Win probability · scoutindex-winprob-v1
The Elo rating going into any game only uses games played before it, so scoring the model across all 1999–2025 games is a fair walk-forward test — not a fit-and-score-the-same-data backtest.
Straight-up
64.2%
7,261 games · coin flip = 50%
Brier score
0.221
12% better than a coin flip
Log loss
0.631
coin flip = 0.693
Margin error
10.5 pts
bias -0.09
2025 season: 64.4% straight up, 0.223 Brier over 284 games. About 66–67% straight-up is the practical ceiling for any pregame NFL model.
Calibration
When the model says a team has a 70% chance, does it win about 70% of the time?
| Predicted | Games | Model | Actual |
|---|---|---|---|
| 50–55% | 1,443 | 52.5% | 52.0% |
| 55–60% | 1,370 | 57.4% | 58.8% |
| 60–65% | 1,259 | 62.5% | 60.9% |
| 65–70% | 1,125 | 67.4% | 66.4% |
| 70–75% | 838 | 72.6% | 72.1% |
| 75–80% | 615 | 77.4% | 75.8% |
| 80–85% | 397 | 82.3% | 82.6% |
| 85–90% | 172 | 87.1% | 87.8% |
| 90–95% | 41 | 91.5% | 90.2% |
By season
| Season | Games | Straight-up | Brier | Margin MAE |
|---|---|---|---|---|
| 2025 | 284 | 64.4% | 0.223 | 10.1 |
| 2024 | 285 | 68.8% | 0.210 | 10.1 |
| 2023 | 285 | 59.3% | 0.233 | 10.5 |
| 2022 | 282 | 63.1% | 0.223 | 9.1 |
| 2021 | 284 | 60.2% | 0.232 | 11.2 |
| 2020 | 268 | 65.7% | 0.217 | 10.1 |
| 2019 | 266 | 63.5% | 0.224 | 10.5 |
| 2018 | 265 | 62.6% | 0.222 | 10.3 |
| 2017 | 267 | 65.5% | 0.217 | 10.6 |
| 2016 | 265 | 64.9% | 0.218 | 9.4 |
| 2015 | 267 | 67.0% | 0.225 | 10.2 |
| 2014 | 266 | 68.4% | 0.208 | 11.3 |
| 2013 | 266 | 65.8% | 0.215 | 10.0 |
| 2012 | 266 | 63.2% | 0.219 | 11.1 |
| 2011 | 267 | 64.8% | 0.213 | 10.7 |
| 2010 | 267 | 61.8% | 0.232 | 11.0 |
| 2009 | 267 | 69.3% | 0.207 | 11.6 |
| 2008 | 266 | 63.9% | 0.218 | 11.1 |
| 2007 | 267 | 65.2% | 0.215 | 11.0 |
| 2006 | 267 | 58.1% | 0.234 | 10.8 |
| 2005 | 267 | 66.7% | 0.211 | 10.4 |
| 2004 | 267 | 63.7% | 0.222 | 10.6 |
| 2003 | 267 | 64.8% | 0.221 | 10.9 |
| 2002 | 266 | 59.4% | 0.226 | 10.5 |
| 2001 | 259 | 65.6% | 0.227 | 10.4 |
| 2000 | 259 | 64.9% | 0.222 | 10.7 |
| 1999 | 259 | 61.8% | 0.222 | 10.5 |
Player projections · scoutindex-projection-v1
Projections are generated one week at a time and scored after the games play. Once the current season is under way this table fills in with, per stat: mean absolute error, how it compares to just guessing the positional average, forecast bias, and how often the real result landed inside the stated 10th–90th percentile band (a calibrated model hits ~80%).
Playoff odds · scoutindex-playoffs-v1
The playoff simulation is a season-long forecast, not a per-game prediction, so it can't be scored the same way. The honest check is end-of-season: did the teams it favored actually make it, and were the stated probabilities calibrated across many seasons. That needs weekly snapshots the model only started keeping this season — so this section is a placeholder until there's a full season of history to grade. Method: /methodology.
Recomputed nightly with the rest of the data. This week's predictions · Playoff picture · Methodology · Not for wagering — see Terms.