FanPost

World-class irrelevant statistics: Chiefs vs. Broncos

Justin Edmonds

From the FanPosts -Joel

Hello AP members and the inevitable cross-traffickers from MHR,

I put together this FanPost for a couple of reasons:

  • First, I have seen several posts like this on both sites over the past few weeks which all suffered from similar mathematical errors, so I thought I would gen up my own
  • Second, after my attempt to compose an inspirational video failed miserably, I figured I had better stick to the stats
  • Third, this is Chiefs vs. Broncos! It's the "Irresistible Force" meets the "Immovable Object!" With a combined record of 17-1! In a game with obvious playoff implications! Why would we not over-analyze it?

Now the Disclaimer:

This post can tell you almost next to nothing. Continue reading for entertainment value only. What I can assure you is that my math is the best I can make it (which may or may not be "world class" but, hey, I get to write the title, right?) On the spectrum of Lies, Damn Lies, and Statistics, these numbers fall squarely in the "statistics" category. What's the great thing about statistics? When done properly, they tell you how accurate they are. In fact, that's whole point of proper econometrics - determining relevancy. What these numbers tell me is that their predictive power is just a little better than your chances of winning at black jack. Of course, I wouldn't have known that without running through the exercise. Bottom line: all of these numbers are real, yet do little to actually predict anything. But why should that stop us?

So here it is in all its glory: your completely meaningless, yet highly analyzed statistical prediction for the game!

The first thing I had to assess was which stats to look at. People have argued over Defensive and Offensive Yardage vs. Split Stats vs. Points Allowed. Here's what I know, the winner is determined by the score board, thus the ultimate goal is to predict the score. But how can we tell if yards are important or not? We look at correlation:

Yardage

KC’S OPP

Pass Yards

Rush Yards

KCC Yards Allowed

PTS vs. KCC

DEN’S OPP

Pass Yards Against

Rush Yards Against

Total Yards

DEN PTS SCORED

JAX

157

71

228

2

BAL

462

65

527

49

DAL

298

37

335

16

NYG

307

107

414

41

PHI

201

264

465

16

OAK

374

164

538

37

NYG

217

98

315

7

PHI

337

141

478

52

TEN

247

105

352

17

DAL

414

103

517

51

OAK

216

121

337

7

JAX

295

112

407

35

HOU

271

73

344

16

IND

386

64

450

33

CLE

293

57

350

17

WAS

354

107

461

45

BUF

229

241

470

13

SDC

313

84

397

28

CORR:

0.77

0.11

0.59

CORR:

0.48

0.17

0.60

This spreadsheet shows the correlation factor between each yardage stat and points scored on the two sides of the ball we care about: The Chiefs Defense and the Denver Offense. These correlation values actually increased by quite a bit after the DEN vs. SDC game Sunday Night, but they are still fairly weak. A "strong" correlation is 0.8 to 0.99. Moderate correlation is around 0.6 to 0.8. Anything below 0.6 is pretty weak for predicting results (In fact, most true statistical analysis won't report something as statistically significant until P-Values reach 95%).

This tells me that I could go through the whole exercise of predicting the yardage of next Sunday's game, but even if we got the yardage perfect, we would be no better off predicting the score from that yardage than we would flipping a coin. But can correlation tell us everything? No, and the 0.6 correlation factor does show some relevance for the stat, so lets look at them side by side. Here are charts of the Chiefs Yards Allowed and Points Allowed side-by-side, and the Broncos Yards Gained and Points Scored side-by-side:

Cyo4_medium

via img203.imageshack.us

What do we notice? First, we notice that although the individual data points don't correlate very well, the trend lines look very similar. We also see that although the trend lines are both regressing to the norm, that regression is pretty low for both teams. If you look at the "R^2" values that are printed on each graph, they tell you what percentage of the graph is "explained" by the trend line.

For instance, in terms of KCC Points Allowed, the regression to the norm accounts for only 23.21% of the Chiefs trend from week to week. Roughly, it's like saying they are regressing a bit, but 77% of their game is not regressing. What does this mean? It means that again, looking at yardage won't tell us anything additional to what looking at points is going to tell us. It also tells us that we should expect the Chiefs to give up more points than usual, and the Broncos should score less than usual, but not by big margins, both teams are still really good - not exactly ground breaking stuff there, huh? So, enough with yardage.

Win Percent Added

A post at MHR claimed that the Broncos have a better chance of winning the game because, on average, they have spent more time at 100% WPA then the Chiefs. What that tells me is that when the Broncos lock up a game, they do it earlier than the Chiefs. Again, is that news given our contrasting styles of play? Not to me. So instead, I want to look at how much time during each game do the Chiefs have the odds in their favor, and compare that to the Denver numbers.

Therefore, I went to pro-football-reference.com and looked at the box scores for every Chiefs and Broncos games. I then counted WPA by quarter and broke it into three categories:

  1. Quarters where KCC/DEN had a "winning" probability, meaning that the WPA graph was above 50% for the entire quarter
  2. Quarters that were "in-doubt," which are any quarters where the WPA line hovers right around the 50% mark or crosses from one team to the other during the quarter
  3. Quarters in which KCC/DEN were "losing," in that their WPA was never above 50% during the entire quarter

Here's what I found:

# of Qtrs WPA:

CHIEFS

BRONCOS

OPPONENT

Winning

In Doubt

Losing

OPPONENT

Winning

In Doubt

Losing

vs. JAC

4

BAL

2

2

vs. DAL

2

2

NYG

2

2

vs. PHI

3

1

OAK

4

vs. NYG

4

PHI

4

vs. TEN

2

2

DAL

1

3

vs. OAK

3

1

JAX

4

vs. HOU

3

1

IND

1

1

2

vs. CLE

4

WAS

3

1

vs. BUF

1

2

1

SDC

4

TOTAL:

26

9

1

TOTAL:

25

9

2

(You can look these numbers up yourself using this link here, and clicking on each game's "boxscore" to find the WPA graph)

Yes, the Chiefs have sustained a winning probability in exactly the same number of quarters as the Broncos, and have only had 1 "losing" quarter of football which came in the Buffalo game. Denver had 2 against Indy, including the most important quarter, the fourth. What does this tell us? It says that even though the Chiefs play in a lot of close games, they have rarely been "out" of a game -- just like the Broncos. It also says that because the two teams are so alike, WPA won't help us differentiate them, or affect the score much. So, onto the points.

The Score

Now, several posts have tried to calculate the upcoming scores by looking at the percentage of points that each team scores above or below other teams, averaging those percentages, and then calculating points against each team. There is a problem with using percentages to do this, however. What's the problem? I'm glad you asked.

The problem is that you can't average the percentages because they are different units (now, we could use the Harmonic Means to average unlike numbers, but that would favor the large numbers, which in football are the statistical outliers--meaning that rare games, like our 2 point game against JAX would skew our numbers even more than using arithmetic means, but I digress...). But Bull, you ask, point are points...how are these different units?

"Points against Jacksonville" are not the same thing as "Points against Philly." Why? Mainly because they play different teams - it's the old, "who have you played?" dilemma. If every team in the NFL played every other team exactly once in every season, we wouldn't have that unit problem. Even so, we still couldn't use "percentages above or below mean" to define a team's overall relative strength because those units vary by team also. Seven extra points against JAX may not be the same as 7 extra points against PHI, even if their average scores are the same.

Wait, what? Why not? Look at it this way: Let's imagine that the points scored against JAX in all of their games has been: 20, 0, 20, 0, 15, 5, 15 and 5. Philly has allowed scores of 10, 10, 10, 10, 11, 9, 11 and 9. Both teams have an average of 10 points allowed. Now, the Chiefs score 17 against JAX and 17 against PHI. Which one is more "impressive?" Which one tells us more about the Chiefs?

Well, the extra 7 points against the Eagles is a more "unusual" feat then the ones against the Jags. Luckily, we can capture that "unusualness?" (yes, that's a word... if you can say it, it's a word) and fix our "units" issue with one thing: standard deviations. Stnd Devs give us a normalized error distribution around a determined mean. No matter what group of numbers you have, if you find the standard deviation, around 68% of your values will fall within one Stnd Dev of the average, about 95% of your values will fall within two Stnd Devs, and 99% of your values will fall within three standard deviations - regardless of the units or size of the sample. Presto, our problems are solved, it just takes more work.

To find our two teams' performance using standard deviations, we have to look at the points scored for and against each of our opponents against each of their opponents. We then find each team's mean and standard deviation for points scored and allowed, using all of the games except for the ones against KCC and DEN. If we used our own games in our calculation, it would induce an endogeneity problem, which basically means that we would be using our own performance to measure our own performance. We don't really want to do that. Once we have the Mean and Stnd Dev for each team, we look at the Chiefs' and Broncos' performance against them, calculating how many standard deviations above or below average we scored. We can then average our standard deviations across our 9 games to determine how well our Offenses and Defenses perform as compared to the mean. All of that is here in this table (scroll all the way to the right for the final values):

Opponent:

JAX

DAL

PHI

NYG

TEN

OAK

HOU

CLE

BUF

BAL

COLTS

WASH

SDC

Averages

Stnd Devs

Week

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

1

36

31

33

27

31

36

16

9

17

21

31

28

10

23

21

23

21

17

27

33

28

31

2

9

19

30

33

24

30

19

9

30

24

6

14

24

23

14

6

20

24

20

38

33

30

3

17

45

31

7

0

38

20

17

9

30

31

27

20

27

30

9

27

7

20

27

17

20

4

3

37

21

30

38

13

14

24

20

23

17

6

23

20

20

23

37

3

24

14

30

21

5

20

34

36

21

21

36

27

17

3

34

37

24

24

37

26

23

34

28

17

27

6

31

16

31

20

21

27

13

20

13

38

17

31

24

27

17

19

9

19

16

31

19

9

7

6

24

17

3

3

17

23

7

17

31

13

31

23

21

16

19

45

41

24

6

8

10

42

30

31

7

15

15

7

21

18

17

35

9

27

23

49

20

28

21

20

49

24

27

24

18

18

24

27

24

30

24

24

30

10

29

27

27

13

24

20

27

29

20

24

24

27

10

23

20

17

8

38

27

34

20

28

AVERAGE

13.43

32.57

27.57

20.14

27.00

20.75

19.29

24.43

22.88

21.25

19.71

23.14

19.25

28.88

19.38

21.75

20.67

26.22

20.13

17.50

22.88

20.00

26.13

30.25

23.56

22.44

STD DEV

8.40

8.89

6.04

10.86

14.15

6.10

9.02

12.45

7.59

7.66

3.69

11.58

9.40

4.70

9.91

8.12

4.37

5.69

5.01

6.24

9.89

10.54

8.33

8.00

5.48

8.81

Against CHIEFS

2

28

16

17

16

26

7

31

17

26

7

24

16

17

17

23

13

23

+/- DEF STND DEV

1.36

1.91

0.78

1.36

0.77

3.44

0.35

0.24

1.75

1.33

0.93

+/- OFF STND DEV

-0.51

-0.29

0.86

0.53

0.62

0.07

-2.53

0.15

-0.57

-0.18

0.95

Against DEN

19

35

48

51

20

52

23

41

21

37

27

49

39

33

21

45

24

21.75

+/- DEF STND DEV

-0.66

-3.38

0.49

-0.41

-0.35

-1.37

-1.63

0.62

-0.08

-0.75

1.17

+/- OFF STND DEV

0.27

2.84

5.12

1.33

1.20

5.04

1.23

1.84

-0.08

2.09

1.78

What does this tell us? First, it tells us that Denver's Offense is "better" than our Defense is "good." Our defensive Stnd Dev is 1.33, which essentially means we've performed better than around 80% of others. Denver's Offensive Stnd Dev of 2.09 means they outplay around 95% of others. But, we also see that Denver's defense is "worse" than our offense is "bad." So, which matters more? Well, we don't know that until we know what our own average and standard deviations are. So, repeating the process above for Denver and KC, we get this:

CHIEFS

DONCOS

Week

PTS SCORED

PTS ALLOWED

PTS SCORED

PTS ALLOWED

1

28

2

49

27

2

17

16

41

23

3

26

16

37

21

4

31

7

52

20

5

26

17

51

48

6

24

7

35

19

7

17

16

33

39

8

23

17

45

21

9

23

13

10

28

20

AVERAGE

23.89

12.33

41.22

26.44

STD DEV

4.38

5.25

8.07

9.62

Now, if we multiply our Offense Average Stnd Dev against their Defense numbers, we get a predicted score of 24.67. If we look at their Defensive Average Stnd Dev against our Offense numbers, we get a score of 27.19. These numbers are pretty close, and thus we get a pretty stable prediction of the Chief's score on Sunday: 25-27 points.

A Pause in the Action to Look at Sacks

Before we look at the other side of the ball (the one with the intriguing storyline, Manning vs. The Red Crush), I wanted to take a quick look at sacks. One key to this game is going to be getting pressure on Manning--but will it happen? We lead the league in sacks, the Broncos are very good at protecting Manning. Again, we can't compare the two directly - our sacks have come against different offenses than Denver's, and their protection has been against different front 7's then KC. I therefore normalized the sack numbers across the league using total number of sacks allowed and sack percentage.

I took each of the teams we have faced, and normalized them on a scale from 0 to 1, where 0 means you have the best pass protection and 1 means you have the worst, based on allowed sack percentage. I then used that number to compute a "sack ratio," which means that a sack against the worst team equals "1 sack," but a sack against the best team would be worth "3.2 sacks." Why 1 and 3.2? It's arbitrary, but the best team in the league had given up 10 sacks through 9 weeks and the worst had given up 32, so the 1 to 3.2 ratio matched the league extremes. With this number, we can compute the "adjusted sacks" the Chiefs have gotten against each team, and find our average number of adjusted sacks: 7.65. Comparing that to Denver's "3.2 adjusted sacks to sack" ratio, we predict the Chiefs will sack Manning 2.4 times. Here's the work:

BY OPPONENT

SACK

FO ASR RANK

FO ASR

Normalize

Ratio

Adjusted Sacks

vs. JAC

6.0

22

7.8

0.4872

2.1282

12.7692

vs. DAL

3.0

7

6.0

0.2564

2.6359

7.9077

vs. PHI

6.0

26

8.8

0.6154

1.8462

11.0769

vs. NYG

3.0

20

7.6

0.4615

2.1846

6.5538

vs. TEN

3.0

18

7.4

0.4359

2.2410

6.7231

vs. OAK

9.0

32

11.8

1.0000

1.0000

9.0000

vs. HOU

5.0

9

6.3

0.2949

2.5513

12.7564

vs. CLE

1.0

23

7.9

0.5000

2.1000

2.1000

vs. BUF

0.0

17

7.3

0.4231

2.2692

0.0000

vs. DEN

2

4.0

0.0000

3.2000

2.3919

That's not great, and a bit below our average, but our trend suggests that our pass rush has been slowing down, as you see here:

9o3x_medium

via img833.imageshack.us

Why did we take this little detour about sacks? Because our prediction for the Denver score is not as stable as our prediction for the Chief's score was. If we look at it from our Defenses perspective, we give up an average of 12.33 points a game with a Stnd Deviation around 5. Since Denver scores about 2 Stnd Devs above average, this suggests they will score 23.31 points against us.

But, their offense scores an average of 41.22 with a Stnd Dev of 8, and since our defense normally holds teams to 1.33 Stnd Devs below, this predicts a Denver score of 30.49. That 23 to 30 point range is a much larger spread than the 25 to 27 for KC. Thus, math says: Denver: 23-30 vs. Kansas City: 25-27

I figured up the predicted sacks to tell me which ends I should favor. Although 2.4 sacks is below our average, it is well above Denver's average of 1.4 per game (but not quite the 4 sacks Indy got in Denver's only loss). If it were four sacks, I'd go with the KC number - if it were 1.4 sacks, I'd go with the Denver number. Since it splits the difference - I'll split it, thus my prediction for Denver is 26.90 points. If we compare to our earlier prediction of 25-27, we can see it's anyone's game, but to be fair, I took our average as well. This gives us our:

Final Score

Thus, after all this, I mathematically predict:

DENVER - 27, KANSAS CITY - 26

Now, once again, the certainty of this prediction is so far out of statistical significance that you have a better chance of falling in a manhole tomorrow than you do betting this line, but hey, it was fun. What does my gut tell me? Honestly, I think this season could wind up a lot like 2010, where the score in Denver was a 49-29 loss, but the score in Arrowhead was a 10-6 win. Two teams playing in two games that took on completely different characters. Either way, I think this Sunday is going to be a close one. Either way, we will all know for sure on Sunday.

Hope you enjoyed this.

Go Chiefs!

Edit: Author's Post-Script About the "Back-Up QB" Argument

In response to a few comments in other threads, I have decided to add a quick blurb about the inevitable critique of these stats that will claim the KC numbers are skewed because we play against back-up quarterbacks. While they are right, the Chiefs did play back-ups, so did many of the other teams that set our baseline averages and standard deviations. Most of the back-ups have played at least three games, making up a third of our sample size which is more than enough to "unskew" any numbers--with one exception: Buffalo. Tuel played only one game, which means the Buffalo numbers are less reliable than all of our others. We can also see that our defense's Stnd Dev against Buffalo was one of our highest at 1.75. So, in fairness, I deleted the Buffalo data as a statistical outlier to see the change in results. What happens? It changes our "final score" from 26.9-25.93 to 27.11-26.16. In other words, it makes no difference what-so-ever.

This is a FanPost and does not necessarily reflect the views of Arrowhead Pride's writers or editors. It does reflect the views of this particular fan though, which is as important as the views of Arrowhead Pride writers or editors.

X
Log In Sign Up

forgot?
Log In Sign Up

Please choose a new SB Nation username and password

As part of the new SB Nation launch, prior users will need to choose a permanent username, along with a new password.

Your username will be used to login to SB Nation going forward.

I already have a Vox Media account!

Verify Vox Media account

Please login to your Vox Media account. This account will be linked to your previously existing Eater account.

Please choose a new SB Nation username and password

As part of the new SB Nation launch, prior MT authors will need to choose a new username and password.

Your username will be used to login to SB Nation going forward.

Forgot password?

We'll email you a reset link.

If you signed up using a 3rd party account like Facebook or Twitter, please login with it instead.

Forgot password?

Try another email?

Almost done,

By becoming a registered user, you are also agreeing to our Terms and confirming that you have read our Privacy Policy.

Join Arrowhead Pride

You must be a member of Arrowhead Pride to participate.

We have our own Community Guidelines at Arrowhead Pride. You should read them.

Join Arrowhead Pride

You must be a member of Arrowhead Pride to participate.

We have our own Community Guidelines at Arrowhead Pride. You should read them.

Spinner.vc97ec6e

Authenticating

Great!

Choose an available username to complete sign up.

In order to provide our users with a better overall experience, we ask for more information from Facebook when using it to login so that we can learn more about our audience and provide you with the best possible experience. We do not store specific user data and the sharing of it is not required to login with Facebook.

tracking_pixel_9341_tracker