Why look past the score
The final score tells you who won. It does not tell you whether the winner played well, because a game of sixty or seventy plays turns on a handful of them: a tipped pass, a fumble at the goal line, a fourth-down stop. The numbers on the team pages look at every play instead, and ask how each one went. Most of the time they agree with the scoreboard. When they do not, that is the interesting part.
The projected margins
Before a game, the top of a team page shows up to three models' projected margins side by side, with the pregame line beside them for comparison. Each model is one method's estimate of how the game will go, and when they disagree that is worth knowing. Above them sits one prediction, Our call, the average of the three models.
SP+ is a rating of how good a team is, expressed in points better or worse than an average team, built from its play-by-play and tempo. Because the ratings are already in points, the gap between two teams' ratings is a projected margin.
Elo is a rating that goes up with every win and down with every loss, more for beating a strong team and less for beating a weak one. It knows only who beat whom and by how much. The gap between two ratings is divided by 25 to turn it into points.
The line is the pregame line, the margin the market expected. The site shows it from the page team's side, so a negative number means that team was the underdog. It is not a model and is never counted as one: it is shown for comparison only, as a margin benchmark. It never counts when the site says how the models line up or how often a model named the winner, and the site gives it no win chance.
Our model is this site's own experimental rating. It uses each team's play-by-play, adjusted for who it has played, blended with a preseason estimate built from last season, returning players and recruiting. The card says prior 67% when 67 percent of the rating is still that preseason estimate; the share shrinks every week as real games replace it.
In SP+, Elo and our model the home team gets 2.5 points; a neutral site gets nothing. The line needs no adjustment because the market already priced the venue in.
Our call
Our call is the site's one prediction for a game: the simple average of the SP+, Elo and our model projected margins, naming the team it favors and by how much. The pregame line is deliberately left out. It stays a benchmark, so the call can be measured against the market instead of borrowing from it. The call is made only when all three models have a number for the game; otherwise the page says when the call will post (our model's number for a week is set that Monday morning) or why the game gets none, such as an FCS opponent that SP+ and Elo do not rate.
The label is the size of the average: Toss-up under 3 points, Lean from 3 to 7, Strong at 7 or more. One line under the call says how the three models line up and then gives the line on its own, for example "SP+ favors Georgia; Elo and Our model favor Alabama. The line, shown for comparison only, has Georgia by 2.5."
Like every model, the call is fixed at kickoff and scored afterwards: its average miss and how often it named the winner, next to each model on the same games, on the Projections page and in each week's review.
Example. SP+ has the home team by 4.0, Elo by 1.0, our model by 7.0: the average is 4.0, so Our call is the home team by 4.0, a Lean, whatever the line says.
Track record: avg miss and closest
Under each card, once a model has 20 finished games behind it, two lines show how it has done this season. avg miss 9.8 means that model's projections have been 9.8 points from the actual margin, on average. A miss is the distance between the projected margin and the real one, whichever direction. closest 10/39 means that of the 39 games that have finished, this model came closest to the real margin in 10; when two models tie, both get credit. The line, a benchmark, shows its average miss only and is never counted as closest.
Only games where every model existed before kickoff are counted, so each model is scored on the same games. The Projections page shows the same numbers for every game, week by week, and the season scoreboard at the bottom of this page shows the totals.
After a game, the line under the score on a team page, and each finished game's card on the Projections page, mark each model ✓ when it named the winner and ✗ when it did not (the line, a benchmark, gets no mark), and name the model that came closest among the models present, never the line (Closest:), our model included when it had a projection before kickoff; an exact tie goes to the first listed. Closest is a per-game read. The counts, "closest 10/39" and the season tables, keep the accuracy report's scoring of SP+ and Elo, where a tie credits both, so a game's Closest and its count can differ.
What a good number looks like. Even the pregame line, the best public number there is, misses the real margin by about 12 points on average over a season. A model that misses by 10 is doing well; the point is not the miss but which model misses least, and by how much.
Play value: PPA and EPA
Every spot on the field is worth something. An offense at its own 20 will, on average, come away with a little under a point on that drive; an offense at the opponent's 5 will come away with nearly six. Those averages, learned from thousands of past drives, are the expected points of a situation.
A play's value is simply how much it changed that number. A 9-yard run on third-and-8 turns a likely punt into a fresh set of downs, so it adds expected points. A sack on first down takes them away. Add up the values of a team's plays and you get the points its play-by-play earned, whether or not those points showed up on the board.
The team pages call this PPA, predicted points added, which is the name the data source uses for its model. The live scoreboard's feed calls its version EPA, expected points added. Same idea, two different models, so do not compare a live EPA number with a postgame PPA number directly. The live feed also assigns no value to touchdowns, field goals, punts or turnovers, so its totals stay small.
Example. Tennessee beat Georgia Tech 38–17 in week 2. Tennessee averaged +0.24 points added per play, Georgia Tech +0.02. Over the whole game that came to +16.5 points earned by Tennessee's plays against +1.1 for Georgia Tech's. Those totals are in points, which is what lets the verdict compare them with the score.
Success rate
A play is a success if it keeps the offense on schedule: at least half the needed yards on first down, at least 70 percent on second down, and all of them on third or fourth down. Success rate is the share of a team's plays that met that bar. It measures consistency, not size: a 4-yard run on first down and a 40-yard run on first down are both one success.
Example. In that same game Georgia Tech's success rate was 45.8 percent and Tennessee's 43.5 percent. Play for play, Georgia Tech stayed on schedule slightly more often. It still lost by three touchdowns, and the next two numbers say why.
Explosiveness
Success rate counts how often a play worked. Explosiveness asks how big those plays were: it is the average value, in points added, of only the successful plays. A team can be steady but small, or erratic but dangerous, and the two numbers together tell you which.
Example. Tennessee's successful plays averaged 1.37 points added, Georgia Tech's 1.02. Georgia Tech moved the chains a little more often; when Tennessee's plays worked, they went a lot further. That gap in explosiveness, plus finishing drives, is the 38–17.
Havoc
Havoc is the share of the opponent's snaps on which the defense wrecked the play: a tackle for loss, a sack, a takeaway, or a pass broken up. It is counted from the play-by-play text, so it is approximate, and it says nothing about the plays in between. It is a good quick read on whether a defense is disruptive or just bending.
Example. Tennessee's defense produced 14 havoc plays on Georgia Tech's 71 snaps, 19.7 percent: one tackle for loss, four sacks, two takeaways and seven pass breakups. Georgia Tech's defense managed 10 on 66 snaps, 15.2 percent.
Standard downs and passing downs
A standard down is one where the offense can still choose to run or pass: first down, second down with 7 yards or fewer to go, and third or fourth down with 4 or fewer. A passing down is everything else, the second-and-longs and third-and-longs where the defense knows what is coming. Every offense does worse on passing downs; the question is how much worse, and how often it lets itself get there.
Example. Tennessee's success rate was 47.8 percent on standard downs and 34.8 percent on passing downs. Georgia Tech's was 52.0 and 31.8. Both fell off; Georgia Tech fell off harder once it was behind the chains.
Scoring opportunities and points per opportunity
A scoring opportunity is a drive that reached the opponent's 40-yard line, or that scored. Once a drive gets that far, the offense should come away with something. Points per opportunity is how much it actually came away with, on average: a touchdown counts 7 with the kick, a field goal 3, a turnover or a missed kick 0. The first number says how often a team got into range; the second says what it did once there.
Example. Tennessee had 9 scoring opportunities against Georgia Tech and scored 4.2 points on each, which multiplies out to its 38. Georgia Tech had 8 opportunities, nearly as many, but only 2.1 points on each, for its 17. The game was not about getting into range. It was about finishing.
The verdict: flattered, undersold, or matched
The verdict at the top of a team page compares two margins. One is the real one from the scoreboard. The other is the margin the plays earned: the team's total points added minus the opponent's, using the totals described above. If the two are within 7 points of each other, the page says the score matched the play. If the scoreboard margin is more than 7 points better for the team than its plays earned, the score flattered it. If it is more than 7 points worse, the score undersold it.
Regression watch, the line at the top of Next up, carries the last game's verdict forward: a team flattered by more than 7 points, or undersold by more than 7, is flagged, with a small tag on its Projections-page card. The numbers quoted come from a study of 5,632 team-games from 2022 to 2025 on this site's own verdict rule: flattered teams fell about 3 points short of their season-to-date form in the next game, undersold teams beat it by about 1.7, both several standard errors from zero. The note is a read on form only.
A verdict is not a prediction and does not say a team got lucky. It says the score was unusual for that play-by-play, and the sentence usually names the reason, most often turnovers, which swing the score far more than they swing the play values. The index shows a verdict only when the gap is 10 points or more, so the ones you see there are the clear cases.
Undersold. Michigan beat Western Michigan 13–12 in week 2. Its plays earned a margin of about +14, with +0.24 points added per play against −0.04, but Michigan turned the ball over twice. The one-point win ran 13 points behind the play.
The live scoreboard shows a running version of the same comparison during a game, from the live play feed. It is an estimate: that feed values fewer plays (none of the scoring plays) and includes garbage time, while the final verdict on the team page uses the full postgame stats with garbage time excluded, so the two can differ.
Flattered. Penn State beat Temple 27–9 the same day. Play for play it was close, +0.21 per play against +0.18, and the plays earned a margin of only about +3. Temple's turnover and Penn State's finishing turned that into an 18-point win, 15 points ahead of the play.
Strength of schedule
The Schedule line at the top of Next up averages our model's ratings of a team's opponents, once over the games it has played and once over the games left, and ranks each average among all FBS teams, so "12th-toughest so far" means eleven teams have faced stronger opponents on average. The ratings come from the newest weekly our model snapshot, in points against an average team; FCS opponents share one pooled rating; conference title games are not counted. It is a read on the opposition faced, not a correction to any other number, and it carries the same experimental label as the our model card.
Series history
Under the next opponent, a collapsed line such as "Series history · 14 meetings since 1968" opens the record between the two schools: the all-time record, the last meeting and its score, the current streak, the record at this week's stadium, the biggest margins each way, and the meetings since 2000. Every fact is computed from the list of past games in the data source, not written by anyone, so when the source's own totals disagree with its list of games the page says so and uses the games.
How the game swung
After a game, a collapsed line under the score such as "Tennessee was as low as 31% and won" opens a chart of win probability across the game: at every play, the chance the team wins from that situation, as estimated by the data source's model from the score, the clock, the down and distance and the field position. A line that never leaves the top of the chart was a comfortable game whatever the final score says; one that crosses the middle several times was a genuine contest.
Below the chart the biggest single swings are listed: the plays that moved the chance the most, with the quarter, the clock and what happened. A game's decisive moment is usually there, and it is often not the play the highlights chose.
Injuries: official and AI-assisted
The Injuries section on a team page comes from one of two places, and the label on the section says which.
Official report means the conference's own availability list, filed by the school and published by the conference before every conference game. The SEC, Big Ten, ACC and Big 12 publish one, usually the evenings before the game and once more on game day. The site reads each one when it is posted, lists every player marked out, doubtful, questionable or probable, and compares it with the previous report for the same game, so a player is marked New when he was not listed before, Was Questionable when his status changed, or Carried over when it did not. Players the school has declared out long term are listed separately. The report gives no reasons; when public reporting had already explained an absence, that explanation is shown under the player with its own links.
AI-assisted means gathered by an AI model from public reporting: beat writers' posts and news articles, each entry with the links that support it. Every link is checked before an entry is shown, and an entry with only a paywalled source is held back until a second source confirms it. When there is no official list for a game, a non-conference game for instance, this is what the section shows. When an official list exists it takes precedence for every player it names; AI-assisted entries remain only for players the list does not mention.
Neither source is used in any number on the page.
Where the numbers come from
Everything on this page is computed from CollegeFootballData, the same source as the rest of the site: its per-game advanced stats for PPA, success rate, explosiveness and the down splits; its play-by-play for havoc; its drive log for scoring opportunities. The "season to date" matchup block on each page uses the season versions of the same numbers, which are not adjusted for who a team has played, so a light schedule inflates them. Injuries and coverage are gathered separately and never touch any of these figures.