Pascal Institute
Track Record
We aim for forecasts that are accurate, calibrated, and scored in public. We publish a win probability for every NFL and college football game before kickoff and score it against the closing line afterwards. This page collects those records, the retrospective studies behind the ratings, and our published 2026 World Series forecast. Weak results are shown as plainly as strong ones.
This season against the closing line
The closing line is the betting market's final price at kickoff, with the bookmaker margin removed. It is a strong forecast, and we do not claim an edge over it. These records are prospective: every probability was published before its game.
- NFL. Over 60 scored 2026 games, the rating's log loss is 0.652 against the closing line's 0.675 (lower is better), so the rating is ahead of the market so far. That is a small sample, too few games to separate skill from luck. NFL ratings and this week's games
- College football. Over 227 scored 2026 games, the rating's log loss is 0.577 against the closing line's 0.550 (lower is better), so the rating is behind the market so far. College football ratings and this week's games
NFL: Pascal Rating vs the closing line · 2026 season
Every completed game with a rating available before kickoff and a closing line, scored on the result. The closing line is the median no-vig moneyline across sportsbooks at kickoff. Lower log loss and Brier scores are better. The rating's lean is the side it rated higher than the market, settled at the closing price for one unit.
60 games scored. Not scored: 32 without a pre-kickoff rating, 1 tied or unscored.
- Log loss (lower is better)
- Pascal Rating0.652Closing line0.675
- Brier score (lower is better)
- Pascal Rating0.230Closing line0.240
- Picked the winner
- Pascal Rating61.7%Closing line55.0%
- Average gap to market
- 7.6 pp
- Rating lean at the close
- 28–32+12.05 u · ROI +20.1%
- Lean of 5 pp or more
- 14–22+3.93 u · ROI +10.9%
| Week | Games | Rating log loss | Market log loss | Lean record | Lean units |
|---|---|---|---|---|---|
| Week 3 | 16 | 0.668 | 0.699 | 9–7 | +6.41 u |
| Week 4 | 16 | 0.585 | 0.625 | 8–8 | +1.64 u |
| Week 5 | 14 | 0.648 | 0.648 | 5–9 | -0.18 u |
| Week 6 | 14 | 0.716 | 0.730 | 6–8 | +4.18 u |
Rating probabilities use calibration rating-win-probability-v2. Past results do not predict future results.
Calibration
Each game is counted once, from the side each source favored, and grouped by the probability it gave that favorite. A calibrated forecast's favorites win about as often as it said. Bins with fewer than 20 games are marked small; their win rates can swing widely.
| Favorite's probability | Pascal Rating | Closing line | ||||
|---|---|---|---|---|---|---|
| Games | Average forecast | Favorite won | Games | Average forecast | Favorite won | |
| 50% to 60% | 31 | 54.5% | 51.6% | 25 | 55.3% | 44.0% |
| 60% to 70% | 19(small) | 64.6% | 73.7% | 17(small) | 64.0% | 52.9% |
| 70% to 80% | 10(small) | 73.7% | 70.0% | 13(small) | 74.9% | 76.9% |
| 80% to 90% | 0 | — | — | 5(small) | 83.1% | 60.0% |
College football: Pascal Rating vs the closing line · 2026 season
Every completed game with a rating available before kickoff and a closing line, scored on the result. The closing line is the median no-vig moneyline across sportsbooks at kickoff. Lower log loss and Brier scores are better. The rating's lean is the side it rated higher than the market, settled at the closing price for one unit.
227 games scored. Not scored: 63 without a closing line, 220 without a pre-kickoff rating.
- Log loss (lower is better)
- Pascal Rating0.577Closing line0.550
- Brier score (lower is better)
- Pascal Rating0.198Closing line0.186
- Picked the winner
- Pascal Rating70.0%Closing line71.8%
- Average gap to market
- 6.9 pp
- Rating lean at the close
- 85–142-17.83 u · ROI -7.9%
- Lean of 5 pp or more
- 38–85-21.84 u · ROI -17.8%
| Week | Games | Rating log loss | Market log loss | Lean record | Lean units |
|---|---|---|---|---|---|
| Week 4 | 55 | 0.520 | 0.459 | 14–41 | -20.40 u |
| Week 5 | 56 | 0.515 | 0.501 | 24–32 | -11.90 u |
| Week 6 | 57 | 0.618 | 0.600 | 23–34 | +4.39 u |
| Week 7 | 59 | 0.649 | 0.633 | 24–35 | +10.08 u |
Rating probabilities use calibration rating-win-probability-v2. Past results do not predict future results.
Calibration
Each game is counted once, from the side each source favored, and grouped by the probability it gave that favorite. A calibrated forecast's favorites win about as often as it said. Bins with fewer than 20 games are marked small; their win rates can swing widely.
| Favorite's probability | Pascal Rating | Closing line | ||||
|---|---|---|---|---|---|---|
| Games | Average forecast | Favorite won | Games | Average forecast | Favorite won | |
| 50% to 60% | 75 | 54.8% | 61.3% | 53 | 55.3% | 58.5% |
| 60% to 70% | 67 | 64.7% | 62.7% | 61 | 65.4% | 62.3% |
| 70% to 80% | 43 | 74.5% | 76.7% | 50 | 75.0% | 72.0% |
| 80% to 90% | 29 | 84.7% | 86.2% | 45 | 84.4% | 91.1% |
| 90% to 100% | 13(small) | 93.4% | 100.0% | 18(small) | 93.0% | 100.0% |
Retrospective studies
These results are retrospective, not prospective. They replay today's methods over past seasons using only games completed before each target game. They show how the methods would have done, not how forecasts published at the time did, and the methods were developed in research that looked at some of those seasons.
Holdout tests against standings
Each sport's method was checked on a later chronological holdout it had not seen, against a simple baseline built from each team's pregame record. The NHL rating was validated separately: on its 2024-25 and 2025-26 holdout, the team favored by rating and home ice won 56.8% of 2,624 games.
| Sport | Games | Correlation with later margin | Higher-rated team won | Compared with standings |
|---|---|---|---|---|
| MLB | 4,823 | 0.109 | 54.6% | Mixed: higher AUC, lower correlation |
| NFL | 570 | 0.416 | 65.4% | Higher on both measures |
| College Football | 1,602 | 0.623 | 70.5% | Higher on both measures |
| NBA | 1,230 | 0.459 | 66.7% | Higher on both measures |
Football against open benchmarks
We scored the Pascal Rating against open rebuilds of two public methods, FiveThirtyEight's Elo and the Simple Rating System (SRS), on the 2024 and 2025 seasons. Each rating was turned into a probability by its own fit on 2020 to 2023, so the comparison is about the information in each rating rather than calibration settings. In the NFL, the Pascal Rating and Elo were within noise of each other and SRS was a little weaker. In college football the Pascal Rating scored better than both, but that comparison favors it: the rebuilds see only games between FBS teams and start without history in 2020. Blending the three did not improve on the Pascal Rating by more than noise.
| Rating | Log loss | Brier | AUC | Picked the winner |
|---|---|---|---|---|
| Pascal Rating | 0.626 | 0.218 | 0.708 | 66.1% |
| Elo, 538 method (open rebuild) | 0.623 | 0.216 | 0.717 | 66.3% |
| SRS (open rebuild) | 0.636 | 0.223 | 0.692 | 64.7% |
| Blend of all three, fitted weights | 0.622 | 0.216 | 0.715 | 66.6% |
| Blend of all three, equal weights | 0.624 | 0.217 | 0.712 | 66.1% |
| Rating | Log loss | Brier | AUC | Picked the winner |
|---|---|---|---|---|
| Pascal Rating | 0.544 | 0.185 | 0.788 | 70.9% |
| Elo, 538 method (open rebuild) | 0.568 | 0.194 | 0.761 | 68.9% |
| SRS (open rebuild) | 0.564 | 0.192 | 0.765 | 70.0% |
| Blend of all three, fitted weights | 0.543 | 0.184 | 0.787 | 70.5% |
| Blend of all three, equal weights | 0.550 | 0.187 | 0.780 | 70.7% |
The same study also compared the ratings with nfelo and with the betting market. Those sources have no declared license, so we do not publish their figures. Its conclusion was that nfelo's rating did about as well as ours, and that every rating, ours included, trailed the market and added no information beyond it. Differences in log loss below about 0.01 cannot be ranked on samples of this size. Source: docs/validation/rating-benchmark-ensemble.
2026 World Series forecast
On October 6, 2026, before the playoffs began, we published title odds for all 12 playoff teams from the Pascal Rating and each team's projected starting pitchers. This is a prospective forecast. Our simulation method made the Milwaukee Brewers the favorite at 22.5%; World Series futures favored the Los Angeles Dodgers at 27.6%.
The result is pending: the postseason is still being played. We will score the forecast here once the World Series is final.
What this shows and does not show
Sample sizes matter more than any single number here. An NFL regular season is 272 games, and a few dozen of them cannot separate skill from luck: a lead or deficit against the closing line can appear and vanish within a few weeks. College football has more games each week, but a season is still one season. Each record shows how many games were scored and how many were not, and why.
Log loss and the Brier score measure the quality of the probabilities, not how many winners were picked. Both reward giving the eventual winner a high probability and penalize confident misses, log loss more heavily. Lower is better on both. Calibration asks a related question: whether favorites given about 70% won about 70% of the time.
The lean record shows which side of the closing line the rating was on and how that side fared at the closing price. It is a diagnostic, not a betting record or a recommendation. Nobody placed these bets, and one season of leans is well within what luck can produce.
Past results do not predict future results. The ratings and the conversion from rating to probability are fixed in advance and versioned, but teams and leagues change, and a method that did well last season can do worse this one.
We publish game forecasts only for the NFL and college football. MLB, the NBA, the NHL, and men's college basketball have ratings but no game forecasts yet, so they have no record against the closing line. A forecast needs a tested conversion from rating to probability and a reliable record of closing prices to score it against. We are building MLB game forecasts for the 2027 season, with a 2026 backtest against closing lines first. The NBA and the NHL follow once we capture their prices. Men's college basketball ratings are published only once an approved data source is in place.
Pick of the Day is not part of this record. Those are members' picks, scored on their own leaderboards.
Published by Pascal Institute. Edited by Christopher Merry, Chief Decision Scientist.
Newsletter
Pascal Ratings Weekly
Each week, the latest team ratings, Pascal Power Rankings, and how our forecasts scored against the results. It is free, and every email has a one-click unsubscribe.
Subscribe to Pascal Ratings WeeklySigning up takes you to the Pascal Institute newsletter page. SendFox sends the newsletter and asks you to confirm by email.