The modelCan a simple model call the Palmetto Bowl?
The experiment: take a basic team-strength rating, add what we learned about rivalry hosts, lock in a pick before the 2026 game, and grade it against the result and against Vegas. The pick on this page was locked on October 17, 2026 and updates only as the inputs change.
How it works
host margin = 3.76 home field
+ 0.0436 × (host Elo − visitor Elo)
− 2.19 rivalry host penalty
Fit on 5,069 FBS regular-season games from 2007 to 2019 with a betting line, using CollegeFootballData's pregame Elo ratings. The win chance assumes the real margin lands within a normal spread of 16.5 points around the prediction, which is how far off games typically land.
About 100 Elo points is worth 4.4 points on the scoreboard. Normal home field is worth 3.8 points. In a rivalry game the model takes 2.2 of those back, which matches what the data shows: rivalry hosts win far less often than other hosts.
Does it beat Vegas?
No, not overall. The betting line is the better predictor, which is what you would expect from a market that has far more information than one rating. The rivalry adjustment does help in rivalry games, and the model's picks against the line in held-out rivalry games are promising but not proven.
| Games | Count | Model miss | Elo only | Vegas miss | Picks vs. line |
|---|---|---|---|---|---|
| All FBS games, 2007 to 2019 (fit) | 5,069 | 13.2 | 13.2 | 12.4 | 49.3% of 4237 |
| All FBS games, 2021 to 2025 (test) | 3,730 | 13.0 | 13.0 | 12.1 | 49.6% of 3158 |
| Rivalry games, 2007 to 2019 (fit) | 315 | 12.7 | 13.0 | 11.8 | 49.2% of 266 |
| Rivalry games, 2021 to 2025 (test) | 221 | 11.5 | 11.7 | 11.0 | 53.9% of 180 |
"Miss" is the average number of points between the prediction and the final margin. "Picks vs. line" counts games where the model disagreed with the line by at least a point and took the other side; a bettor needs 52.4% to break even. 2020 is left out because empty stadiums distorted home field.
Every Palmetto Bowl, replayed
For each game, the model was refit using only earlier seasons, then asked to pick. It called the winner 12 of 17 times and was closer than Vegas 8 of 12 times when a line existed.
| Year | Host | Model | Vegas | Final | Winner |
|---|---|---|---|---|---|
| 2008 | Clemson | CLEM by 6.5 | – | CLEM by 17 | Called it |
| 2009 | Columbia | CLEM by 10.3 | – | SC by 17 | Missed |
| 2010 | Clemson | SC by 3.1 | – | SC by 22 | Called it |
| 2011 | Columbia | SC by 6.8 | – | SC by 21 | Called it |
| 2012 | Clemson | CLEM by 3.1 | – | SC by 10 | Missed |
| 2013 | Columbia | SC by 5.7 | SC by 3.5 | SC by 14 | Called it |
| 2014 | Clemson | CLEM by 2.4 | CLEM by 4.5 | CLEM by 18 | Called it |
| 2015 | Columbia | CLEM by 14.9 | CLEM by 18.8 | CLEM by 5 | Called it |
| 2016 | Clemson | CLEM by 25.0 | CLEM by 25.3 | CLEM by 49 | Called it |
| 2017 | Columbia | CLEM by 17.8 | CLEM by 13.2 | CLEM by 24 | Called it |
| 2018 | Clemson | CLEM by 31.0 | CLEM by 25.9 | CLEM by 21 | Called it |
| 2019 | Columbia | CLEM by 34.4 | CLEM by 27.7 | CLEM by 35 | Called it |
| 2021 | Columbia | CLEM by 16.5 | CLEM by 11.4 | CLEM by 30 | Called it |
| 2022 | Clemson | CLEM by 11.7 | CLEM by 14.0 | SC by 1 | Missed |
| 2023 | Columbia | CLEM by 5.9 | CLEM by 7.3 | CLEM by 9 | Called it |
| 2024 | Clemson | CLEM by 0.8 | CLEM by 2.7 | SC by 3 | Missed |
| 2025 | Columbia | SC by 1.2 | SC by 2.5 | CLEM by 14 | Missed |
The 2026 test
- Vegas line
- Not posted
- Elo
- CLEM 1541 · SC 1699
- Host penalty
- Built in
- Records
- CLEM 3–2 · SC 3–3
Snapshot from Oct 17, 2026, through Week 7.
What counts as a pass: the model picks the winner, and its miss is smaller than Vegas's. After the game this card shows the final and the grade. The cron job keeps the inputs current until kickoff, so the pick can move as Elo ratings and the line move.
What it leaves out
Injuries, quarterback changes, weather and motivation. Elo only knows wins, losses and margins. One game is one data point, so a single right or wrong call says little about the model; the backtest tables above are the better guide.