About buffalo win probability

This site shows the Bills' chance of winning, and updates it after every play. It runs live on game day, and you can replay any Bills game since 2009, one play at a time.

Under the win bar there are two little boxes. The first is about the next play: what teams usually do in that spot, and how much each option would help or hurt Buffalo. Once the game's over there's no next play, so that box sums up the game instead. The second box is the play that just happened, and how much it moved things.

There are four more pages up top. The season page shows a whole season at once, one little chart per game. Versus does the same for every game against one team. Records has the biggest plays and players since 2009, plus the best comebacks. And standings has the standings for any season since 2002.

The win chance comes from a formula built on past games. It looks at five things: the score, the time left, who was favored going in, where the ball is, and who has it. Basically it asks: when games looked like this one, how often did this team win?

The last 5 minutes of a close game get special treatment. When it's a one-score game, the clock is everything, so a second formula takes over. It starts from the first one's number and adds what the clock allows: how far the team with the ball has to go in the time it has, how many timeouts each side has left, and whether a field goal is enough. Two spots are simpler. If the team ahead can just kneel it out, they've won. And a field goal try on 4th down in the last 40 seconds comes down to how often that kick is good, and what the other team can do with whatever time is left.

Who was favored comes from the point spread, how many points Vegas expected one team to win by, right before kickoff. If there's no spread, the site uses its own team ratings, built only from games played before that one.

The formula learned from every regulation play in every NFL season from 2001 to 2025, except for a batch of games it never got to see. Those were saved for testing. Anything looks smart on games it's already seen, so the real test is games it hasn't.

The easy check: stop each test game at kickoff, at halftime and with five minutes left, see who the site had ahead, and see who won. The last row is what you'd get by just picking the home team every time.

It gets better as the game goes on, which makes sense, since there's more to go on. Even at kickoff it beats picking the home team.

The percentages also mean what they say. When it says 70%, that team wins about 7 times out of 10. Same at 60%, 80%, 90% and so on. To check, in each test game we found the first time the site had either team at 60% or better, then 70% or better, and so on, and counted how often that team won. The last play of each game doesn't count, since by then everyone knows who won.

A team is rarely at exactly 90% the first time it gets there, so the middle column is what the site was actually saying at those moments, on average, and the last column is how often those teams won. They're never more than 0.4 points apart. The biggest gap is at 80%, where the site was a little high. Ties (8) and 2 games with no plays on record are left out.

Just the Bills, every game since 2009 (including games the formula learned from, so this is history, not a test): Buffalo got to 90% in 168 games and won 158 of them. They fell to 10% in 145 games and still won 12.

Overtime is its own thing. Whether a score ends the game depends on whose possession it is and what the other team already did. The formula can't see that, so overtime runs off the NFL's actual overtime rules instead, and those have changed a lot.

The site tracks each overtime possession. It knows how often a drive ends in a touchdown, a field goal or nothing, and figures the win chance from there. A score that ends the game is 100%. A regular-season tie counts as half a win.

The first box under the win bar has three lines: the play that's about to happen, what NFL teams usually do there, and what each option is worth to Buffalo's chances.

BUF 4th and 1 at HOU 34
teams here: go 76% / FG 22%
BUF win%: > go 60.5% (+0.6) | FG 60.3% (+0.4)

After the final whistle there's no next play, so the box tells the story of the game: where it started, how low it got and when, and the one play that swung it most. Tap it to read a long play in full.

BUF opened at 58.0%
low: 4.0% at Q4 6:12, BUF 17, HOU 31
biggest play: +28.0 Q4 1:40, J.Allen pass

On first, second and third down, the second line says whether teams usually run or pass there. It only says something when one is clearly more likely. If teams pass anywhere from 35% to 65% of the time, it's a toss-up and the line stays blank. It shows up on 42% of plays, and when it does, it's right 80.1% of the time.

On fourth down, it shows how often teams go for it, kick or punt. On the test games it called go-for-it or not correctly 89% of the time.

Anything teams do less than 1 time in 20 from that spot isn't shown. Nobody goes for it on 4th and 9 from their own 5, so the box doesn't pretend they might. That's why the numbers don't always add up to 100.

On first, second and third down, the third line shows how the drive could end: touchdown, field goal, punt or turnover, and how much each would move Buffalo's chances.

And when the team ahead can just kneel until time runs out, because the other team doesn't have enough timeouts left to stop the clock, the box says "kneel it out" instead.

To figure out what an option is worth, the site plays out each way it could go (the new score, who has the ball and where, how much time is left) and works out the win chance for each. To do that it needs a few football facts.

Those facts come from 2023 on, and so do the habits on the second line (2023 on). Older seasons made both worse, because the game has changed. Teams go for it on fourth down way more than they used to: going by every season back to 2001 you'd expect 18%, and now it's 22%. Kickers are better too: going by every season, a 50-yarder would look like a 68% kick.

Every play moves the win chance up or down. The site calls that the play's swing, and credits it to the players who made the play. It counts for the player's own team, so a Houston touchdown is a plus for the Houston guy who scored it.

A score and the kickoff after it count as one swing. Kickoffs and punts by themselves go to nobody, and neither does a made extra point. The win chance often dips a bit on a made extra point, because the other team is about to get the ball, and blaming the kicker for that once left one Bills kicker 169 points in the hole on 118 good kicks. A missed extra point is still on the kicker, though.

Defense is the weak spot. A defender only gets credit for a sack, a pass breakup, an interception or a fumble. Tackles get nothing, because the guy who makes the tackle on a long gain is usually not the guy who gave it up. Good coverage and pressure never show up in the play-by-play, so they get no credit here.

The standings page covers any season since 2002, sorted the way the NFL sorts it. Nothing on it is a guess, it's all from games already played: the record, home and away, division and conference, points for and against, and the current streak.

Ties are broken the way the league breaks them, in the league's order: head to head, division or conference record, common games, strength of victory, strength of schedule. After that the league has a dozen more tiebreakers based on points; here, point differential stands in for all of them. If it's still tied, the site goes by the order the teams are listed, where the league would flip a coin. Run on the real 2022, 2023 and 2024 seasons, it gets all fourteen playoff teams and their seeds right, and on 2018 and 2019, back when six teams per conference got in, it gets all twelve. Right every time it can be checked.

A finished season shows its playoff teams: each seed, and whether they won the division, got a wild card or had a first-round bye. Through 2019, six teams per conference got in and two got byes; since 2020 it's seven, and only the top seed gets a bye. Each season is sorted by its own year's rules.

It only goes back to 2002, because that's when the league last reshuffled its divisions. Before that there were 31 teams in six divisions, which doesn't fit the eight-division table.

The site runs on public NFL play-by-play. Each game is saved once it's final. The plays go back to 2001, but the site only offers games from 2009 on, because before that the play descriptions are just labels like "Pass Completion", with nothing to read and no player to credit.

The data has some problems, so four things get cleaned up on the way in.

Wrong clocks. About one game in twelve before 2025 has a play with the wrong time on it. If the play's description has the time, the site uses that. Otherwise it takes the time from the plays around it.

Field position. It's read from the text, like "PIT 24", because the yard-line number in the data sometimes counts from the wrong end of the field.

Postponed games. A game that got postponed and played later only shows up once, not twice.

Wrong scores. Sometimes a play gets listed with a score from some other point in the game. The site ignores it and keeps the real score going, more on that under "where it misses" below.

It doesn't watch the game. All it knows is the play-by-play: the score, the clock, the down and distance, and where the ball is. It doesn't know who's hurt or what the weather's like. The pregame point spread helps a little, since Vegas already took things like a missing starter or a snowstorm into account. But if the quarterback gets hurt during the game, the number won't change until the score does.

Timeouts only matter late in a close game, so that's the only place they're used: in the last 5 minutes of a one-score game, and to tell when a team can kneel out the clock. Earlier than that, adding them didn't make the forecast any better.

Whether the starting quarterback got pulled was tried too, and left out because it made things worse. Teams mostly pull their quarterback when they're already getting blown out, so the formula was really learning about blowouts, not quarterbacks.

The habits shown for the next play are league-wide averages, not Buffalo's or head coach Joe Brady's.

The live number runs about 15 to 30 seconds behind the game. Replays have no delay.

Nobody gets these right every time. Here's where this site is known to be off right now.

Needing a long field goal at the end. When a team leads by one to three with under 20 seconds left, and the other team has the ball 40 to 60 yards from the end zone, the site gives the leader 84.8%. Those leaders only won 75.2% of the time (129 plays). The team that's behind just needs a few yards and a long field goal, and kickers have been making long ones more often lately than in the older games the formula learned from.

Going for it when a field goal would do. On 4th and 1 or 2 with under 30 seconds left, when a field goal ties it or wins it, teams almost always kick: 27 times out of 28 since 2001. The next-play box says teams go for it there 15% of the time. That's a bit high.

Too confident, or not confident enough? Neither, really. Back in the table that checks 60%, 70% and so on: when the site first said 80% or more, it averaged 82.9% and those teams won 82.5%. That's the biggest miss in the whole table, and it's small. From 90% up it runs a hair low instead.

Bad scores in older games. The play-by-play data sometimes lists a play with the wrong score on it. The site ignores a score that doesn't add up, but a few still sneak through. You can spot one as a jump of more than 20 points that snaps right back. In games from 2009 to 2013, about 16% still have one (it was about 40% before the fix). Some are bad scores and some are real swings, and the chart can't tell which. In 2025 it's about 5%, and every one of those is real, usually a long pass in the last minute that the next play answered. The records page skips all of those plays, real or not.

This site is independent. It is not affiliated with, endorsed by, or connected to the NFL, the Buffalo Bills, DraftKings, or any other team, league, sportsbook, media company or business. Team names and colors are used only to identify the teams.

It is not a betting site and is not meant to be used for betting. Nothing on it is betting advice. It takes no bets and links to no sportsbook.

The point spread and the total shown are a sportsbook's last numbers before kickoff, as publicly reported, when a sportsbook posted them for that game. The sportsbook is named so you know where the number came from, not as a recommendation. The site never says a team covered the spread unless that spread came from a sportsbook.

The play-by-play is public NFL game data.