Dallas Tops Final 2021 DVOA Ratings

Seahawks lost five games by a field goal or less.  That could be the difference between a dropped TD pass and a subsequent field goal, a missed kick or any other single play per game.  DVOA would not change much either way if five plays over the course of Seattle's season changed slightly, but they would be a playoff team in one scenario, and ended up with a losing record in reality.

It's because DVOA measures efficiency on a play-by-play basis, but the Seahawks ran so few plays compared to their opponents. They only ran 954 plays all season, their opponents ran 1,201! The Seahawks outgained their opponents in yards per play 5.8 to 5.4 but lost a lot of games because the other team ran far more plays. Part of this was due to the offense rarely turned the ball over and had the most three and outs in the league, while the defense rarely took the ball away and gave up long drives with tons of plays but was very good on 3rd down and in redzone.

Aaaarrgh!  This is so frustrating - I know one shouldn't feed the troll, but on the off-chance you actually are interested in how things work...

DVOA correlates extremely well with success on the field.  In fact, it's pretty much measuring whether each play on the field is successful or not.  This is more predictive than drives or scores (or win/loss) because there are more plays than drives, or scores, or game results.  Football, like most sports is a weighted random number generator.  It's random, because there is luck involved, and it's weighted, because some people are better at football than others.  So, the weighting is important, but the randomness isn't (because it's random).  The more things we measure, the better we get at understanding what is weighting (which is repeatable) and what is randomness (which isn't).

So, if we can agree that measuring plays is a good idea (because there are more of them), we're at success on the field.  Now, if you just want successful plays to correlate exactly with game results, then you are following the wrong sport (or, in fact, any sport - see above about randomness).  However, what's nice about football, with it's play by play nature, is that on any given play, each team is trying to 'win' that particular play.  If we can measure which team 'won' the play, that's helpful, because we know that's what teams want to do [1].  If a team 'won' every play, then you'd have to assume that this has a high correlation with drive success, and with game success (and, ultimately, season success).

DVOA isn't perfect - but the imperfections within the system do not have anything to do with a lack of correlation to 'success on the field'

[1] Actually, it's slightly more complicated, as there are times that teams are happy to 'lose' a given play because of game situation - e.g. running up the gut to drain the clock, but DVOA does try to measure these, as well.  There are also some situations early in games where although a team would prefer to 'win' the play, they may accept a 'loss' if it tells them something about how to move forwards in the future.  DVOA assumes independent outcomes of plays, which isn't completely true.

...is that I think SF blowing out DAL is highly unlikely, on the basis of recent Cowboy offenses having a habit of digging out of blowout holes (sometimes even salvaging the game entirely). I could see a blowout for 2.5ish quarters! But Dallas's offense tends to find itself even after bad starts and thus prevent full-game blowouts.

I noticed that too. It definitely could make for some one-sided (on paper anyway) Divisional matchups. Or for a team like the Bengals or Raiders to go on a run if they match up with the Titans.

I think the Chiefs aren't terribly interested in routing opponents these days.  Indeed, they seem to be the most casual of playoff teams, perhaps thinking that they are the two-time defending AFC champs and that they'll be ready when needed.  (It's a dangerous mind set, but they might be able to pull it  off.)

If anybody looks at Buffalo vs. KC and thinks the DVOA disparity means KC should be the underdog, well, I'd disagree.  

Good post but I think a comparison based on weighted DVOA might be more revealing. But it's easier criticize than do the work  🙄

More than a little luck in both one-score games and when they played their opponents, too. Baltimore is much more difficult to beat with Jackson at QB, much less beat twice because of a failed two-pointer when going for the win - and they still needed an epic choke by the Colts to get in. This season would be difficult to repeat for the Steelers.

Every time I've seen them they looked damn good, even when their top receivers are out. They're consistently good, but not dominant. Story of the year.

Dallas came on so strong at the end, not only were they DVOA champs, but Weighted DVOA champs by a more substantial margin.  Funny how models work, GB is a prohibitive favorite in the NFC right now, in large part because they will automatically survive next weekend.  Once the results from next weekend are in, there will be some teams right on their tail, especially if DAL, TB, and LAR win.

Fans make presumptions that certain teams will move forward.  Looking at another model, DAL is 80% to move on past PHL, 80% is hardly 100%.  

Dallas is in the same spot as Buffalo IMO. They've done a lot of their damage in games they were "supposed" to win (Washington, vs. Philly's backups, New York), and struggled in the games vs top opponents (Tampa, Arizona, KC), with one explicable loss on their schedule (Denver). They're a legit Super Bowl contender, but public narratives are built on a handful of games.

Dallas is the DVOA champ  in part because they just left their starters in most of the game to whup the Eagles' second string.  In contrast, the Packers had a comfortable lead over the Lions before they pulled Rodgers, and the Lions took advantage in the second half to win.

So, no.  I don't take the Cowboys seriously as a contender. I did before the Broncos went into Dallas and clobbered them.  They could have problems with SF and beyond that, I put them with the Rams as "teams that are very talented that I don't have confidence in", as opposed to Green Bay and Tampa. 

My guess is that, on average, the data will say that the blowout wins are predictive and don't need to be discounted.  the reason will be that winning by 40 points instead of 30 points will be a characteristic of teams with better depth and a more robust "system" (i.e. winning on fundamentals rather than raw talent).  Plenty of exceptions will no doubt be found, but overall as a predictive element I suspect the large wins will highlight the most solid teams.

That said, 2021 (and 2020) was more likely than most years to provide some of those exceptions, for the covid-related reasons you pointed out.  Key players (usually a QB) missing a game here or there have always put a bit of a wrench into opponent adjustments, especially when measuring the quality of a team's D.  And some teams have always dressed backups in meaningless end of season games, which throws off their opponent measurements in those games.   But depth concerns from a concentration of injuries in a single unit usually accumulate over time and those have a chance to be reflected in that team's scores over multiple games.  Covid-whacked rosters that cause a team to field a radically different unit than started the game before or the game after may impact the validity of opponent adjustments in any given game more in 2021 (and 2020) than in past years.  That may be partially offset in 2021 by those adjustments and all team DVOAs accumulating over 17 games rather than 16 games.

Their offensive DVOA for the final 6 games was -23.4%, -33.5%, -41.4%, -82.5%, -90.2%, and -72.9% which averages to an offensive DVOA of -57.3% in those games.

DEN only had 166 offensive drives, 3rd fewest in the league. The offense was 14th in yards per drive, 19th in points per drive, 16th in drive success rate, 7th in avoiding turnovers and 4th in time of possession per drive, meaning they generally ran long drives that resulted in fewer possessions, causing the low scores in their games.

The site used to be more rigorous in emphasizing that the stats were really "Aaron Jones with this scheme and these teammates produced XXX DYAR". I understand why repeating that every single time would get tedious. So yes, DYAR has the exact known flaw you mention. Especially with ONLY being able to use play-by-play data to build the metric. More data is out there now but you won't have any of it for historical and a lot of it likely proprietary and potentially in a difficult to integrate format. I'm going to use a lot of Packers example because I know them best but other teams show the same things.

So the questions becomes, are great or bad RB's really great or bad or are they simply a function of the team for which they play?

Both. That doesn't mean player DYAR is useless. But the the results of what a player does are more and more dependent on what the other 21 players on the field are doing the value does diminish. Some running plays are affected by every player on the field if there is a big cut back. Most still involve around 16 with heavy influence. But as you said. Why pay running backs big money when a hard to determine but likely large percentage of their value comes from their teammates. My guess is that about 70% of what a running back produces has just about nothing to do with the running back. DYAR clearly illustrates that. You can get similar value from very different players. I also agree with that and didn't think the Packers should have paid Jones even if the 30% of his value he controls is better than Kylin Hill or Patrick Taylor. I'm not sure how, especially with only play-by-play data you could come to a number of how much a running back matters and how much the sheme and other players matter. It would be nice. Even the stats here, adjusted line yards and RB yards are making guesses to come to those values. But I've seen Allen Lazard, a WR, make a block that got Jones 20 more yards which even the stats that are trying to separate things isn't going to see. In part because that data isn't in the play by play and in part because great or bad WR blocks are less likely to affect the play.

It isn't that running backs don't matter and that they don't have different skills and value. It's more that play-by-play data doesn't have enough information. Even Jones and Dillon who had very similar advanced and traditional stats got there differently in the same system. You need to start looking at things like yards before and after contact. Dillon had 2.1 yards before contact and 2.2 after. Jones was 2.4 before and 2.3 after. So while Jones had more ypc his portion of yards went down after contact, Dillons went up. Then you look at other things Jones had more 20+ yard runs. Dillon had more success in the "power running" game, etc. Even in the same system they are used differently. They ultimately brought about the same total value to the team but they did it in different ways.

With their (and likely all RB) receiving it's a bigger difference and I agree that the DVOA/DYAR is harder to use there. Dillon had 36.5% DVOA as a receiver and Jones only 11.9% but basically the same DYAR again (110 to 97) because Jones was targeted in situations that Dillon simply wasn't. That wasn't the whole reason that Jones had 65 targets and Dillon only 37 but it was part of it. Rodgers would checkdown to Jones even when he was covered if he was the best option and take the 2 or 3 yards. Dillon didn't see the ball unless he was wide open, and that lead to bigger gains when he did see it. This actually illustrates one of the reasons DYAR is still valuable, at least when assessing players on the same team. DYAR said they gave about the same total value, and you can then watch the games and see that Jones was good enough to give small positive or slightly negative value in situations that if Dillon was in would have been large negatives (a sack or throw away vs a 2 or 3 yard gain) but that Dillon sees none of the negative value applied. Jones was also way more likely to draw a better player in coverage if motioned out wide. It doesn't mean that they should throw to Dillon more. Though with his improvements as a route runner and his catching that may also be true. But it's value you can get from a limited metric anyway.

Now lets go to WR.  Take any top 10 receiver that you want and pair him with the above mentioned Darnold, Z. Wilson, Glennon, Fromm and Book.  Now what do you have?

WR DYAR likely has better value when looking at players on different teams than RB DYAR in part because fewer players influence the value a WR can produce. You'll have the offensive line and pass rush playing a role so 10 players there. The QB, WR, and at least 1 defender so 3 more. So yes it's still 13-16 like with a running play. But the role of the offensive and defensive lines can be nearly eliminated. Some routes are designed to make that not matter because you do things so quick that it's basically just the center, QB, WR, and 1-3 defenders that really impact the play heavily. Also if the QB is skilled enough the impact of the lines on the play diminish as well. So yes Davante Adams has an advantage over Justin Jefferson because Rodgers>Cousins and trying to parse the 421 vs 414 DYAR and 18.8 and 18.4% DVOA differences, this year, would be tough. But it's still probably fair to say that Jefferson on GB and Adams on MIN are likely still going to produce 400+ DYAR. Of course this was the best DYAR season Adams has ever had. But he also produced 215 DYAR in 2017 while 54.5% of GB's totals passes were thown by Brent Hundley. He only had 139 DYAR in 2019. So yeah Adams has been a top receiver catching passings from a -28.3% DVOA QB which isn't quite as bad as breaking the -30% threshold that Darnold and Wilson did, but still shows that while his DYAR may have been depressed his ability to get himself open and hold onto contested passes still means a lot. Rodgers 334 DYAR was completely wiped out by Hundley's -396 but Adams still had his 215 and he even made quick reads with 54 with Hundley throwing to him so it wasn't just the passes he got from Rodgers.

Sure when I look at the QB's, I can not argue that Brady, Rodgers, Prescott and Mahomes are great, while Darnold, Z. Wilson, Glennon, Fromm and Book gave us some ghastly QB play this season which is well reflected in DYAR.

QB DYAR as mentioned is probably the most accurate DYAR gets. The system a QB plays in still matters of course. There is a reason Rodgers has 4 DVOA titles (2011,2014, 2020, 2021) and no DYAR titles he's usually in the 8-14 range in attempts (4 and 5 are his highest ever rankings). But of course when you take the two numbers together even with system differences you can get a pretty solid idea of who is and isn't a good NFL QB. You also have the injury replacement numbers that show that even more. In the years he's been hurt the drop off to his back-ups has been in the 30 - 60% DVOA range (2017 Rodgers 7.8 to Hundley -28.3 and 2013 Rodgers 25.4 to Tolzien -9.1, Wallace -54.3, Flynn is harder to tell since he played for 3 teams that year but was -24.5 overall).

Finally the running QB.  During the season, a post was made by an FO writer who sent a 2018 article explaining Lamar Jackson.  If Lamar breaks the system so will others.  Lets look at this season.

 I joked in several Quick reads about Brady having the best QB rushing DYAR that week. I think Pat has addressed part of the fundamental issue with this. Run vs pass is arbitrary and doesn't really describe what the play is trying to accomplish. The easiest example of course is a toss to a running back (run) vs a short out in the flat to a running back. Just because in one the ball went backwards 1 yard and in the other it went forward 1 yard even if the blocking and everything else is identical one is a run, one is a pass but they are trying to do the exact same thing. Get the ball in the hands of your running back who is going towards the sideline outside of the offensive line and all the defenders near the line. There are plenty of other examples.

I think there are other offensive systems that could "break" DVOA as well. But the data DVOA has is play-by-play and so what is a pass and what is a run is defined by the NFL and even if they wanted to try and change those categories they couldn't. Adjusting baselines could help but it can't solve it.

DVOA works because it makes sense that if you are more efficient on your plays, on average, than your opponent you are more likely to win. Obviously there are other factors that go into W-L. It's possible that a good machine learning system could help tweak the baselines and success values to help make more of this measurable from the play-by-play. I also suspect that the predictive value of the model could be improved by running two models the purely descriptive that may include penalties and turnovers differently so that more individual games line up with the W-L result and then another model that uses that as an input along with some other potential factors to give the descriptive. I think there might be something you could do with snap counts to change the weightings that get applied to things as well. But testing all that takes time I don't have and they might not have and even if it does improve things it might be a very small improvement and it very likely doesn't change the DVOA/DYAR per player. Though if you could suss out a bit better how much of a players value comes from himself and how much comes from other players you could apply some crude weights. If you are pretty sure that the offensive line is worth about 50% of the value of a run you could start giving 10% of every run to each lineman and 2% to the TE, and 5% to the QB (or 15% to the QB if you find that running QB offenses help RB more). That's actually some of the stuff that a good machine learning system could help try and figure out. You might even be able to get more refined since the play-by-play data does have some some granularity on deep, short, and direction etc. A QB or WR may be worth 80% of the value on a deep pass but only 50% on a short.

What we really want is probably the data that is currently being captured by systems like AWS next gen stats. They don't just have the play by play they have player and ball velocities and can generate the route combos and actually see if the blocking failed on a run. That data you can start to use to build much more refined metrics around teams and players. But you can't go back in time with it. The standard of a play-by-play means that if you can get the data or video of a game you can still use it in the same system with mostly apples to apples comparison. DVOA still loses a lot of history only going back to 83, but that's better than just the 2 -5 (at most) of the data level I was just talking about could give. We all appreciate DVOA and we all love that feedback has helped make it better, but you output can only be as good as your inputs. It's better than things like SRS because it has better inputs, but we can't forget the input is still limited. I think you can massage a bit more out of it but eventually there will have to be another line like the current 1983 one. Base play-by-play can get us to here and they can keep generating that. But I think you'll need another model that can pull in stuff already mentioned and that will likely be even more predictive and/or help with describing what we saw.

Is there any way that a feature can be added so that we can type in our screen name and get to all of our prior posts?

Older versions of the site used to have this. I used it all the time to help follow threads I had posted in. Now the best you could do is go to google and do a site:footballoutsiders.com jheidelberg Lamar Jackson for you search terms. I'm also saddened they have gone away from formulaic URLs for the DVOA and quickread articles. I was able to flip through all the 2017 Quick reads in a couple minutes just by changing the url to find the Adams link I used earlier. It would have taken probably 10 times longer to find that in 2021 format since the URLs are now the article titles and even using the menus to just go to Quick reads I'd have to guess a bit since if I go to say page 8 of those results I'll get titles like Week 11 Quick reads but a date of November 19, 8:39pm ET. I have no idea what year that is. Turns out it's 2018 and that is only because it's in the URL. Well that article has some graphs and stuff that say 2018 but not all. It has always driven me nuts that the datelines never had the year. Give us 3 years of the new URL formats and finding older references will get even trickier.