The Top Dog: Dallas Cowboys
I know what you're thinking: "how can a 2-2 team with a negative point differential be the top team in the league?" While both of those facts are true, the Cowboys have also had an unfavorable schedule, as they have faced three top-12 clubs. Early in the season, this model will heavily weigh "strength of schedule," so there you go. Also, Dallas feels like one of those teams that can put up good numbers on a regular basis without winning many games. A paper team, if you will. The 'Boys get a break this week to prepare for the Patriots offensive attack.
The Undefeateds: Green Bay and Detroit
Green Bay being ranked outside the Top-5 is due to them allowing a huge amount of passing yards. In fact, they are second worst in the league in terms of yards allowed through the air while they are second best in the league when it comes to yards allowed on the ground. They probably aren't the second worst pass coverage team or the second best at stopping the run, as their offensive firepower forces teams to pass on them more often than they would against other opponents. Detroit's problem, on the other hand, is their strength of schedule. They've gotten to face the Chiefs already, and they barely sneaked out a win against the lowly Minnesota Vikings.
The Winless
According to these rankings, the Colts are the best winless club, with the Dolphins, Rams and Vikings all following suit. The Rams will still be in this group next week since they have a bye, and the Colts are most likely to pull out a victory against the Chiefs this weekend.
Other Fun Facts:
- The Falcons, who were picked to win the Super Bowl by many, are ranked second to last. Sure, they're 2-2, but one of those wins came against the Seahawks, who had a chance to win the game on their final drive. Watching the Falcons with my own two eyes, they didn't look like anything special. Julio Jones is nice and all, but they really could have used their other picks to improve this club in a couple other places (such as the defensive interior).
- The 49ers are the worst team in sole possession of first place in their division, but the Buccaneers are ranked much lower and are tied for the NFC South lead with the number two Saints. San Diego is the next worst, perched right in the middle of the rankings.
- Oakland is ranked in the top half despite being the worst team in all of football when it comes to how often they are called for penalties. Tenessee and New England are also higher on the list than you'd expect to find a title contender.
RANK | TEAM | LAST WK | GWP | Opp GWP | O RANK | D RANK |
1 | DAL | 2 | 0.67 | 0.55 | 2 | 6 |
2 | NO | 1 | 0.65 | 0.52 | 6 | 7 |
3 | HOU | 8 | 0.62 | 0.52 | 5 | 8 |
4 | BAL | 3 | 0.61 | 0.53 | 11 | 2 |
5 | TEN | 7 | 0.61 | 0.49 | 3 | 18 |
6 | GB | 9 | 0.60 | 0.50 | 4 | 22 |
7 | NE | 5 | 0.60 | 0.50 | 1 | 28 |
8 | DET | 10 | 0.59 | 0.46 | 12 | 4 |
9 | PIT | 14 | 0.57 | 0.49 | 14 | 11 |
10 | WAS | 6 | 0.56 | 0.52 | 17 | 5 |
11 | NYJ | 4 | 0.56 | 0.57 | 20 | 1 |
12 | BUF | 11 | 0.55 | 0.51 | 15 | 14 |
13 | NYG | 13 | 0.55 | 0.47 | 8 | 21 |
14 | OAK | 12 | 0.55 | 0.55 | 9 | 12 |
15 | CIN | 19 | 0.52 | 0.48 | 18 | 17 |
16 | SD | 21 | 0.50 | 0.44 | 16 | 20 |
17 | PHI | 15 | 0.48 | 0.43 | 10 | 29 |
18 | SF | 29 | 0.47 | 0.50 | 21 | 16 |
19 | ARI | 17 | 0.47 | 0.47 | 13 | 23 |
20 | DEN | 24 | 0.47 | 0.57 | 23 | 13 |
21 | CAR | 25 | 0.45 | 0.49 | 7 | 32 |
22 | CLE | 28 | 0.44 | 0.49 | 26 | 15 |
23 | JAC | 20 | 0.44 | 0.57 | 32 | 3 |
24 | CHI | 18 | 0.44 | 0.51 | 27 | 9 |
25 | IND | 23 | 0.43 | 0.51 | 25 | 26 |
26 | MIA | 16 | 0.41 | 0.54 | 22 | 27 |
27 | TB | 27 | 0.41 | 0.43 | 19 | 31 |
28 | STL | 22 | 0.38 | 0.55 | 30 | 10 |
29 | MIN | 26 | 0.38 | 0.47 | 24 | 25 |
30 | KC | 30 | 0.37 | 0.51 | 29 | 19 |
31 | ATL | 31 | 0.32 | 0.41 | 28 | 30 |
32 | SEA | 32 | 0.32 | 0.46 | 31 | 24 |
TEAM | OPASS | ORUNSR% | OINT% | OFUM% | DPASS | DRUNSR% | DINT% | PENRATE |
ARI | 6.8 | 49 | 3.1 | 1.6 | 7.0 | 56 | 2.6 | 0.57 |
ATL | 5.9 | 35 | 2.4 | 1.9 | 7.7 | 54 | 4.3 | 0.40 |
BAL | 6.3 | 36 | 2.1 | 2.1 | 5.3 | 63 | 4.0 | 0.38 |
BUF | 6.8 | 47 | 2.7 | 0.0 | 7.2 | 52 | 5.3 | 0.31 |
CAR | 7.8 | 42 | 3.1 | 0.5 | 8.0 | 55 | 2.1 | 0.51 |
CHI | 5.8 | 40 | 3.1 | 0.6 | 6.9 | 55 | 1.8 | 0.43 |
CIN | 6.2 | 48 | 2.9 | 0.5 | 5.4 | 57 | 0.8 | 0.29 |
CLE | 5.3 | 45 | 1.7 | 0.5 | 6.2 | 57 | 2.6 | 0.46 |
DAL | 7.8 | 35 | 4.3 | 1.4 | 5.7 | 66 | 2.7 | 0.41 |
DEN | 5.9 | 37 | 4.2 | 2.6 | 7.6 | 64 | 1.4 | 0.37 |
DET | 7.2 | 36 | 1.9 | 0.0 | 5.5 | 63 | 4.6 | 0.46 |
GB | 8.6 | 41 | 2.1 | 0.9 | 7.7 | 60 | 4.9 | 0.33 |
HOU | 7.8 | 44 | 2.7 | 0.5 | 5.8 | 52 | 3.0 | 0.46 |
IND | 5.0 | 45 | 0.7 | 3.2 | 7.4 | 59 | 2.3 | 0.20 |
JAC | 4.5 | 37 | 5.4 | 0.5 | 6.4 | 58 | 3.7 | 0.33 |
KC | 5.3 | 31 | 4.5 | 2.7 | 7.0 | 54 | 3.7 | 0.40 |
MIA | 6.3 | 42 | 3.6 | 1.0 | 8.0 | 50 | 1.4 | 0.41 |
MIN | 5.1 | 45 | 1.8 | 0.5 | 6.9 | 55 | 1.9 | 0.54 |
NE | 9.2 | 48 | 3.1 | 0.0 | 8.5 | 54 | 4.2 | 0.50 |
NO | 7.3 | 47 | 2.3 | 0.9 | 5.8 | 59 | 1.2 | 0.24 |
NYG | 7.2 | 39 | 1.6 | 1.0 | 6.3 | 51 | 2.8 | 0.38 |
NYJ | 6.0 | 34 | 3.4 | 2.7 | 5.6 | 61 | 5.0 | 0.38 |
OAK | 7.4 | 46 | 2.5 | 1.0 | 6.2 | 52 | 1.8 | 0.68 |
PHI | 7.1 | 50 | 3.4 | 2.3 | 6.5 | 54 | 1.7 | 0.39 |
PIT | 6.9 | 40 | 3.6 | 3.0 | 5.0 | 57 | 0.0 | 0.38 |
SD | 7.6 | 45 | 3.8 | 0.9 | 6.7 | 59 | 3.7 | 0.39 |
SF | 5.9 | 34 | 0.9 | 1.0 | 6.8 | 65 | 3.8 | 0.49 |
SEA | 5.0 | 33 | 3.0 | 1.1 | 6.9 | 63 | 1.5 | 0.44 |
STL | 4.3 | 34 | 0.6 | 2.7 | 6.1 | 54 | 2.2 | 0.56 |
TB | 6.1 | 48 | 2.7 | 0.5 | 7.1 | 57 | 1.4 | 0.31 |
TEN | 8.2 | 31 | 2.3 | 1.1 | 5.1 | 56 | 3.2 | 0.50 |
WAS | 6.1 | 46 | 3.5 | 0.9 | 5.4 | 59 | 2.1 | 0.34 |
Avg | 6.5 | 41 | 2.8 | 2.2 | 6.6 | 57 | 2.7 | 0.41 |
I have been working on a similar model as this and was also getting the Cowboys as a top team but was wondering how much weight to give to schedule. I ran a regression with last years numbers and this years of actual winning% as the dependent by my modeled winning% and modeled strength of schedule and the coefficients I got weighted the model winning% about 4-5 times higher than the schedule strength. I would recommend doing that with yours to see what you come up with.
Stephen,
The opponent adjustment ranking is really straightforward if you are familiar with logit models.
Each team has its own logit value based on the sum of the product of their stats and regression coefficients.
I average the logit values of each team's set of to-date opponents. That value is added to the team's logit.
Then the probability estimate is calculated (either GWP or game probability) using the standard math.
You can do second and third-order adjustments too, because as each team's opponent-adjusted logit changes, their opponent's logit values change.
I've probably asked this question before, but do the ORANK and DRANK numbers reflect strength of schedule adjustments?
Or are they based on the raw logit calc?
Brian - thoughts on the Bengals being ranked better than the Chargers?
That leads me to another idea - one which may take years to get any meaningful information from. Despite the objective nature of the model (which, by the way, I am a huge advocate of), I find it extremely hard to believe the Bengals have performed better than the Chargers this year (and thus I don't believe they will continue to perform better). Of course, everyone will nitpick here and there about different teams, but I think the vast majority of us can agree with the Bengals/Chargers disparity.
Proposition: why don't you (or anyone else) make some subjective assessments each week, and then track them as the season goes on? It's something that can be quantified. Of course, the subjective predictions become bolder the further apart the two teams are, and should be "rewarded" more. Predicting the Chargers will finish the season better in the rankings than the Bengals isn't so difficult to do since they are already so close.
Someone may want to predict that Atlanta will finish the season ranked better than say, Buffalo - if it comes to fruition, it would be valued greater than the Chargers/Bengals prediction.
Granted, if enough people do this, you're guaranteed to get some football "gurus" out there just by chance...but I think something meaningful can be attained by doing subjective predictions using your objective rankings as a baseline.
I agree with you that it's more likely than not SD outranks CIN by season's end. Additionally, even with these rankings, SD has a clearer path to the playoffs by far than CIN. But, as you mention, the value here is in their blind objective ignorance of what we're 'supposed' to think. There is no shortage of subjective opinions out there already. Mine would be no better than everyone else's. But maybe a "model-informed" wisdom-of-crowds readers' poll would be cool, and possibly the most dead-on accurate ranking out there.
In case anyone is wondering, CIN is superior to SD in: ORUN, OINT, OFUM, *DPASS*, and PENRATE.
CIN has had a slightly tougher schedule to date.
The real value to me is that, until your comment, I had no idea CIN def pass eff was so good--4th in the league behind BAL, TEN, and PIT--pretty good company.
Does on of those columns help us see the strength of schedule?
Opp GWP
Do you know if penalty rates are generally consistent throughout the season?
@Anonymous - According to this post (http://www.advancednflstats.com/2010/11/predictivity.html) even and odd weeks correlate at 0.39. So pretty well.
The problem with these rankings is they don't take into account the fact that Romo is a choker.
Why is the league average offensive run success rate 41%, but the average defensive run success rate is 57%? Shouldn't they be the same?
Disregard the previous comment. They shouldn't be the same, they should add up to 100. Sorry about that. It's early and I haven't had my coffee yet.
I know you factor in home field advantage when calculating game probabilities, but do you factor that in for the efficiency rankings?
I can't recall reading if there is a home/road efficiency split, but it seems possible that since teams win at home more they might have better efficiency stats at home. This factor would be more significant early in the season when playing three road games to one home represents 75% of your total games played.
Another way this might affect things is Opp GWP. Two teams could have played the same four opponents and have the same Opp GWP, but if one team played three of them on the road while the other played them all at home, you would expect the latter to have more wins. Again I'm not sure if it would change the efficiency stats, but it might help explain the difference between GWP and record.
James-Good point, but no. The model doesn't consider to-date home/away splits. HFA does affect efficiency stats in the same way it affects other things. However, once we get past the first few weeks of the season the effect is really tiny, as teams usually have at most 1, possibly 2, surplus home or away games.
If you like, you can mentally subtract or add a point of GWP to the teams with a home road imbalance so far.
A question, and I should have asked this when the notion of run success % as part of the rankings system was first introduced. Q: Have you compared net yards per attempt with passing success %, and of the two, which is better suited to a logistics model?
Thanks for any reply.
David Myers, of Code and Football.
Brian, have you considered using data from past seasons in a continuous time series rather than using averages from the current season alone? Whilst the teams do change over the off-season, I think it is incredibly unlikely that there is no useful data to carry forwards. A time series that reacts more quickly to continued unexplained performances would do away with any fear you might have of teams that have changed a lot over the off-season.
I only ask because in my work using points and yardage analysis rather than efficiency stats, I have found that throwing away the data from the past season has a negative effect on the quality of predictions.
ditto what tom said from my experience
I'll have to take a look at that. I haven't done time series stuff before, so I'll have to take some time to get more familiar. I know it's not terribly different from what I've been doing.
I think that treating each parameter of the model as a random variable within its own elo-style model (obviously requiring a suitable prior distribution, though the logistic distribution usually is suitable) may be most effective. As such you get the time-series element and the strength-of-schedule built in, so no adjustments need be made for who has played who.
Obviously you'd have to generate the data for each game again and rerun the logistic regression, but I think the results may be very good.
Curses...I use your percentages to pick straight up winners in my work pool...and I have a bad feeling w/ home field advantage the model is going to say Denver over San Diego. Denver is awful, right? Help me feel better about this. What is San Diego doing poorly that will give Denver the chance to win.
any method that has the Cowboys at the top over the Packers and several of the others needs reworked.
Yes. As does your capitalization and grammar education.
I kid, but have you considered that the Packers defense has been pretty poor so far this season, as in 7.7 net YPA allowed? (4th worst)
"Early in the season, this model will heavily weigh "strength of schedule," so there you go. "
Why would you heavily weight strength of schedule early in the season when strength of schedule is only 1/4 known. That is giving too much weight to an unknown
It considers to-date strength of schedule, which is a known.
It weighs it equally throughout the season. It's just that teams tend to have widely varying average strength of opponent early in the season compared to later, when it tends to even out. Therefore, the adjustments tend to be larger early in the season.
Hope that helps.
On the SoS angle, I assume you perform an iteration until stability is found? What I am wondering is whether you end up with closed loops, and a generally incomplete network this early in the season - another possible reason for using past-season data.
The only problem I have is that last season's games are still relevant. For example; the Falcons may have performed terribly this season, but you always want to use as much data as you can and last season's results are still good data (if less relevant)
Winning teams moving down and losing teams moving up. That is a bad outcome for any system of ranking teams because the most important factor in ranking teams should be wins and losses. After all it is why they play the games.
Are you a Cowboys fan?
Anonymous immediately above-I used to think like you. But it turns out exactly the opposite of what you said is true.
Wins and Losses are highly random. Relative team ability is only one part of the equation that determines which team wins.
But randomness, by definition, is not predictable, so the idea here is to throw out the past randomness and focus on what is true repeatable skill.