Stack Smash with QB and plausible bring-backs; use Stream for correlated leverage/salary relief.
  • Props: The model outputs PPR, but tier + projection delta is a quick screen before drilling into target share and route rate.

🔧 What changed this week (that matters)

  • Schedule mapping and Vegas implied are injected directly from the current-week schedule slice (fix for earlier home/away nulls).
  • Robust live-pool fallback so weekly runs don’t block on data lag.
  • Calibration is now baked into inference; we’ll experiment with isotonic next.

🛣️ What’s next

  1. Isotonic calibration (non-linear) for flatter decile bias.
  2. Ceiling signals for elites: end-zone/inside-10 targets, air-yards share, route participation, team PROE, QB efficiency/pressure, coverage matchups.
  3. Tier cutpoint tuning to keep a consistent “Smash/Start” count given weekly scoring climates.
  4. Auto-switch to in-season rolling features as soon as 2025 weekly data is reliably posted.

🔗 Project Links

If you’ve got ideas for opponent context or how you want tiers to look (e.g., target ~15 Smash each week), open an issue. On to the next slate. 🏈📈

Posted by

Nathan

in