Browse this week's slate
Every live game, tracked against the call
How the model works, and whether it does
See where every team stands, replay 2025
Who makes the twelve-team field
What we said, what happened, and where we split from the market
Every projection, week by week, as a CSV
Your call against the projected margin
HOW SATURDAY THINKS
Saturday rates every team in points. That rating starts as a preseason projection and moves with the games actually played, and any matchup is then run thousands of times, which gives back a range of outcomes rather than one score. No game is ever priced using anything that happened after kickoff. That rule is what makes the record on the second tab worth reading. Start with how it works.
EVERY GAME IS ONE EQUATION
Home strength minus away strength, plus home field. That is the projection. Every team carries exactly one number, and the distance between two of them is the margin Saturday expects. The rating itself starts in the spring, built off last year's team, the recruiting class, the portal haul, who came back and who is coaching. Then the season takes over and each week of results moves it. The one game a rating never sees is the one it is being asked to project.
Saturday rated Ohio State about 11.1 points stronger than Penn State on a neutral field. Add 3.3 for playing in Columbus and you get Ohio State by 14.4. Both of those ratings were fit on games those teams had already played, and nothing else.
A real, worked example from a completed game (week 10, 2025), fixed for illustration, not this week's projection.
There is a real formula underneath all of this, with real coefficients, and it is published. See what goes into a rating.
THE OUTCOME CURVE
“Ohio State by 14.4” is the middle of a range, not a call on the final score. Play that game a thousand times and it lands everywhere from a rout to a Penn State win, because that is what this sport does on a given Saturday. So the answer is a curve. The win probability is the share of it sitting on the winning side of zero.
The curve is about 16 points wide, and it has to be. Good teams lose games they have no business losing, often enough that a model shrugging it off would be overstating what it knows. The percentage on screen carries that risk instead of rounding it away.
THE CURVE IS WIDEST IN AUGUST
On opening weekend the model has watched none of this season. It is working off last year, a recruiting class and a portal cycle, so it widens the curve to say so. Results narrow it. By midseason the ratings are built on real games and the range settles near its floor, where it stays. Here, the uncertainty is a number you can see on the chart.
Chart's y-axis is σThe Greek letter sigma (σ), used here for how wide the range of likely outcomes is, in points. A bigger number means more uncertainty; it shrinks as the season goes on and Saturday sees more games., in points.
That is what the “early season · high uncertainty” flag on a game page is telling you. Fewer games behind a rating, wider range in front of it. As the season builds, Saturday also blends in per-play efficiency from the games already played, a signal that reads a team better than final scores alone. It carries no weight on opening weekend, when there are no plays from this season to read, and grows to about a third of the projection by midseason.
TESTED AGAINST THE MARKET
Saturday prices games off its own ratings and never looks at a betting line to do it, which is what makes the market a fair scorekeeper. There is one exception, disclosed on purpose: preseason market win totals help set each team's opening rating, an influence that washes out as real games get played. The number it is graded against is where the sportsbooks opened a game, not where they closed it. Across four seasons (2022 through 2025), Saturday finishes within 0.1 of a point of that opening number from week 5 on, and on plenty of games it had the better number. Everything below was priced before kickoff.
The books opened USC a home favorite. Saturday leaned the other way, and Washington won by ten.
Florida opened a home favorite. Saturday called it a coin flip.
The line opened Georgia by two and a half touchdowns at Auburn. Saturday said one, and the final was a one-score game.
The line opened TCU by a field goal in the rivalry game. Saturday had two touchdowns, and it finished wider than that.
Those are the wins, and the market still beats Saturday more often than not in September. The whole record, calibration and all, including where it trails, is on the .
WHAT MOVES A PROJECTION
Nobody eyeballed these sliders. Each toggle on a game page moves the margin by what that factor is actually worth, measured from real games. A few of them barely move it at all, because that is all the data supports.
Fit stadium by stadium out of the margins themselves, from 2.3 at the quiet venues to 4.3 at the loudest
Per net giveaway. The biggest single swing any toggle can apply
Scaled to what the starter is worth per play. An elite quarterback costs far more than a replacement
What a shanked kick costs, on average
Per net sack, backtested
Per degree away from 60°F, and only on the projected points. Real, but small
WHAT SATURDAY IS NOT
No game is priced off what a sportsbook has posted. Every projection on this site comes out of Saturday's own ratings. The single exception is preseason market win totals, which help set the opening ratings in August and fade to nothing once real results arrive. (In a season where the recruiting-talent feed hasn't published yet, those same win totals also set how spread out the opening ratings are, under the same fade.)
These are probabilities. A 75% favorite is supposed to lose one Saturday in four, and across a season it will. When that happens, the model is doing its job.
This is here to help you understand a game. Beating a sportsbook was never the point, and nothing on this site promises it.
THE FULL TECHNICAL SPEC
Every setting and every constant behind Saturday is written out: the preseason projections, the weekly rating fit, how the curve gets built and taken apart, the simulations, the scenario adjustments, and the four-season validation. Enough detail to rebuild the model from scratch.