Track Record
2026 so far
pre-game predictions, scored once games are finalWeeks 1–4 were played before this model launched; their predictions were made afterward using only information from before each week, and are included here.
Are the win chances honest?
dots near the line mean yesWinners picked by week
How Vegas gaps have done
backtest, model’s side vs. the closing line, pushes excluded| Gap size | Games | Record | Model side won |
|---|---|---|---|
| 0 to 3 pts | 2,444 | 1,183-1,261 | 48.4% |
| 3 to 5 pts | 625 | 318-307 | 50.9% |
| 5 to 7 pts | 157 | 90-67 | 57.3% |
| 7 to 10 pts | 39 | 21-18 | 53.8% |
| 10+ pts | 8 | 3-5 | 37.5% |
| Gap size | Games | Record | Model side won |
|---|---|---|---|
| 0 to 2 pts | 1,894 | 949-945 | 50.1% |
| 2 to 4 pts | 1,067 | 527-540 | 49.4% |
| 4 to 6 pts | 312 | 155-157 | 49.7% |
| 6 to 9 pts | 54 | 28-26 | 51.8% |
| 9+ pts | 1 | 0-1 | 0.0% |
52.4% is the break-even rate at standard odds. A gap means the model disagrees with Vegas, not that it is right; small samples swing a lot.
Backtest
3,359 games, 2014 to 2026, each season predicted using only earlier seasons| Model | Avg miss, margin | BEAVER side |
|---|---|---|
| EnsembleIn use | 10.15 | 49.3% |
| Ridge | 10.17 | 49.8% |
| Elo baseline | 10.27 | 50.2% |
| LightGBM | 10.29 | 51.1% |
| Vegas closing line | 9.92 | n/a |
Average miss is points between the prediction and the real result. Brier score measures win-chance accuracy: 0.25 is a coin flip, lower is better.
Average miss by season
Weeks 1–4 were played before the touchdown model launched; their predictions were made afterward using only information from before each week, and are included here. Results are graded for players who played.
BEAVER Best Bets
2–5 anytime TD picks a week: chance 50%+ and at least 5% above his usual, at most 1 per team| Week | Picks | Scored | |
|---|---|---|---|
| Week 1 | ✓ Derrick Henry 59%✓ Jonathan Taylor 53% | 2 of 2 | Board |
| Week 2 | ✓ Christian McCaffrey 60%✗ Javonte Williams 57% | 1 of 2 | Board |
| Week 3 | ✓ Jahmyr Gibbs 64%✓ Christian McCaffrey 63%✓ Javonte Williams 54% | 3 of 3 | Board |
| Week 4 | ✓ Christian McCaffrey 61%✓ Jonathan Taylor 59%✓ Kenneth Walker III 58%✗ Cam Skattebo 53%✗ Chase Brown 51% | 3 of 5 | Board |
| Week 5 | ✓ Javonte Williams 57%✗ Bijan Robinson 56% | 1 of 2 | Board |
| Week 6 | ✗ Kyren Williams 67%✓ Javonte Williams 65%✓ Christian McCaffrey 64%✓ James Cook 62%✗ Bijan Robinson 59% | 3 of 5 | Board |
| Week 7 | ✓ Christian McCaffrey 67%✓ D'Andre Swift 56%✗ Woody Marks 53% | 2 of 3 | Board |
| Week 8 | · Derrick Henry 58%· Josh Jacobs 54%· Javonte Williams 53% | pending | Board |
| Week 9 | · Adam Prentice 3% | pending | Board |
Weeks 1, 2, 3, 4: picks made afterward with only earlier information (before the touchdown model launched).
Anytime TD
a rushing or receiving touchdownThis season: dots near the line mean the chances are honest. Small samples swing.
| Backtest, log loss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER TD modelIn use | 0.4035 | 0.3834 |
| Usage only (carries and targets) | 0.4113 | 0.3935 |
| Player’s own TD rate | 0.4199 | 0.4037 |
| Position average | 0.4539 | 0.4419 |
Log loss measures how good the chances are; lower is better. The first period was used for tuning, the second only to confirm. ▲ = 12.5%+ above his usual chance, usual 15%+, at most 2 per team per game.
First TD
the game’s first touchdown, any teamThis season: dots near the line mean the chances are honest. Small samples swing.
| Backtest, log loss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER first-TD modelIn use | 2.9361 | 2.8324 |
| Straight share of expected TDs | 2.9349 | 2.8465 |
| Share of the anytime chance | 2.9339 | 2.8383 |
| Every player equal | 3.2405 | 3.2514 |
Log loss per game: the chance given to whatever happened, including defensive or special-teams first TDs and games with none. Even the top pick in a game scores first only about 1 time in 7. ▲ = 20%+ above his usual chance, usual 5%+ (tested, not shown on the site: they didn’t hold up in 2023–25).
2+ TDs
two or more rushing or receiving touchdownsThis season: dots near the line mean the chances are honest. Small samples swing.
| Backtest, log loss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER 2+ TD modelIn use | 0.1114 | 0.0993 |
| Simple formula from the anytime chance | 0.1125 | 0.0994 |
| Position average | 0.1287 | 0.1180 |
Lower log loss is better; 2+ TDs is a rare event (about 3% of player-games), so the numbers are small. ▲ = 20%+ above his usual chance, usual 5%+, at most 2 per team per game.
Touchdowns are noisy: even a 50% player fails half the time. Touchdown props usually carry a large sportsbook margin, first-TD and 2+ props most of all, so a good matchup is not the same as a good bet.
Weeks 1–4 were played before the player-props model launched; their projections were made afterward using only information from before each week. Graded for players who played.
Passing Yards
starting quarterbacks| Backtest, average miss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER projectionIn use | 60.0 | 58.6 |
| Recent weighted average | 66.9 | 65.9 |
| Season average so far | 70.4 | 69.7 |
| Last 3 games | 72.7 | 71.4 |
| Likely range held (50% / 80% is right) | 50% / 80% | 51% / 82% |
| Milestone | Predicted | Happened |
|---|---|---|
| 200+ yds | 66.4% | 66.8% |
| 250+ yds | 42.9% | 41.0% |
| 300+ yds | 21.5% | 19.7% |
2019-2022: ▲ beat usual 55% (274) vs. 43% overall · 2023-2025: ▲ beat usual 55% (133) vs. 44% overall
Rushing Yards
running backs and starting quarterbacks| Backtest, average miss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER projectionIn use | 15.8 | 15.9 |
| Recent weighted average | 16.8 | 16.9 |
| Season average so far | 17.1 | 17.1 |
| Last 3 games | 17.7 | 17.8 |
| Likely range held (50% / 80% is right) | 50% / 80% | 52% / 80% |
| Milestone | Predicted | Happened |
|---|---|---|
| 50+ yds | 19.5% | 20.5% |
| 75+ yds | 9.0% | 9.9% |
| 100+ yds | 4.0% | 4.6% |
2019-2022: ▲ beat usual 58% (434) vs. 50% overall · 2023-2025: ▲ beat usual 57% (567) vs. 50% overall
Receptions
receivers, tight ends and running backs| Backtest, average miss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER projectionIn use | 1.20 | 1.12 |
| Recent weighted average | 1.27 | 1.20 |
| Season average so far | 1.29 | 1.21 |
| Last 3 games | 1.33 | 1.25 |
| Likely range held (50% / 80% is right) | 53% / 83% | 55% / 85% |
| Milestone | Predicted | Happened |
|---|---|---|
| 3+ catches | 31.7% | 32.1% |
| 5+ catches | 13.9% | 14.1% |
| 7+ catches | 5.1% | 5.4% |
2019-2022: ▲ beat usual 51% (102) vs. 50% overall · 2023-2025: ▲ beat usual 50% (4) vs. 50% overall · no ▲ flags shown for receptions
Receiving Yards
receivers, tight ends and running backs| Backtest, average miss | 2019-2022 (tuning) | 2023-2025 (confirmation) |
|---|---|---|
| BEAVER projectionIn use | 15.0 | 14.0 |
| Recent weighted average | 16.3 | 15.3 |
| Season average so far | 16.7 | 15.4 |
| Last 3 games | 17.2 | 16.1 |
| Likely range held (50% / 80% is right) | 53% / 82% | 56% / 84% |
| Milestone | Predicted | Happened |
|---|---|---|
| 50+ yds | 14.9% | 15.5% |
| 75+ yds | 6.8% | 7.3% |
| 100+ yds | 3.1% | 3.2% |
2019-2022: ▲ beat usual 55% (135) vs. 50% overall · 2023-2025: ▲ beat usual 60% (68) vs. 49% overall
Backtest: every season 2019–2025 projected using only earlier seasons (2019–22 used for tuning, 2023–25 only to confirm). The projection is the typical game (median); the miss is the gap between it and what happened.