Revamping Win Probability for 2025

Imagine this: it’s the fourth quarter, tie game, your team has the ball on the opponent’s 1-yard line with one second left. What are the odds they actually win?

That’s the kind of question a win probability model is built to answer and it’s one I’ve been calculating for years. But until now, the system powering those numbers was showing its age. For 2025, I’ve completely overhauled win probability on CollegeFootballData.com, replacing outdated models with a modernized, better-calibrated engine that understands not just the flow of regulation, but also the unique dynamics of clutch time and overtime.

The Old Way (Retired Models)

The previous version of win probability was powered by two models: one for regulation and one for overtime. These were built years ago using a now-obsolete JavaScript library called SynapticJS. They've generally worked well enough, but they had serious drawbacks:

They were essentially black boxes, with no good way to measure calibration or error.
There was no mechanism for handling rare, high-leverage scenarios, like the one-second, goal-line situation above.
And practically, SynapticJS is no longer maintained, making the models brittle and hard to improve.
Lastly, how many people are training machine learning models on JavaScript? There's a reason JS is used primarily for web while most ML happens in the Python (and R) ecosystem.

In short, they were due for replacement.

The New Models (2025 Revamp)

For the new season, I’ve rebuilt the system from the ground up in Python using XGBoost, a modern machine learning library that’s fast, well-supported, and ideal for structured sports data.

Instead of two opaque models, there are now three specialized models:

Regulation Model – trained on all non-overtime plays, handles the bulk of game situations.
Clutch Time Model – trained specifically on close games in the final minutes, where every play can swing the outcome.
Overtime Model – trained only on overtime possessions, which are fundamentally different because of college football’s unique rules.

The regulation and clutch models are combined into a blended approach: the regulation model drives most of the game, while the clutch model gradually takes over in high-leverage late situations. This way, the system is both broadly calibrated and sharply tuned for the moments that matter most.

Calibration and Results

A major advantage of the new models is that they can be tested and evaluated, which is something the old SynapticJS models couldn’t do.

For each of the three models, I generated calibration curves that compare predicted probabilities to actual outcomes. The closer the line is to the diagonal, the better calibrated the model is. The results show:

Regulation Model

Clutch Model

Overtime Model

The bottom line: the new system doesn’t just look smarter; it actually measures smarter.

Clutch Time in Action

One of the biggest improvements comes from how the new system handles endgame situations.

In the old model, a tie game with one second left and the ball on the opponent’s 1-yard line might have been treated like a coin flip with ~50% win probability. That never felt right.

With the new blended approach, the clutch model takes over, recognizing this as a near-certain win for the offense. Scenarios that used to break the model now produce realistic, intuitive results.

This “clutch awareness” makes the new win probability charts much more believable, especially in the final minutes of close games.

🚀 New Tools on the Site

Along with the revamped models, I’ve added a brand-new Win Probability Calculator to the site. This tool lets you plug in the game situation (score, time remaining, down, distance, and field position) and instantly see the home team’s win probability. Behind the scenes, it uses the new regulation + clutch blended model, so the numbers reflect both the general flow of a game and the pressure of high-leverage moments.

Advanced box scores and data for all 2025 matchups and beyond have been using this new blended model. And every win probability chart you see during the season will now run on the new models. You’ll notice smoother, more realistic shifts, especially late in games where the old system struggled.

Finally, the Excitement Index, a measure of how thrilling a game is based on swings in win probability, has been using the updated engine during the 2025 season. Because the clutch model is sharper, excitement ratings will better capture the drama of close finishes.

Overtime Model

Overtime in college football is a world of its own: possessions starting at the opponent’s 25, alternating turns, and since 2021, two-point shootouts. That structure makes overtime play fundamentally different from regulation, which is why a dedicated model was necessary.

The new overtime model is trained only on overtime possessions and captures those dynamics directly. Its calibration curve shows a solid fit, giving confidence in the numbers when games head into extra frames.

Takeaways & What’s Next

The old SynapticJS models got us this far, but they were opaque and unmeasurable. The new system is:

Transparent – feature sets are clear, and the models can be tested.

Calibrated – probabilities better match reality across all game states.

Clutch-aware – no more 36% win probability in a one-yard, one-second tie.

Specialized – overtime handled with its own dedicated model.

This overhaul powers not just the charts you see on the site, but also the new calculator and updated Excitement Index.

Looking ahead, I hope to extend these improvements into other areas like live win probability updates during games, deeper situational models (e.g., 4th-down decisions), and expanded API access for developers.

Closing

The 2025 season marks a new era for win probability on CollegeFootballData.com. Whether you’re following along live, exploring charts after the fact, or testing “what if” scenarios in the calculator, the numbers you see are powered by smarter, sharper, clutch-ready models.

2025 is here, and win probability just got a whole lot smarter.