NFL challenger comparison
Separate regression and XGBoost forecasts for point spreads and game totals. Live picks and official results are unchanged.
Research, not a replacement model. Historical challenger forecasts were reconstructed today. They were not published before those games. Lower average point error is better; this is not a betting win rate.
Weeks 1–5: identical archived games
73 games with recoverable archived EdgeLine forecasts. Games missing original evidence are excluded.
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 73 | 10.14 | 73 | 11.23 |
| Regression challenger | 73 | 10.11 | 73 | 11.89 |
| XGBoost challenger | 73 | 10.29 | 73 | 12.03 |
Comparison with saved market lines
Only games with both an original EdgeLine line and original market line are included below. Earlier archives lack market lines; they are excluded from this separate comparison.
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 13 | 11.00 | 13 | 9.85 |
| Saved market | 13 | 11.23 | 13 | 9.77 |
| Regression challenger | 13 | 12.22 | 13 | 10.38 |
| XGBoost challenger | 13 | 12.47 | 13 | 11.39 |
Week 1: 14 archived comparisons / 16 completed games
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 14 | 11.79 | 14 | 11.89 |
| Regression challenger | 14 | 11.49 | 14 | 12.66 |
| XGBoost challenger | 14 | 11.70 | 14 | 12.51 |
Missing archived forecasts: NE@SEA, SF@LAR
Week 2: 16 archived comparisons / 16 completed games
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 16 | 12.66 | 16 | 11.38 |
| Regression challenger | 16 | 12.74 | 16 | 12.33 |
| XGBoost challenger | 16 | 13.03 | 16 | 12.34 |
Missing archived forecasts: none
Week 3: 16 archived comparisons / 16 completed games
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 16 | 7.91 | 16 | 11.59 |
| Regression challenger | 16 | 8.08 | 16 | 12.28 |
| XGBoost challenger | 16 | 8.32 | 16 | 12.15 |
Missing archived forecasts: none
Week 4: 14 archived comparisons / 16 completed games
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 14 | 7.36 | 14 | 11.25 |
| Regression challenger | 14 | 6.07 | 14 | 11.57 |
| XGBoost challenger | 14 | 5.96 | 14 | 11.68 |
Missing archived forecasts: PIT@CLE, IND@WAS
Week 5: 13 archived comparisons / 15 completed games
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 13 | 11.00 | 13 | 9.85 |
| Regression challenger | 13 | 12.22 | 13 | 10.38 |
| XGBoost challenger | 13 | 12.47 | 13 | 11.39 |
Missing archived forecasts: NYG@WAS, BUF@LAR
Forward comparison
2 completed · 14 independently saved receipts · 5 pending.
| Forecast | Spread games | Spread error | Total games | Total error |
|---|---|---|---|---|
| Archived EdgeLine | 2 | 3.00 | 2 | 16.00 |
| Saved market | 0 | — | 0 | — |
| Regression challenger | 2 | 6.09 | 2 | 17.88 |
| XGBoost challenger | 2 | 8.32 | 2 | 18.36 |
Independent immutable pre-kickoff receipts. Contemporary EdgeLine fair lines and market quotes captured with the challenger. Candidate inputs may be older; feature timestamps are retained. No official record changes.
Candidate data is within its 48-hour capture limit.
How the test was separated
Train: 2010–2022. Tune: 2023. Probability residual calibration: 2024. Historical test: 2025. Weeks 1–5 of 2026 were used only for evaluation. Features use prior scoring form, scheduled rest, home field and division.
- Retrospective reconstruction from the current version of nflverse game data; not contemporaneously published challenger forecasts.
- Past scoring form, schedule rest, home field and division only. No unverifiable historical injury, starting-QB or weather inputs.
- 2025 market baseline uses historical closing lines; 2026 paired comparison uses the saved EdgeLine market observation, which may be earlier.
- 2024 empirical residual probabilities are experimental. No matched two-sided historical prices: betting ROI is withheld.
- Training 2010–2022, tuning 2023, calibration 2024, untouched historical test 2025. 2026 Weeks 1–5 used only for evaluation.
- No automatic promotion. Historical records and official picks are unchanged.
Historical game source · CC-BY-4.0
Updated 2026-10-18T17:55:46.607615+00:00 · Full research metrics (JSON)