Varsity Numbers: Revisiting Win Correlations

When we last looked at WinCorr -- the correlation between any given statistical category and wins/losses -- we went in two different directions: we first looked at which categories were most directly tied to wins and losses in college football as a whole, and we looked at how the WinCorr differed from individual teams. As the second full year of play-by-play data starts to come together, I thought it was a good time to revisit the concept, and see how the WinCorr figures have changed.

National WinCorr

Let's start with some ground rules. First, unlike with the initial national correlations column, we will only be looking at data for conference play. From a talent perspective, conference games are going to be closer on average than most non-conference games, so these numbers will give a better idea of which stats most directly impact football games in which both teams have a chance to win.

Second, we will be throwing 2nd- and 3rd-level EqPts and PPP (Points Per Play) into the mix. This concept was discussed last week, but looking at the WinCorr values of these stats will tell us just how seriously to take them in the future. These figures will be added to the 2007 stats, but they haven't been yet.

Win Correlations for Varsity Numbers statistics
Stat Scenario 2008 WinCorr 2007 WinCorr Diff
S&P Overall 0.723 0.566 +0.157
Points Per Play (PPP) Close-Game* 0.679 0.638 +0.041
S&P Close-Game 0.664 0.632 +0.032
PPP Overall 0.634 0.579 +0.055
3rd-Level PPP Close-Game 0.633 N/A N/A
Total EqPts Overall 0.622 0.564 +0.058
2nd-Level PPP Close-Game 0.618 N/A N/A
Actual Score Overall 0.617 0.681 -0.064
Passing PPP Close-Game 0.605 0.546 +0.059
Passing PPP Overall 0.603 0.529 +0.075
Passing S&P Close-Game 0.599 0.571 +0.028
3rd-Level PPP Overall 0.597 N/A N/A
Total 3rd-Level EqPts Overall 0.595 N/A N/A
Passing S&P Overall 0.590 0.541 +0.049
Total Rushing EqPts Overall 0.580 0.568 +0.012
Total Non-Passing Downs EqPts Overall 0.578 0.541 +0.037
Total 3rd-Level EqPts Close-Game 0.568 N/A N/A
2nd-Level PPP Overall 0.567 N/A N/A
Total 2nd-Level EqPts Overall 0.562 N/A N/A
Total Non-Passing Downs Rushing EqPts Overall 0.557 0.558 -0.001
*The Close-Game scenario once again represents only plays that took place while the game was close. "Close" is defined as within 24 points in Q1, 21 points in Q2, and 16 points or less in Q3 and Q4.

The first interesting thing to note here is that the Win Correlations for almost all of the primary statistics went up in 2008. Whether there is a specific reason for that or not is hard to gauge, but it probably did not hurt matters that there were fewer upsets in college football's FBS this season. What was supposed to happen did happen most of the time, and that probably helped strengthen the correlations.

A lot of statistics saw relatively minor shifts overall, but it is heartening to see S&P, the measure behind a large portion of Varsity Numbers analysis, assert itself as the single most important stat measured here.

Something else to notice: In such an offensively explosive season, with different versions of the spread offense catching fire in many locations (including much of Big 12 country), you had to be able to pass to win. Obviously, there are exceptions to that rule (Georgia Tech), but proficiency in the forward pass undoubtedly became more important in 2008. Meanwhile, the concept of leverage -- the ability to keep your offense out of Passing Downs -- took on less importance as a host of teams had disproportionate success on Passing Downs. It's still a solid concept, but this season was pretty ridiculous from an offensive standpoint, as exemplified by a Heisman race that saw three quarterbacks (a fourth, if you include Graham Harrell) putting up numbers and wins that would have made them the runaway winner in most seasons.

Other notes:

  • It is a bit surprising to see S&P that high considering how much more weight PPP (the "P" in "S&P") seems to carry over success rates (the "S").
  • It appears that last week's main topics -- 2nd- and 3rd-level EqPts and PPP -- are indeed important, telling statistics that will be focused on more in the future. It makes sense, really. Second-level EqPts take into consideration your ability to move the ball on given downs, while 3rd-level EqPts take into consideration your ability to move the ball on both given downs and given distances. The more you can both move the chains and eat up large chunks of the field at a time, the more successful you will be.
  • When all FBS play-by-play data has been entered (in about another month or so), "+" numbers will play a role in this WinCorr concept again. For now, we are sticking with the straightforward numbers.
  • Also, the plan is still to add the "standard" stats (rushing yards, passing yards, first downs, et cetera) to the database this off-season so we can see how much more useful (if at all) the higher-level stats actually are.

Team-Specific WinCorr

Here's what I said about team-specific Win Correlations back in October:

Unlike a lot of statistics, though, Team WinCorr offers an explanatory viewpoint instead of an evaluative one, and that makes the sample size issue more palatable. Team WinCorr is meant to complement what you already know about a team, rather than tell you something you did not.

I also mentioned that it was impossible to say yet whether the team-specific correlations were somewhat steady from year to year, or whether they changed rather significantly. With another year of data, let's take a look. This still isn't a truly long-term view, but it should tell us something. And for continuity's sake, we'll look at the same two teams we looked at in October: Florida and LSU (just ignore the part in that October column where I talked about the Florida offense's "slight regression"). These are correlations for all games, not just conference games.

Florida WinCorr
2008 Top WinCorr Category 2007 Top WinCorr Category
Defensive Q3 EqPts
Defensive Rushing EqPts, third quarter
Defensive PPP
Defensive S&P
Defensive EqPts
Defensive Passing EqPts, third quarter
Defensive PPP, close games
Defensive Passing PPP
Defensive PPP, non-passing downs
Defensive S&P, close games
Offensive Rushing Success Rate, red zone
Offensive S&P
Offensive S&P+, close games
Offensive S&P, close games
Offensive Success Rates
Offensive Passing Success Rates, close games
Offensive PPP, close games
Offensive Passing S&P, close games
Defensive Rushing S&P+, non-passing downs
Defensive S&P, pressure situations

That whole "continuity from year to year" thing? Throw it out the window. Florida's 2008 numbers could not possibly be any more different than the 2007 numbers. As always, we are dealing with very high correlations here, so the categories at the top of the 2007 list were still quite strongly correlated to wins and losses, but the most important categories completely changed. Part of this can be explained by the consistency of Florida's offense over the last two-thirds of the season, but part of it can also be explained by the fact that Florida only lost one game and only scored less than 75% of a single game's points three times; in Florida's one loss, Mississippi took advantage of repeated breakdowns in the Gator secondary, and that could be why these numbers made it to the top of the list.

Either way, the list changed quite a bit. What about LSU?

LSU WinCorr
2008 Top WinCorr Category 2007 Top WinCorr Category
Defensive PPP, close games
Defensive Line Yards, non-passing downs
Defensive Passing S&P
Defensive Passing PPP, close games
Defensive S&P, close games
Defensive EqPts
Offensive Rushing EqPts, non-passing downs
Defensive Passing PPP, first quarter
Defensive PPP
Defensive Rushing EqPts, non-passing downs
Defensive Passing EqPts
Opponents' score
Defensive Passing S&P
Defensive Passing PPP
Defensive EqPts+
Defensive EqPts
Defensive S&P+
Defensive Passing Success Rates
Defensive Success Rate, Passing Downs
Defensive Passing S&P+

First of all, these are really high correlations. LSU's degree of success was almost directly tied to their degree of success in a series of defensive categories. Part of that is because their offense was pretty consistent. They were almost guaranteed to put together pretty decent (but never spectacular) offensive numbers and give up an interception for a touchdown in every game.

Okay, that's a broad generalization. Regardless, the offense put up good numbers in the SEC, but they had a bad turnover margin (an even worse turnover points margin with Jarrett Lee's seven pick-sixes), and their defense came and went. They missed both the personnel that left after the 2007 season (Glenn Dorsey, Craig Steltz et al), and apparently the coach (Bo Pelini left to take the head coaching job at Nebraska). When the defense put things together, LSU won, no matter what their shaky series of quarterbacks (Lee, Andrew Hatch, Jordan Jefferson) did. When they didn't, they lost.

The categories at the top of LSU's list did not change as significantly as those at the top of Florida's, but it is clear that this list is quite fluid from season to season.


Here are some responses to comments from last week's column.

If the data is there, showing how running backs from the past ~5 years have fared in 3rdPPP and whether it is an indicator of NFL success or at least future college success would be interesting.

This is one of the great things about writing for Football Outsiders--there is endless potential for "college-to-pro" data when I get the college data filled in a bit more. Right now I have less than two full years of play-by-play data, but I will eventually have pulled together most data for at least 2006, 2005 and maybe 2004, depending on how reliable schools are about posting (and keeping posted) their Automated Scorebook play-by-play entries on their websites. Going forward, we should pretty quickly have a healthy amount of data to look at things like this, and with the wealth of NFL data that FO obviously provides, it opens the door for some interesting philosophical "What leads to success in the pros?" questions.

In other words, the answer is "the data isn't there", but it will be.

What do you think about Beanie Wells ranking so low? Do you think he is a talented player who will continue to fail to get yards when they are needed and rack up 5 yard gains on 3rd and 6 a la Willie Parker or is this a failure to take into account that his quarterback is a poor passer and the line is stacked every play? Or is 63rd not so bad on a list littered with quarterbacks?

Being that the sample size only included 86 runners, 63rd is pretty shaky. While the historical data gets compiled, Beanie Wells can certainly serve as a case study next year in the NFL, assuming he goes pro. On the flipside, Shonn Greene can do the same. He's a couple inches shorter than Beanie, but they're both stout, successful backs, and Greene's 3rdPPP numbers are much better than Wells'. Greene is old for college (23), so you have to figure the odds of his declaring for the draft are pretty good, and if 3rdPPP is a good predictor for NFL success, then you should see Greene at least slightly exceed expectations, while Wells struggles a bit. It will be interesting to watch.


2 comments, Last at 20 Dec 2008, 6:16pm

#1 by John (not verified) // Dec 19, 2008 - 4:37pm

The quarter variations seem likely to be inherently random. Florida did not do poorly in games in which their 3rd quarter defense is bad because they are 3rd quarter chokers; they just happened to underperform in the 3rd quarter in the 1 game they lost, which they lost largely due to poor defense.

The other correlations appear more informative about the statistic being correlated with wins than they do about individual teams, making them useful (and arguably interesting) only to those with an inherent interest in the statistics rather than being something the average fan could use or would want to use. This is not in itself bad, but could explain why commentary has been low.

That 3rd-level PPP has a weaker correlation than regular (1st-level?) PPP seems concerning. Why would taking down and distance into account make the correlation weaker? Does doing so punish teams that are winning and running out the clock at the end of games by running conservative plays that do not necessarily help them get to the end zone but may help them win the game? Or are there not enough games in a college football season to get an effective sample size? Something seems wrong when the "actual score" win correlation drops from .681 to .617 in a year.

Points: 0

#2 by Bowl Game Anomaly // Dec 20, 2008 - 6:16pm

I would be wary of drawing any conclusions from win correlation data. If it is meaningful, it probably should correlate year-to-year.

(Formerly "The McNabb Bowl Game Anomaly")

Points: 0

Save 10%
& Support Bill
Support Football Outsiders' independent media and Bill Connelly. Use promo code WRITERS to save 10% on any FO+ membership and give half the cost of your membership to tip Bill.