Data scientist, physicist, and fantasy football champion

Week 4 DEF Predictions

All the models did poorly last week because the results were so weird. NYJ, NO, and JAC all ended up in that top 10 and TB, SEA, BAL, DEN, ARI, and CAR were all in the bottom 10. That was unexpected. Man, I hate modeling these things some weeks.

This week I’m going to follow models A and C closely. They were the top models from last year and now that I finally have enough data to properly run Model A I’m interested in seeing how it does.

ModelTypeScoreOppScoreOppHomeAwayDataYearsPointSystem
ALMXXXX2017IOUH/Standard
CLMXXXX2015-2017IOUH/Standard
FLMXXXX2015-2017TruScor
GGibbs SamplerXXXX2015-2017TruScor

 

Model A:

.Model A-1.png


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 2017
Standard points
Did well in 2016, but didn’t stand out


Model C:

.Model C-1.png


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 2015-2017
Standard scoring
Did well in 2016, but didn’t stand out


Model F:

.Model F-1.png


Model F 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 F Notes:

Uses data from 2015-2017
Truscor ignores defense TDs but weights interceptions and fumbles to try to account for it
I really liked this one in 2016. It was marginally better than A and C


Model G:

.Model G-1.png


Model G 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 G Notes:

Uses data from 2015-2017
Truscor ignores defense TDs but weights interceptions and fumbles to try to account for it
Uses Bayesian statistics (Gibbs sampler) to create a model
Would have done very well in 2016 based on simulation


Week 4 model summary:

RankACFGH
1TENSEASEASEABAL
2BALPHICINCINTEN
3JACJACJACJACJAC
4CHIKCPHIARIKC
5DETARIARINEPIT
6SEATBOAKPHIWAS
7GBCINNETBSEA
8PITNETBOAKGB
9KCTENGBGBCHI
10LACHOUTENDALNYG
11WASGBPITPITARI
12HOUDALKCTENNO
13MINOAKSFKCDAL
14DALATLDALATLNE
15PHIPITLACLACHOU
16NONYGNYGSFDET
17LARDENATLNYGTB
18TBSFHOUDETOAK
19NELACCLEHOUPHI
20ARIDETDETNOLAC
21INDCLEBALCLEMIN
22CARLARNYJDENCIN
23DENMINNONYJIND
24NYGCARDENBALDEN
25OAKNYJCARMINSF
26CINNOLARLARCLE
27SFWASMINCARNYJ
28NYJBALWASWASATL
29BUFCHICHICHICAR
30ATLINDINDINDLAR
31CLEMIAMIAMIABUF
32MIABUFBUFBUFMIA


Conclusions

[NOTE: As of writing, the odds for the DET/MIN game are still OFF, so take those predictions lightly. That QB situation is keeping Vegas from setting the odds.]

Depending on which Jacoby Brissett shows up, Seattle could be in for a big game. If they get the one from week 2, Seattle is the defense to play. If they get the one from Week 3 they can expect an OK week, but nothing special. We’ll see!

Model C thinks this week is a who’s-who of good defenses. If you have a top DST, play them (KC, ARI, JAC, SEA, HOU). rounding out the top 12 for model C we have a few streaming options in TB, TEN, and GB.

My models disagree the most on CIN this week. They’re up against CLE (usually a smart play), Model A really doesn’t like them for a weird reason, so bear with me here. CLE gave up a lot of points to BAL, PIT, and IND. That is not up for debate. HOWEVER, the model takes into account how well these defenses did against CLE compared to how well they did against other teams. And while BAL did well against CLE, they crushed CIN. PIT did well against CLE, but they also did well against MIN. IND did slightly better against ARI than CLE. THEREFORE, my models are arguing that CLE isn’t an offense that gives up a lot of points, they’re just a team that has gone against good defenses. And CIN hasn’t done much to distinguish itself this year (good points but against bad teams). THEREFORE, my models argue that CIN isn’t a good defense, they’ve just been up against easy targets. All this adds up to Model A not really liking CIN.

Model A is pretty tough to defend on only 3 weeks of data. CHI against GB is a tough sell, despite GB giving up a lot of points so far this year. But many of those points came from TDs. If only we had a way to account for this. Wait, we do! Truscor! Here’s a secret fifth model for the week, Model H. It’s Model A, but using truscor scoring, and it’s the model I’d like to go with this week:

.Model H-1.png

This model passes all of my eye tests (BAL, PIT, and KC all pretty high, CIN and CLE both pretty low). It likes BAL, but pulls DET and IND back down to Earth a little. It still thinks defenses will score well against GB, HOU, IND, and PHI, but less than before. It’ll be a little controversial, but this is the model I like this week. It’s not a good idea to search for models until you find one that looks good to you (shut up, Ben), but that’s exactly what I’m going this week.

But the question remains: do I have the conviction to play TEN as my defense this week? I have PHI on my team from last week, and if I believe Model C (the current champion) I should keep them. But none of my models are distinguishing themselves, and Model H is fun and new. It’s also volatile; 3 data points is really not a good sample set.

I think I’ll wimp out and try for Jacksonville since they’re #3 for all my models this week, but if I can’t get them, the rest of those in Model H are unpalatable. DET has been dominant, but MIN is as stingy as they come (literally, they’re the only opponent with a negative modifier in my model). CHI has been good and GB gives up the most points (relative to the opposing defense), but the estimated score (26.5) is intimidating. I’m talking myself out of this now. Time to stop.

Stick with Model C if you’re the kind of person who read the warning label on Q-tips and stopped using them in your ears. Give Model H a whirl if you eat meat from street vendors and prefer the mystery dum-dum pop.

Week 4 QB Predictions

Week 4 K Predictions