Change Log

Every change to the model and the site, newest first. A change applies from the next update on. It never rewrites a week that was already published, so the record always shows what the model said at the time.

Suggested by Werner Barz on Facebook. He asked whether the model uses this season’s player stats or only last season’s. Answering that turned up a blind spot: the model couldn’t know a starting quarterback was out until after the game.

• Picks move to Wednesday night. Picks for games after Wednesday 10 PM Eastern now go up Wednesday at 10 PM, after the official SEC, Big Ten, Big 12 and ACC availability reports come out. Sunday still posts the rankings, odds, brackets, and picks for the midweek games.
• Starting quarterbacks listed Out. If a team’s regular starter (3+ starts, including its last game) is listed Out, the team’s rating for that game drops 1 to 4.5 points, more for a quarterback who has played well. The pick says so.
• Why only quarterbacks. On 2015–2025 games, a backup’s first start left the model 4.3 points too high on that team. A missing running back, receiver or defender moved results by less than a point, too little to tell from chance.
• Tested on past seasons. On the 2015–2025 games it applied to, the model picked 76.3% of winners instead of 75.1%.
• Tested and left out: adjusting after a quarterback change shows up in the box scores. By then the ratings have already caught up, and it added nothing.
• The first Wednesday is Week 9 (Oct 28). Through Week 8, picks were made Sunday without injury reports.

• Weekly game picks for every Top 25 team, graded against the final score and the betting favorite on a season-long Pick Record. Weeks 1–7 were reconstructed using only results known before each week, and they’re labeled that way.
• A locked record. Once a week’s rankings, odds, brackets or picks are published, they never change. The model can improve; what it already said can’t be rewritten.
• Two 24-team formats beside the 12-team bracket. The CFP plan has the Power 4 champions, the top Group of Six champion and 19 at-large teams. All Champions has all ten conference champions and 14 at-large teams. Neither protects Notre Dame.
• Seven pages, automatic Sunday-night updates through the national title game, and a new home at robranksbetterthanyou.com.

CFP Finish Odds went live after Week 7: every FBS team’s chance of finishing at each spot in the final College Football Playoff committee ranking, from 10,000 simulated seasons.

Built in October 2026 on games through Week 6. Every piece was tested on the 2014–2025 seasons, and each season was predicted by a version fit without it.

• Results only, to start. Team ratings came from this season’s games alone: point margin plus per-play efficiency, adjusted for opponent and home field. The rest of the season was simulated 10,000 times, and each simulated season was ranked by a model of the committee trained on every final CFP ranking. At Week 6 it beat the AP poll but trailed the betting market.
• Start from the roster. Each team’s starting rating comes from last season’s production of this season’s players, wherever they played (transfers included), plus recruiting talent. It counts about as much as six games and fades as results come in. It closed most of the gap to the betting market and improved top-12 predictions in 9 of 10 seasons.
• The committee’s preseason habit. Hold the results equal, and the committee still ranks preseason favorites higher. Modeling that improved Week 6 top-4 predictions in all 10 seasons. The weekly polls themselves were tested and added nothing.
• Tested and left out: last season’s team rating (the “helmet”). It helped pick the top 4 but hurt the 5–12 range, which is where the 12-team field is decided.
• Tested and left out: disagreement between the AP and Coaches polls as a sign of uncertainty. Teams the voters split on were no harder to predict.