Data scientist, physicist, and fantasy football champion

Introducing TruScor

I’m trying to get a little more accuracy out of these models, but I keep running into the problem of TDs. They’re worth so much and they’re so random. Or are they…

Some of the TDs come from special teams, some from fumbles, and some from interceptions. My idea is this: if we plot TDs vs the number of fumbles and interceptions, the intercept will be the special teams TDs, and I can get how many TDs we would expect per fumble and interception. In theory there’s a perfect data set out there that has where each TD came from, but I don’t have it.

 

For example, here are the number of TDs a defense scored plotted against the number of fumbles. I was going to do a 3D plot with TDs vs. fumbles and interceptions, it wasn’t clear unless you could rotate it. It looks like there’s a positive relationship between fumbles and TDs, but don’t trust that number (0.5 TDs per 4 fumbles) since it doesn’t take into account interceptions. Do do that, let’s run a linear model:

## 
## Call:
## lm(formula = Defensetd.number ~ Defensefr.number + Defenseint.number, 
## data = data_Def)
## 
## Coefficients:
## (Intercept) Defensefr.numberDefenseint.number
## 0.013700.092240.14995

Now we’re cooking. It looks like special teams score a TD about once every 73 games (1/73 ~ 0.0137), they score a TD about 1 in every 11 fumbles (1/11 ~ 0.09224) and one of every 6.7 interceptions is a pick-six (1/6.669 ~ 0.14995). The numbers for fumbles and interceptions feel about right, but the special teams TDs feel low to me. Let’s try it out anyway.

So what am I going to do with this? Since TDs are a little more random, I’m going to create a fantasy score that doesn’t count TDs, but gives a slightly higher score for fumbles and interceptions. Normally they’re worth 2 points each, but now fumbles will be worth (2 + 6 * 0.09224) there’s a 0.09224 probability that that fumble would have led to a TD. Similarly, interceptions are now worth (2 + 6 * 0.14995). I call this TruScor! (trademark pending, as soon as I figure out how to trademark things or what trademarking even is). Sacks are still 1 point, shutting a team out is still worth 10, etc.

I’m going to try to fit models A, B, and C from week 11 using TruScor (trademark pending) instead of the real fantasy score. That means that teams that got more fumbles and interceptions get more points than teams that only got 1 fumble but turned it into a TD. It’s not perfect, but I’m trying to bring a little order to a complicated world.

In Week 11, here would have been each team's TruScor assuming that TDs aren't worth anything and fumbles and interceptions are worth slightly more:

Rank

True.Order

True.Score

TruScor

1

PIT

22

17.4268

2

MIN

20

9.752

3

DET

19

9.3028

4

BUF

9

10.752

5

CAR

8

9.4268

6

JAC

7

8.1016

7

OAK

7

8.4268

8

NYG

7

7.876

9

SEA

7

8.752

10

LA

7

7.876

11

MIA

7

7.5508

12

WAS

6

7.4268

13

NE

6

6

14

IND

6

6

15

TB

5

6.4268

It looks like Pittsburgh earned those TDs with a bunch of interceptions and fumbles while MIN and DET landed on the luckier side, getting a TD on fewer turnovers. SEA and BUF should have been a little higher than they were if life worked on TruScor. Would I have predicted these scores, though?

Rank

Model.A

Model.A.TruScor

Model.B

Model.B.TruScor

Model.C

Model.C.TruScor

True.Order

True.Score

1

MIN

MIN

MIN

KC

KC

PIT

PIT

22

2

KC

KC

KC

MIN

MIN

KC

MIN

20

3

MIA

WAS

WAS

WAS

IND

MIN

DET

19

4

WAS

SEA

MIA

NO

DET

DET

BUF

9

5

SEA

LA

CAR

OAK

PIT

SEA

CAR

8

6

CAR

MIA

IND

CAR

SEA

OAK

JAC

7

7

LA

ARI

NO

LA

LA

ARI

OAK

7

8

IND

OAK

LA

MIA

PHI

LA

NYG

7

9

PHI

NO

BUF

SEA

ARI

NE

SEA

7

10

ARI

NE

PHI

PIT

WAS

IND

LA

7

11

NO

PIT

SEA

NE

NYG

WAS

MIA

7

12

DAL

CAR

TEN

DET

MIA

CIN

WAS

6

13

TB

TB

DAL

BUF

OAK

NO

NE

6

14

BUF

CIN

NE

CIN

NE

PHI

IND

6

15

TEN

DAL

TB

DAL

NO

CAR

TB

5

So using TruScor would have predicted PIT higher in all 3 models, and in model C it would have even predicted PIT in 1st place! Model C, you are really earning my trust this week. What about the accuracy score? How accurate were these?

Model.A

Model.A.TruScor

Model.B

Model.B.TruScor

Model.C

Model.C.TruScor

52

62

46

73

65

43

Using TruScor (trademark pending, don’t you dare touch it!) made models A and B a little worse and model C better. In fact, Model C with TruScor technology (trademark pending) is the best in week 11. Let’s get a quick week 12 predictions while we’re here.

I have Miami this week. Well… crap. NYG and BUF I would totally believe. PHI is up against GB this week, so I doubt the interceptions will be flowing, but they do a good job of forcing turnovers. Oakland is facing Carolina, and they have given up a lot of points, and not all of them from TDs. Time to stop. It’s Thanksgiving day at 1:00, so before these games start I need to see if anyone is available and how much I trust myself.

Should I switch, or can I accept mediocrity this week? 

New kicker models - GLMs

Data acquisition