Elevation Change prediction model
The Bot
Our own numbers, built from scratch. No borrowed odds, no network power index. Every FBS team rated, every game picked before kickoff, and every miss kept on the record.
Net rating = points better than an average FBS team on a neutral field. Offense is points scored above average, defense is points prevented. Subtract two teams' net ratings for a neutral-site spread. These measure strength for predicting games, not who has earned the most. For a résumé ranking of all FBS, see the National Index.
The Bot is a two-stage model that reruns itself automatically after every weekend of games, with updates through the week. It doesn't read betting lines, ESPN's FPI or anyone else's rankings. It only sees what happened on the field.
1. Power ratings
- Offense and defense, separately. Every team gets a points-per-game rating on each side of the ball, adjusted for every opponent it has played and every opponent those opponents have played.
- Efficiency, not just the scoreboard. Each game is scored partly on what the box score says a team deserved: success rate, yards per play and explosive plays, with garbage time thrown out. A fluky fourth quarter moves a rating less than a dominant three hours.
- Recent games count more, and blowouts are trimmed so running up the score isn't rewarded.
2. Preseason priors
- Before anyone plays, a team starts from last season, two seasons ago and its longer-run program level, pulled toward its current conference's average (so realignment is handled honestly).
- Returning production and the transfer portal adjust that start: how much of last year's passing, rushing, receiving and defensive playmaking is back, and how much arrived from other rosters.
- The prior fades as real games pile up. By midseason a team is rated almost entirely on this year.
3. The game model
- The rating gap becomes a predicted margin, adjusted for home field, altitude (a visitor climbing 7,000 feet to Laramie or Air Force is a measurable edge), travel distance, time zones and rest.
- Totals use both teams' offense and defense ratings, the season's scoring environment, tempo and run/pass mix (so a service-academy game projects lower).
- Win probability comes from how far real results land from the prediction. When the Bot says 70%, the favorite has won about 70% of the time.
4. Season simulations
- Every remaining game is played 10,000 times. Each simulated season first decides how good every team really is, so a team that's secretly better wins several more games, not just one.
- Conference title games are seated from the simulated standings, champions are crowned, and and the 12-team playoff field is picked: the SEC, Big Ten, ACC and Big 12 champions, the highest-ranked Group of Six champion, Notre Dame if it finishes in the top 12, then at-large teams. The committee's ranking is estimated from record first, then how good a team really is, with credit for winning the conference.
5. Keeping it honest
- Every number in the backtest comes from a walk-forward test: for each week of each season since 2010, the model only knew what had happened before that week.
- Picks lock at kickoff. The record can't be rewritten after the fact.
- Data: game results and play-by-play from the public CollegeFootballData.com archive. Betting lines are used only to grade the Bot, never to make its picks.