About v12

How TEMPO works

TEMPO (Team Efficiency, Margin, Probability, and Outlook) is a statistical model of college football. It rates every FBS team, forecasts every game and plays out the rest of the season, all from one data source, CollegeFootballData. Each section below starts in plain words; the equations wait under “The math”.

What it is

A rating of every FBS team’s strength, in points.
A forecast for every game involving an FBS team: scores, margin, total and win chance, with 90% ranges.
Conference and playoff chances from 10,000 simulated seasons.
A public record of forecasts saved before kickoff, each graded once.

What it isn’t

A poll or a résumé ranking: it rates how good a team is, not what it has earned.
Aware of injuries, suspensions, weather or other game-week news.
Built on betting lines, other systems’ predictions or provider analytics such as SP+ and FPI. None of them goes in.
Betting advice. The betting market is usually more accurate.

What a rating means

A team’s rating is the margin it would be expected to win by against an average FBS team on a neutral field. A rating of +14 means about two touchdowns better than that average team; a negative rating means worse. It measures strength, not a season’s résumé.

Offense is the points a team would score against that same average team, and defense the points it would allow, so a lower defense is better. Offense minus defense is the rating. Both are per game, so a fast-paced team scores and allows more than its efficiency alone suggests.

Ranges

Each rating comes with a 90% range for the team’s true strength. A game’s range is wider, because one game adds its own luck. A well-judged 90% range still misses about one result in ten. Ranges don’t only narrow: uncertainty grows between games, because teams change during a season.

The mathOffense, defense and the rating, as equations.

Rating, offense and defense

\begin{aligned} O_i&=\frac{\lambda+\tau_i+\theta_i}{2}\\ D_i&=\frac{\lambda+\tau_i-\theta_i}{2}\\ O_i-D_i&=\theta_i\end{aligned}

θ is a team’s strength, its rating; τ its scoring tendency, how many more total points than average its games produce; λ the league level, the expected total of two average FBS teams on a neutral field. Strengths and tendencies are measured from the FBS average, so an average FBS team has θ = τ = 0.

How a game is forecast

The forecast margin starts from the difference between the two ratings. The home team adds a home-field edge of 2.8 points, learned from 2014–2025 games; a neutral site adds none.

The total comes from each team’s scoring tendency and the league’s scoring level. Margin and total together give each team’s expected score. A team can’t score below zero, so the expected scores allow for that floor, which matters most in mismatches: it shrinks the expected margin and lifts the total.

The win chance is the chance the margin ends above zero, given how uncertain both ratings are and how much a single game varies. Scores, margin and total each come with a 90% range.

The mathThe margin, total, win chance and scores, as equations.

Expected margin and total

\begin{aligned}\mu_M&=\theta_h-\theta_a+\eta_M v\\ \mu_T&=\lambda+\tau_h+\tau_a+\eta_T v\end{aligned}

h and a are the home and away teams. v is 1 at the home team’s site and 0 at a neutral site, and η the learned home effect.

Win chance and scores

\begin{aligned}P(h\text{ wins})&=\Phi\!\left(\frac{\mu_M}{\sqrt{V_M}}\right)\\ \mu_h&=\frac{\mu_T+\mu_M}{2},\quad\mu_a=\frac{\mu_T-\mu_M}{2}\\ \sigma_P&=\frac{\sqrt{V_M+V_T}}{2}\\ E[\max(0,P)]&=\mu\Phi(\mu/\sigma_P)\\ &\quad+\sigma_P\varphi(\mu/\sigma_P)\end{aligned}

Φ and φ are the standard normal distribution and density. The variances V add the uncertainty in both teams, their drift until kickoff and one game’s noise. Each predicted score is the expectation of that team’s points floored at zero; the forecast margin and total are their difference and sum. Ranges are the adjusted centers plus or minus 1.645 latent standard deviations, with lower score and total bounds clipped at zero: nominal Gaussian ranges, not exact quantiles of the floored scores.

The season outlook

The outlook plays out the rest of the season 10,000 times: every remaining game, each conference race with its tiebreakers, the committee’s ranking and the playoff, under the season’s rules. A team’s playoff chance is the share of those seasons in which it makes the field. Each simulated season draws every team’s strength at once, so a team better than its rating wins more of its games, not just one.

Displayed odds are rounded simulation frequencies. An undecided event seen in none or all of the runs reads 0% or 100%; that is not proof it is impossible or certain. The actual selected field and completed rounds fix outcomes already known.

The committee’s ranking is estimated from its past rankings. Each update runs the outlook again on that update’s ratings and results.

What a game changes

A playoff stake compares a team’s chance in the simulated seasons where it wins its next game with those where it loses. It isn’t a prediction that the team will win. Outcomes seen in fewer than 1% of simulations are left out.

Rules and notices

Season rules are confirmed each summer; “assumed from” means a season runs on an earlier year’s confirmed rules. A schedule that fails a check makes the odds provisional, with a note saying what doesn’t match. Where the data doesn’t say where a game is played, a chance that depends on it shows as unavailable, which isn’t the same as zero.

Past seasons’ fields

For a finished season, the committee’s actual field sits beside three others under the same rules: the committee model’s, fitted only on earlier seasons; TEMPO’s strongest teams, by rating; and its most deserving, by strength of record. In the three alternatives, a ring marks a team the committee seeded differently and a dot a team outside the committee’s field. They are comparisons after the fact, not what anyone published before selection.

The mathThe committee model and strength of record.

The committee’s No. 1

P(\text{No. 1}=i)=\frac{\exp(\mathbf{w}^\top\mathbf{f}_i)}{\sum_j\exp(\mathbf{w}^\top\mathbf{f}_j)}

The committee is modeled as a ranking chosen one place at a time: No. 1 from every FBS team, then No. 2 from the rest, through No. 25. f holds a team’s wins, losses, rating, opponents’ average rating, wins over the model’s top 25, conference title (in the final ranking) and power-conference membership; w their weights, fitted on earlier seasons’ rankings. A simulated season ranks teams by w·f, without the committee’s own randomness.

Strength of record

\operatorname{SOR}_i=P\big(W_i^{\mathrm{ref}}\ge W_i\big)

W is a team’s wins by selection and W^ref the wins of a reference team, as good as the average of the 25 best by rating, against the same opponents at the same venues. The chance is computed exactly; a lower value is a stronger record.

How the ratings learn

Before a season, a returning FBS team starts from what was known before week 1: last season’s rating, older program history, talent and recruiting, how much production returns or arrives by transfer, a new head coach, and whether it plays in a power conference. In the 2026 start, a continuing FBS team keeps about 54% of last season’s rating.

Each game then moves both teams’ ratings and, through their opponents, other teams’ too. A game with full play data counts the score together with two measures of play, success rate and yards per play, weighted as earlier seasons taught. A game with only a final score counts through the score. As games arrive, the preseason start matters less and less.

Teams change during a season, so every rating drifts a little each week: recent games count somewhat more than early ones, and a rating’s range widens until the team plays again. Everything the model learns for a season comes from earlier seasons only.

The mathThe preseason start and the weekly update.

Preseason start

\begin{aligned}\theta_i^{(0)}&\sim\mathcal{N}\big(\mathbf{x}_i^\top\boldsymbol{\beta},\;s_i^2\big)\\ s_i^2&=\exp(\gamma_0+\gamma_1c_i+\gamma_2r_i)\\ &\quad+\beta_{\mathrm{last}}^2\operatorname{Var}\big(\hat\theta_i^{\mathrm{last}}\big)\end{aligned}

x holds a team’s preseason features and β their learned weights. c marks a new head coach and r lower returning production; the fit decides whether either widens the start. An FCS team starts from its own last rating when it has one.

Drift and a game’s update

\begin{aligned}P&\leftarrow P+q\,\Delta t\,D\\ S&=\mathbf{h}^\top P\mathbf{h}+\nu\sigma^2\\ \mathbf{k}&=P\mathbf{h}/S\\ \mathbf{m}&\leftarrow\mathbf{m}+\mathbf{k}\,(y-\mathbf{h}^\top\mathbf{m}-\eta v)\\ P&\leftarrow P-\mathbf{k}\mathbf{h}^\top P\end{aligned}

The filter keeps a joint estimate m of every team, with covariance P. Between games P grows by the weekly drift q over the weeks elapsed. A game updates every team at once: on the margin, h picks its two teams (+1 home, −1 away); y is the observation and νσ² its noise. With play data, y combines the margin, net success rate and net yards per play with weights learned from earlier seasons.

β_last, share of last season kept: 0.54
q, weekly drift, points² per team: 2.00
η, home effect on the margin, points: 2.82

Learned for 2026 from earlier seasons, as the season’s model card reports them.

What goes in, and what never does

Everything the model reads comes from one source, CollegeFootballData. Responses are kept as they arrived; a value that can’t be read stays blank rather than guessed, and no game is corrected by hand.

When

Ratings, results and playoff odds update each Sunday at 15:00 UTC (11 AM ET, 10 AM during standard time), from every game final for at least four hours, Saturday’s included. Forecasts are made each Tuesday at 12:00 UTC, after Monday’s games and before Thursday’s, and published at 14:00 UTC (10 AM ET, 9 AM during standard time). A forecast reads a game only once it kicked off more than 24 hours before that run, so it never reads a game that might still be under way or uncorrected; Tuesday’s ratings can move a little from Sunday’s for that reason and for late corrections. Preseason facts count as known before week 1; everything else must be known by the run. Updates go live at these times exactly. Times on this site are Eastern (ET).

Goes in

Schedules, scores, drives and plays of every game involving an FBS team.
Final scores of FCS-vs-FCS games, only as evidence about FCS opponents.
Preseason facts: rosters, coaches, team talent, recruiting classes and last season’s player statistics.
Weekly polls, only to learn how the committee ranks, for the outlook.

Never goes in

Betting lines. They are kept only to compare with.
Other systems’ predictions, also kept only to compare with.
Provider analytics: PPA and EPA, win probability, Elo, SP+, FPI and FEI. TEMPO fits its own expected points from raw plays.

How accurate it is

The record

The live record is the real test. Each game is graded once, on the forecast that stood at kickoff, beside the betting line saved with it. The public record counts games between two FBS teams, each model version keeps its own, and it begins with the forecasts of Tuesday, October 13, 2026. See it on the Record.

Games without a recorded forecast are replayed: forecast afterward by the current version from the run that would have made them. Replays help testing but never enter the record. Historical pages without a live forecast show scores on the public site, not a prediction made afterward. The local Dev server can show replays, explicitly labeled.

How forecasts are graded

Straight up asks whether the predicted winner won. Against the spread and over/under compare the unrounded forecast with the line. For the spread, TEMPO takes the team whose projected margin beats the market’s. For totals, it takes over when its projected combined points exceed the market total, under when lower. These are picks, not cover probabilities. A forecast exactly on the line makes no pick, a result on the line is a push, and a game without a line is unpriced. Misses are measured in points: MAE is the average miss and RMSE weighs big misses more. Win chances are scored by Brier score and log loss; lower is better for all four. Coverage is the share of results inside their 90% ranges, which should be near 90%, not as high as possible.

Backtests

Before going live, TEMPO forecast every game of 2016–2025 as the live run would have, each from the cutoff that run would have used; 2014 serves only for calibration. In those games:

Total-points ranges held about 91% of results; forecast totals ran slightly too spread out.
Margin ranges held about 88% of postseason results.
Profile projection ranges held 89.6% to 91.0% of actual values.
Against other systems, TEMPO was statistically level with several leading public systems and behind every betting line.

These are development results: the seasons were in view while the model was designed, and their preseason facts were retrieved afterward. The live record is the forward test.

The mathError measures and win-chance scores.

Margin errors

\begin{aligned}e_i&=\widehat M_i-M_i\\ \operatorname{MAE}&=\frac1n\sum_{i=1}^n|e_i|\\ \operatorname{RMSE}&=\sqrt{\frac1n\sum_{i=1}^n e_i^2}\end{aligned}

M̂ is the forecast margin and M the result. Both measures are in points; neither is a betting return.

Win-chance scores

\begin{aligned}\operatorname{Brier}&=\frac1n\sum_{i=1}^n(p_i-o_i)^2\\ \operatorname{Log\,loss}&=-\frac1n\sum_{i=1}^n\big[o_i\ln p_i\\ &\qquad+(1-o_i)\ln(1-p_i)\big]\end{aligned}

p is the home team’s win chance and o is 1 if it won, 0 if it lost, on games with a winner.

Team profiles

A profile describes how a team plays on ten measures football analysts use, from points per drive to havoc. Each is adjusted for opponent and venue: what the team’s offense would produce, or its defense allow, against an average FBS team on a neutral field.

Percentiles place each value among this season’s FBS teams, so a high defense percentile means holding opponents down. Charts draw a measure by its distance from the FBS average, in standard deviations of FBS teams, clipped at three either way. Pace has no better direction and stays off the radar.

A matchup projection sets one offense against the other team’s defense at the game’s venue, with a 90% range for that game’s own value. Profiles and projections describe teams; they never feed the score forecasts.

Plays count when they are readable regulation scrimmage plays outside garbage time (a margin over 43, 37, 27 or 21 points in the first to fourth quarter). Pace counts every regulation scrimmage play.

The mathAdjusted measures and their percentiles.

One offense against one defense

\widehat y=\alpha+o_{\mathrm{offense}}-d_{\mathrm{defense}}+\eta w

α is the measure’s league level, o the offense’s effect and d the opposing defense’s; w is +1 at home, −1 away and 0 on a neutral field. A positive d holds opponents below average.

Percentiles

100\,\Phi\!\left(\frac{x-\alpha}{s_O}\right),\qquad 100\,\Phi\!\left(\frac{\alpha-z}{s_D}\right)

x is an offense value and z what a defense allows; s_O and s_D are the spreads of this season’s FBS offenses and defenses. Havoc flips both directions, and pace reads upward on both sides.

The ten measures

Offensive points per drive

points per possession

Offensive points per regulation possession: 7 per offensive touchdown and 3 per field goal. Return touchdowns and safeties are excluded; a possession ending the half counts as zero. Possessions starting in garbage time are excluded.

Points per scoring opportunity

points per scoring opportunity

Offensive points per possession that reaches the opponent’s 40-yard line or ends in a touchdown. Uses the same 7-point touchdown and 3-point field goal convention as points per drive. A team-game needs at least two opportunities. A value of 4.22 means 4.22 points per opportunity, not a percentage.

Expected points added (EPA) per play

expected points added per play

The mean change in expected points across counted plays with a readable value, using TEMPO’s own expected-points model. Positive values improve the offense’s scoring position; negative values worsen it.

Success rate

percent of counted plays

A play succeeds when it gains at least 50% of the yards needed on first down, 70% on second, or all the needed yards on third or fourth down, without a turnover.

Rushing success rate

percent of counted rushes

The same down-and-distance success test applied only to rushing plays: 50% of needed yards on first down, 70% on second, 100% on later downs, without a turnover.

Passing success rate

percent of counted pass plays

The same down-and-distance success test applied to passing plays, including sacks and interceptions. Turnovers never count as successes.

Yards per play

yards per counted play

Mean yards gained on counted plays, with each gain capped at −10 and 20 yards. This limits one extreme play’s influence; it is not the uncapped box-score average.

Explosive play rate

percent of counted plays

Share of counted plays that are rushes gaining at least 12 yards or passes gaining at least 16 yards. A turnover cannot be an explosive play.

Havoc rate

percent of counted plays

Share ending in a sack, safety, tackle for loss on a rush or completed pass, interception, or any fumble. Pass breakups are not included. Lower is better for an offense; higher is better for a defense.

Plays per game

regulation plays per game

All regulation scrimmage plays, including garbage time. This is a play count, not seconds per play. Faster or slower is not inherently better.

Limits

Game-week news isn’t in the model: injuries, named starters, suspensions, weather. The data has no history of it to learn from. This is where the betting market is furthest ahead, most of all before teams have played.
FCS teams are rated only as opponents, without full profiles.
2026’s preseason facts were first retrieved on September 30, five weeks into the season, so the ratings shown for 2026’s first weeks were computed afterward. Live forecasts use only what was known at their run.
Data can be corrected after a game. A replay may see corrections a live forecast did not.
The outlook approximates the committee and some tiebreakers; it can’t reproduce every discretionary decision.
Ranges assume the model’s learned settings and can’t include every source of uncertainty.
Fit checks reject failed fits but can’t prove the best fit was found. Different machines can produce slightly different numbers; the published ones are authoritative.
Earlier seasons’ pages show their saved final ratings, not week-by-week snapshots of what was published at the time.

Data and credits

CollegeFootballData and its API: schedules, scores, drives, plays, preseason facts, and the betting lines kept as benchmarks.
The Prediction Tracker: other systems’ predictions, as benchmarks only.
Bill Connelly’s analytics glossary: garbage time, success rate, explosive plays, havoc and finishing drives. TEMPO’s definitions follow its own implementation.
Established statistics: the Kalman filter, Gaussian marginal likelihood, Plackett–Luce rankings, natural cubic splines and the week-resampling bootstrap. TEMPO’s own: its expected points, preseason start, committee features and strength of record.
KaTeX sets the math; Bricolage Grotesque and JetBrains Mono (SIL Open Font License 1.1) set the text; Material Symbols (Apache License 2.0) draws the icons.

Privacy and notices

TEMPO has no accounts and sets no cookies. To count visits and measure page speed it uses Vercel Web Analytics and Speed Insights. For each page view they record the page address (without anything after a ? or #), the site you came from, an approximate location (country, region and city), your browser, operating system and device type, and how fast the page loaded. Visitors are counted with a code made from the request that changes every day, not anything stored on your device. Your theme and chart edits stay in your browser.

TEMPO is an independent project with no ties to the NCAA, the College Football Playoff, any conference or school, or any sportsbook. Team and conference names and logos are trademarks of their owners, used only to identify the teams and conferences. The data comes from CollegeFootballData and is not official NCAA, conference or school data.

Forecasts are for information and entertainment, not betting advice. No result is guaranteed, and the betting market is usually more accurate than any model, this one included.