Patriots offensive success dwarfs Giants
Brian Burke, of the NFL Stats Blog, notes today that the more possessions each team has in a game, the more likely the favorite will beat the underdog. That makes sense, of course, and for the same reason that the Royals may have a 30% chance of beating the Yankees in one game, but roughly a 0% chance of beating them in 162 games.
Burke's simulation shows that, with 9 possessions per team, the Patriots have a 71.4% chance of beating the Giants. But with 13 possessions per team, their winning percentage rises to 76.5%.
That is: the Giants can improve their chances of winning by almost 25% if they try to play really, really, slow. (Of course, they can improve their chances even more by scoring lots of points, and preventing the Patriots from scoring.)
Anyway, as I wrote, this isn't really all that surprising. But what IS surprising – at least to me – is how much better the Patriots are per possession:
The Patriots scored a touchdown on 42% of their possessions
The NYGiants scored a touchdown on 21% of their possessions
The Patriots were *100%* better at scoring touchdowns than what is presumably the second-best team in the NFL!
On defense, they were only 10% better at *preventing* touchdowns. I wonder what that means ... does it mean the Patriots have a better offense than defense? Is the Giants defense especially good (assuming 90% as good as the Patriots must be exceptional)? Or does it mean that, in general, defenses in the NFL vary a lot less than offenses do?
Geez ... twice times as successful on offense. That doesn't include field goals, but still ...
posted by Phil Birnbaum @ 1/29/2008 04:07:00 PM
16 comments
16 Comments:
Phil-Actually the Giants scored on about 21% of their possessions. The Patriots gave up TDs in 17% of opponents' drives. My post might not have been very clear.
Your observation stands, however. The Patriots scored TDs just over twice as often as the Giants. That's a remarkable difference by any measure.
Oops, my fault. Your post was perfectly clear. Will fix.
Take a look at Football Outsiders advanced statistics, DVOA (which uses play by play data) in particular. The Patriots had a historically great offense this year, and defensive efficiency had a small standard deviation across the NFL (compared to past seasons). So in short, yes, the Patriots offense is much better than their defense, but their defense is also pretty good.
Go back far enough and its clear that by the game's nature defensive efficiency varies less than offensive efficiency. In college you don't even need a schedule correction to see that.
What would be interesting is to study the relative persistence of team O/D efficiency across seasons.
Eight possessions seems like a major change in the game in terms of big plays, punting, first downs, etc. By comparision a change in the chance of winning of only 5% doesn't seem as extreme
Was a simulation even needed? Granted the changing play in the final few minutes of the game, what if we just looked at half-time "win%", where presumably you only have half the possessions?
I'm going to guess that if you looked at the 8 teams with the highest points differential for a season, that their win% if the game ended at the half would be at the midpoint of their actual overall win% and .500.
Specifically, the win% in Brian's data has an almost perfect match to this equation:
Wins divided by Losses
= 1.1 ^ possessions
So, if you have 12 possessions, the above equation will give you 3.14. That is, 3.14 wins per loss, which is a win% of .758. Brian's simulator said .755.
Here's who Brian's simulator compares to my above equation:
9 0.714 0.702
10 0.727 0.722
11 0.740 0.740
12 0.755 0.758
13 0.765 0.775
So, a 12-possession game for the Pats has the equation at .758. If I extend back to only 6-possessions, I get a win% of .639. And that is pretty much the half-way point between .500 and .758.
Tango: very nice equation! Never thought it would work that way, but it makes sense.
" ... their win% if the game ended at the half would be at the midpoint of their actual overall win% and .500."
Actually a little higher, right, because of the way odds ratio works? It would be whatever you get from the square root of the odds ratio, wouldn't it?
It would be exactly that, but I didn't want to confuse any of your other readers.
(1.1^6) ^ 2 = 1.1 ^ 12
Which is of course:
(1.1 ^ 12) ^ (1/2) = 1.1^6
See kids? Math is useful.
Math rocks! And sometimes I even understand some of it.
I agree possession frequency is very important. The same argument applies in basketball - the better team wants a greater sample size so that the standard deviation of the mean is reduced (at rate of N^0.5)
But during the game, competing incentives come into play in that the leading team also wants to slow the game down (in terms of # possessions). So when a favored team is losing, there is an additional push by them to speed the game up, while the dog wants to slow it down... i would like to study which team succeeds in imposing their preferred rate of possessions in the second half of games.
Back to football, I've found the % of TDs to be a statistic with too much noise (small sample size - dependent on field position...). However a very good predictor of TD% is avg yards per play. And teams have about 1000 plays per season. (With OLS a 1 yard/play increase in offense results in 9.5% increase in TD % -- FG% is virtually flat at 16% for all teams)
Given this statistic, NE led the league in yards/play at 6.2, while NY was exactly average at 5.1 (defenses were both just above average at around 4.95). This tells me that of the 21% TD% differential seen this year, about half was due to NE having a truly more productive offense, and half was luck/noise/error. (6.2-5.1)*9.5% = 10.5%
Thus, my money (sad to say) is on Eli and the "not as bad as you think" Giants.
I agree with Nate. In fact, any extreme statistic is probably due, in large extent, to luck. If there are several reasons for the luck, they should all be expected to be somewhat positive.
The big question, for me, is: does the betting line take that into account? It gives the Patriots an 81.6% chance of winning. That seems like it does realize the Pats aren't as good as their record, considering (a) they haven't lost a game, and (b) Eli Manning has never been considered that great.
But I don't know for sure. What I do expect is that whatever statistic you use to judge the Patriots, you have to regress it to the mean to get an accurate estimate. *How much* to regress is, of course, the big question.
And the fact that you have to regress doesn't mean you don't admire the actual record. You can still marvel at 18-0, and Reggie's three home runs in one WS game, and Barry's 73 home runs.
good point, phil. in sports we love many different things about the games while in the moment, but it is the outlier event that we remember years later.
fwiw my model (based on the yards/play metric) has ne at 70%. but my model has been down on ne all year.
a quick check of my stats for the past 5 years shows ne as the 3rd best offense in that period of time in terms of yards/play. they are behind 04 colts (6.7!) and 04 vikings (6.45)
Now, of course, the more granular the statistics you use for a team, the less the noise, and the more the resultant w% gets "automatically" regressed.
The worst "handicapper" will use teams' w/l % to handicap a contest. That is horrible of course, and each team's w/l record would need to be regressed a lot (not to mention normalizing or adjusting for strength of schedule).
The second worst handicapper is going to use average point differential per game, which is the same thing as or similar to using a pythag record for baseball. Even then, you would still have to regress quite a bit, I would think. And adjust for context (that goes without saying for any stat or stats).
The next best handicapper will look at yards per possession on offense and defense and add in special teams, field goals, turnovers, etc.
And the "almost best" cappers will adjust that data for context, relevance, etc., such as kneel downs, spikes, etc.
But, there are two critical things or mistakes that most if not all handicappers miss/ignore.
That is, one, regression, regression, and regression. Even at the most granular level, you HAVE TO regress all of the stats, some a lot more than others. For example, turnovers are mostly luck (some turnovers more than others), and thus have to be regressed a lot.
And two, and I have not read the blog that Phil refers to, and this is the one thing that almost all handicappers err on, is the idea that one year of data is "magic." That relates to the first thing of course. If you are going to use one year of data to evaluate a team, you are necessarily going to have to regress a lot, especially when you are talking about only 16 or 18 games (I don't know how many "trials" that correpsonds to in the NFL, as compared to the other sports - Tango had a post a while ago wherein he showed how many NFL games is equivalent to how many NBA games, NHL, MLB, etc.). And more importantly, why not use more than 1 year of data on NFL teams and/or the individuals on the teams? Can you imagine using only one year of data for MLB player and team projections? It wouldn't get you very far. Yet, almost every handicapper and (betting) analyst known to man does this.
It ain't going to get you too far. That is how the line is created in the first place, which makes the line quite accurate as compared to any "system" that also uses one year of data (and especially with little or no regression - the linesmakers do not FULLY comprehend the notion of regression, although they realize to some extent that NE is not "really" a 19-0 team, etc.).
mgl, i agree with much of what you said. note that using some sort of yards/play metric (cleaning up for kneel downs etc... and maybe using a log transformation) will eliminate the random effects of turnovers.
It also provides 1000 data points, which leads me to wonder if 1 year of data is in fact enough. Maybe you can blend in part of the previous season into your analysis, but there is sufficient player turnover to make me believe that what happened 2 or 3 years ago isn't particularly relevant -- this is especially true if you are using a granular data set like per-play.
Nate, without turning this into a handicapping thread, the more data, the better. So yes, what happened (let's say at least) one year ago is important. In baseball, we weight the "importance" of data by recency. That is true in all sports. The problem with using multi-year data in football, as opposed to baseball, and to some extent NBA, is that it is hard to break the data down into individual contributions. Football is of course much more of a "team" sport than is baseball (which is not at all, and that which has a "team" element can still be broken down into the individual components).
Which means that you are left with several choices if you wish to use multi-year data. One, figure out how to use individual data and put it all together at the team level. Or use team data from previous years (the nice thing in football is that you don't have as much player movement as you do in baseball - I don't think) and adjust for player movement. Or some combination thereof.
But I still think that a multi-year approach is the right way to go, which is one reason why the average good handicapper (who does not use more than one year of data) does not do particularly well. Go to this web site, and you will see that even the best models break about even after the juice in NFL, and worse in the NBA. THat is assuming that they accurately track these handicappers/systems and that they accurately track the prevailing Las Vegas lines (which is always a dubious assumption, even on the honest tracking web sites). The reason is usually three-fold: One, not using granular enough data and properly adjusting for context; two, not regessing enough or at all; and three, not using enough data.
Now here we are talking about handicapping a Super Bowl game after the season is almost over. What about handicapping over an entire season? At what point is your data adequate enough in terms of size? You would almost have to use prior year data at the beginning of the season. If you don't, you have almost no shot...
When I said, "Even the best sytems," I meant ALL of the systems combined being tracked on this site, with the assumption being that the systems/handicappers that are being tracked on this site are some of the best ones out there, or at least exclude most of the "scam" handicappers, which comprise 90-98% of the field...
Post a Comment
<< Home