Total points is one way to gauge a comeback, but it may not be the best way. A comeback from a 10-point deficit in the 1st quarter is not the same as a 10-point comeback with under 5 minutes to play in the 4th quarter. An alternative way to measure the

*comebackiness*of comebacks is by using the win probability graphs.

One of the two cryptic acronyms on the bottom right of each of the WP is CBF, which stands for Comeback Factor. CBF represents the chance that the winning team would be victorious at their low point in the game. For example, the Giants beat the Cowboys last night despite having a WP of only 0.02 at their nadir. That translates into a CBF of 50, because the Giants had about a 1 in 50 shot at winning. WP is not measured below 0.01, so the maximum CBF is 100.

This past weekend's average CBF for all games is 14.7, meaning the winning teams averaged a low-point WP of 0.14. That's good enough for the 8th biggest comeback regular season week since 2000. The biggest comeback week in the data was week 12 in 2003. In that week, there were seven highly improbable comebacks, including two games with a 100 CBF.

NYJ won 13-10 over JAX, despite a 0.11 WP--trailing by 4 points with 3:33 to play and facing a 4th down at their own 40.

IND beat BUF 14-10 on a TD from 4 and goal at the 1 with 1:42 to play.

Billy Volek led TEN to a 38-31 win over ATL despite trailing 21-0 in the first half.

MIA beat WAS 34-23 despite muffing a punt on their own 15 while trailing 20-10 late in the 3rd quarter.

The STL-ARI game was crazy.

*Both*teams came back from WPs of 0.03. STL ultimately won thanks to a 23-yd pass from Bulger to Holt on 4th and 7 that led to the tying TD with only seconds remaining. STL went on to win with an OT FG.

The NE-HOU game was an even bigger roller coaster. NE tied it up on a 4th and goal Tom Brady TD pass in the final seconds to force OT. Then in the first play in OT, Tony Banks threw an interception inside HOU's own 30. HOU managed to block the NE FG attempt, then drove to the NE 35 with a 1st and 10. HOU was driven back out of FG range, punted, and allowed NE to drive for the winning FG with only 45 sec remaining in OT.

BAL beat SEA 44-41 despite trailing 41-24 while on defense as late as 7:19 to play in the 4th quarter. Blocked punt for a TD, a fumble recovery on their own 30, a heroic effort by QB Anthony Wright, and a clutch FG by Matt Stover brought BAL back from the brink to win in OT. This wasn't one of those games where a team's WP spiked to a low-point before a clutch conversion or score. This was a type of game I call a "flatline," where one team has such an insurmountable lead the WP graph is pegged at 0.99 WP for an extended period.

Although it's not something you ever thought you'd think about again, 2003's week 12 stands out as the biggest comeback week in recent memory.

For an index of the biggest comebacks by year or by team, you can use the Top Games tool.

I think the WP model is making some assumptions about late game 4th down decisions. A lot of the really big comebacks include a 4th down decision that seem to overestimate the degree of comebackiness involved.

Take, for example, the 2003 NE-HOU game with NE down by 7 inside of 2 minutes remaining.

NE ball @HOU 9, 3rd and 6... (17%WP)

and after a 5 yard gain...

NE ball @HOU 4, 4th and 1... (1%WP)

I suspect that the model's assumption is that NE gets a field goal there. I'm not sure whether that 5 yard gain is negative WP, but even if it is the result should not be a dramatic 17% to 1% drop. This is just one example, but I'm seeing a similar effect a lot of the other comebacks

With yesterday's Broncos comeback, it looks like there's an issue with the WP model not knowing that the game was being played at Denver altitude. According to the model Denver was down to a 1% WP when they had 4th down at the Bears 41, and on the next play Prater tied the game with a 59-yard field goal.

I don't think the model assumes NE is going for a field goal, but getting a touchdown does not get them the win, only a tie. So whatever the probability that they don't make the first down is all on the side of HOU winning. If they do score, HOU still has time to get a field goal or touchdown before the game ends, however unlikely. If they don't, it goes to 50% at the start of OT.

Also, NE can get a first down but not a touchdown, there is a three yard difference, so even if they convert, they could still come up short. So the significant drop in probability is due to the time running off the clock and loss of down, not the yardage. Any yardage that did not result in a first down or touchdown would drop the probability that much.

Fun stuff; I enjoyed this read.

However, I must point out the intense mathematical weirdness of averaging CBFs, when CBF is the inverse of a probability. There's just not a reasonable argument that I see that that's the most logical way to measure the overall comebacky-ness of the week.

It would make more sense to just average the lowest WP the winning team had in every game, and figure out what week has the lowest average. This gives you a nice, simple measure that ranges from 50% (i.e. the team with the first positive play won every game) to 1% (every winner was dead to rights at some point).

If you (reasonably) want to argue that our perception of how comeback-heavy a week is is fueled by the extreme cases, as opposed to the average - that is, that one 50% game and one 2% game is more remarkable than two 26% games. But even then, that doesn't argue that 1/(lowest WP of winner) is the best measure of CBF.

My first instinct for a meaningful formulation of CBF would be to to measure the statistical distribution of winner's lowest WP. Obviously the minimum value is 50%, and there would be a ton of games at or near 50%, with a long tail heading out towards 1%. I would measure CBF as the number of standard deviations away from 50% the game was. a 50% game would get a zero, and you could scale it such that a 1% game gets a 100.

@Eddie

It may be that the 3rd down scenario is preferable, but only slightly. Even if they've lost half their chance of winning on the 5 yard run (an overestimate) that should not account for losing ~95% (17% to 1%) of their WP.

Whatever their WP is after scoring (let's say 40%), 1% WP would indicate that the model believes that there is a 1/40 chance of scoring a TD there.

If you randomly choose some other examples you'll see similar issues. Teams facing a 4th down (and needing a conversion, not a kick) have long odds that are disproportionate to the situation. I thought the NE-HOU game best illustrated this.

I really don't understand this algorhythm. I simply don't see how a team can have a WP of almost 75% in the first quarter after a single score and before the other team has touched the ball. It seems to me that the algorythm should not allow for a WP greater than what is average for that differential for that time of the game. How many teams win when they are up 7 points with 8 minutes gone in the first quarter? Not 75% me guesses. It seems to me that the WP fluctuations should look like the expanding universe, narrow at the beginning and wider to approach 100% as the game gets closer to end.

75% when up by 7 in the 1st quarter sounds about right to me. Though there is the slight effect of selection bias. Teams that jump to 7 pt. leads tend to be superior teams to begin with.

I think the 1% at 4th and 1 are an artefact of the caculation method.

Since there is not enough sample size for every time plus yard plus distance, Brian was extrapolating the exact curve from other data. Especially in goal to go, where the scoring probabilities rise relativ sharp with each yard gained, this does probably not give the correct numbers.

Looking at the NE game you would guess 60% of Conversion* 80% of stopping HOU * 50% in OT gives a rough estimate of 25% chance of winning

I simply don't see how a team can have a WP of almost 75% in the first quarter after a single score and before the other team has touched the ball.Well, the Pythagorean formula says a team that's better by 7 points has a 69% chance of winning even before the game starts, with the score an even 0-0. That actually scoring first and taking a 7-point lead might add another 6% doesn't seem unreasonable on its face.

For what it's worth, using the NE example of 4th and 1 on the 4 with 1:14 left down by 7, the 4thDownulator says that's a WP of .25 by going for it, and .03 with a FG.

More investigation...

Using the WP calculator, setting the down as 4th and the time to 0:01 (last play of game) at the 1 yard line-

When the score difference is -3, we get WP=.48, as expected for a FG attempt.

When the score difference is -4, we get WP=.01, which I believe is the minimum WP.

Also, by adding some time to the clock (1:30, in this example) a team will have a WP=.05 with 4th and goal from the 1 yard line, losing by 4 pts.

As we move the ball away from the end zone, we see that WP stays at .05 all the way until the 20 yard line, at which point it drops slightly along with FG percentage. This is indicative of an assumed FG attempt.

I think this shows some issues with the end game WP calculations (at least on 4th down, which might have only slight effects on 3rd, 2nd...). I'd offer a stab at the solution, but I'm not sure of the details of the calculation.

I love the EI and CBF. Personally, I'd like to see other stats as well, such as how close the game was or late game importance

@Jim Glass re: Frank Day's comment:

I think the point that Frank is making is that without seeing the data set, it seems somewhat absurd that team A, being the only team to have possessed the ball in the game at point X, and being ahead by 7 points, will win 3 of 4 games. Does a FG yield a 60+ WP? I mean, in this past week's JAX/TB game, TB was up by 14, and got blown out--and they are pretty equal teams. IIRC, CLE took the opening possession and scored a FG--and lost.

Now, if you told me that a team winning by 7 at any point in the 4th Q would go on to win 75% of the time, I'd believe you--maybe even in the 3rd Q. But just examining this week's games (small sample size, cherry picking data, etc.), by my count 5 out of 16 teams leading at the half (or in CHI's case, leading after the 1st score) LOST--plus TB was ahead by 14 mid-2nd.

Sorry--I'm not buying it. If I had time, I'd check every game THIS year only--just to see how often a team with a 7 point lead at ANY point in the game won the game. (Obviously, in some games there would be the potential that both teams would be included). I'd bet that because of these "see-saw" games it wouldn't be higher than 2/3. Mind you, I said at ANY point.

@Joseph

You don't have to check. That's what WP is for. The WP system uses data from actual game situations similar to the one in question (smoothed, of course). WP=75% means that if you went back and checked all the games in the past 10 years similar to the game situation in question that team would end up winning 75% of the time.

To be clear, Brian Burke is not sitting at his computer guessing who is going to win the game after each play. It's all data-driven.

Yes. A TD advantage following the kickoff yields about a 67%-70% WP right up until the 4th quarter. A FG advantage yields a 60% WP.

Keep in mind that fielding a kickoff is only worth about 0.6 Expected Points. Possession is obviously important, but at your own 20 or 22 is not terribly far from break even.

You forgot the biggest comeback of the week: the Vikings were down 21 and came back to win by 1 on the last play of the game. Oh wait, no, sorry, the refs stole it from them on that play, never mind.