Data scientist, physicist, and fantasy football champion

Week 11 DEF Predictions

Last week wasn’t a total mess, so let’s try this again!

Model A:

Model A parameters:

  • The team that’s playing
  • Who their opponent is
  • Their predicted score
  • Their opponent’s predicted score
  • Whether the team is home or away

Model A Notes:

  • Uses only data from 2016
  • Best model from weeks 8 and 9, 3rd best in week 10

Model B:

Model B parameters:

  • The team that’s playing
  • Who their opponent is
  • Their opponent’s predicted score

Model B Notes:

  • Uses only data from 2016
  • Comparable results to Model A while reducing overfitting

Model C:

Model C parameters:

  • The team that’s playing
  • Who their opponent is
  • Their predicted score
  • Their opponent’s predicted score
  • Whether the team is home or away

Model C Notes:

  • Uses data from 2015 and 2016
  • OK performance weeks 8 and 9, best model in week 10

Model D:

Model D parameters:

  • The team that’s playing
  • Who their opponent is
  • Their predicted score
  • Their opponent’s predicted score
  • Whether the team is home or away

Model D Notes:

  • New model this week
  • Uses only data from weeks 5 through 10 (inclusive) for 2016
  • What have you done for me lately?

Model E:

Model E parameters:

  • The team that’s playing
  • Who their opponent is
  • Their predicted score
  • Their opponent’s predicted score
  • Whether the team is home or away
  • NEW: “Team:week” term (does the team have momentum?)
  • NEW: “Opponent:week” term (does the opponent have momentum?)
  • NEW: “week” term (basically a bookkeeping term if I want to use the other two)

Model E Notes:

  • New model this week
  • Uses data for all of 2016 (through week 10)
  • includes week + Team:week and Opponent:week terms for each team

Week 11 model summary

Rank

Model.A

Model.B

Model.C

Model.D

Model.E

1

MIN

MIN

KC

MIA

WAS

2

KC

KC

MIN

SEA

MIN

3

MIA

WAS

IND

KC

NE

4

WAS

MIA

DET

MIN

TB

5

SEA

CAR

PIT

WAS

CAR

6

CAR

IND

SEA

PHI

GB

7

LA

NO

LA

TB

CIN

8

IND

LA

PHI

ARI

KC

9

PHI

BUF

ARI

PIT

NYG

10

ARI

PHI

WAS

IND

DET

11

NO

SEA

NYG

GB

DAL

12

DAL

TEN

MIA

CAR

JAC

13

TB

DAL

OAK

BAL

IND

14

BUF

NE

NE

DAL

BUF

15

TEN

TB

NO

OAK

OAK

16

NE

OAK

DAL

BUF

NO

17

CIN

ARI

CAR

NYG

PIT

18

OAK

CIN

TB

DET

BAL

19

GB

DET

TEN

LA

CHI

20

CHI

PIT

CIN

CIN

PHI

21

NYG

GB

BUF

NE

TEN

22

DET

NYG

GB

NO

SEA

23

PIT

CHI

BAL

CHI

HOU

24

BAL

BAL

HOU

SF

MIA

25

SF

SF

CHI

TEN

CLE

26

HOU

HOU

CLE

HOU

ARI

27

CLE

CLE

SF

CLE

SF

Like last week, Models A and B give similar results. Model C is a bit different and makes a few bold predictions such as IND and DET toward the top and PIT fairly high as well, though that’s more about facing CLE than anything I’d guess. Note also that the prediction range is decreased in this model (max and min scores are pulled toward the center a little). Despite being the best model last week I still don’t trust this one as much. I think it takes into account past performance too much and gives a much more conservative estimate of the DEF scores. I imagine that if I had a decade of data it would just cluster everyone around 5 or 6 points.

In fact, I distrust the past so much I’ve created 2 NEW models this week: Model D and Model E. Model D uses only the last 6 weeks of data to predict how well a defense will do. Model E uses Team by week and Opponent by week terms that are distinct for each team to see if they have “momentum”. It understandably puts TB very high (24 points last week) but confusingly drops MIA pretty low (30 points in the last 2 weeks and LA has been giving up decent points to team defenses lately). This is a crazy model. If nothing else it should provide a nice contrast for the other models. All of my models have done very well so far so it’s not clear to me how I should proceed to improve them. If I have any flaws it’s that I’m too good at this.

That and kidney stones. I’d call those a flaw too.

At the time of writing this I have LA. They were good last week (10 points), but facing MIA this week who haven’t given up an INT in a while. Models D and E recognize this, but Model E might overcompensate a little, predicting LA for dead last. That’s probably a bit of a stretch, but I’ll still pick up someone else this week. Hmm… It looks like TB is available…

Conclusions / Trash talk

When compared to D and E, models A, B, and C are all actually very similar. I’ll probably find a team that shows up toward the top of those lists. The good news is that it doesn’t really matter who I pick up this week since I’m going up against the worst player in my league. Her team has the same record as mine, she’s scored the most points so far this year (by more than 100 points), and I’ve already lost to her once this year, but I’m not worried. This is a case where the numbers hide the truth rather than reveal it. Numbers are liars sometimes, and I don’t trust the past.

Good luck to everyone (but MF) this week!

Week 11 K Predictions

Week 10 K - Evaluation