PICK'EM
Methodology

A model that has to keep receipts.

Pick'em separates the quantitative model from Bryan's final decision. That makes it possible to measure whether human judgment adds value or just adds Cowboys bias.

Production model

EPA Baseline v1.0.0

  • • 55% full-season net EPA/play from the 2025 regular season
  • • 30% net EPA/play across each team's final five 2025 games
  • • 10% turnover margin per game
  • • Small fixed home-field adjustment

The source is nflverse weekly team-stat data. Week 1 has no 2026 regular-season sample, so v1 deliberately starts with the complete 2025 regular season. Each prediction stores the exact feature snapshot used when it was generated.

What v1 does not claim

  • • No injury adjustment yet.
  • • No sportsbook or betting-market input.
  • • No hidden AI judgment or manual team boosts.
  • • Confidence is model separation, not a guarantee of calibrated real-world probability.

Accountability rules

  1. 1. Bryan picks publish before kickoff.
  2. 2. Published Bryan picks receive timestamps and cannot be edited.
  3. 3. Model predictions are versioned and stored with their feature snapshots.
  4. 4. Existing v1 predictions are never overwritten by a later rerun.
  5. 5. Every completed result counts.
  6. 6. Third-party benchmarks are sourced, not silently copied.
Cowboys Bias Index

Yes, this is a real metric.

The index rises when Bryan selects Dallas despite increasing model disagreement. Those overrides are tracked separately so the end-of-season record can show whether Cowboys loyalty created wins, cost wins, or somehow defeated mathematics.