Read on Twitter
DEFENSES DON& #39;T MATTER FOR FANTASY FOOTBALL 101
Class is now in session.
thread 1/n https://twitter.com/rbkeeney/status/1309902391240732673">https://twitter.com/rbkeeney/...
Class is now in session.
thread 1/n https://twitter.com/rbkeeney/status/1309902391240732673">https://twitter.com/rbkeeney/...
Fantasy football is a game of skill,,, and luck.
a LOT of luck.
Injuries, weekly variation, draft-to-draft variation, and many more things all drive variance, or noise.
2/
a LOT of luck.
Injuries, weekly variation, draft-to-draft variation, and many more things all drive variance, or noise.
2/
Because there is so much noise, it& #39;s just as easy to be right for the wrong reasons as it is to be right for the right reasons.
And this drives our narrative.
3/
And this drives our narrative.
3/
As analysts (film or data), the foundations of our recommendations come from understanding the noise and variance:
(High Variance) Play > Game > Season > Career (Lower Variance)
The more data we have, the more confident we can be.
4/
(High Variance) Play > Game > Season > Career (Lower Variance)
The more data we have, the more confident we can be.
4/
For example, I have a model for WRs that looks at the top 200 ADP players. On a SEASONAL basis, I can predict a player& #39;s PPR points per game at a 0.8 rsq clip using their airyards, target, ADP and team information from the year prior.
5/
5/
However, that drops to < 0.3 rsq if when we look at each game individually.
Results on a per-game basis are very, very noisy.
6/
Results on a per-game basis are very, very noisy.
6/
Projecting weekly points is challenging, and very profitable if you get it right... so we have an incentive to look for any edge we can get.
Enter defensive data.
7/
Enter defensive data.
7/
However, there is danger lurking...
For our analysis go beyond entertaining and become useful, we have to understand the difference between what& #39;s descriptive... and what& #39;s predictive.
And this is the trap that defensive data lays.
8/
For our analysis go beyond entertaining and become useful, we have to understand the difference between what& #39;s descriptive... and what& #39;s predictive.
And this is the trap that defensive data lays.
8/
Defensive data can be VERY descriptive.
Here& #39;s my defensive model looking at the average expected fantasy points allowed to WRs and RBs allowed by a defense a that team& #39;s average defensive rating over the season.
Here& #39;s my defensive model looking at the average expected fantasy points allowed to WRs and RBs allowed by a defense a that team& #39;s average defensive rating over the season.
Now let& #39;s look at the models rating of team strength at the LAST game of the season. It& #39;s more descriptive than the mean rating over the same season.
(it& #39;s a type of elo model for you nerds)
(the chart is defensive xFP allowed to RBs by def/year)
9/
(it& #39;s a type of elo model for you nerds)
(the chart is defensive xFP allowed to RBs by def/year)
9/
But... what happens if we want to use this data to predict future data?
Well, it turns out that isn& #39;t a good idea.
(chart: defensive xFP allowed to RBs by def/year)
10/
Well, it turns out that isn& #39;t a good idea.
(chart: defensive xFP allowed to RBs by def/year)
10/
But wait.
The model is designed to best at predicting the "next" game, does it to that okay?
It& #39;s not great, but it can describe 3-5% of the variation in xFP.
Not much, but that& #39;s a small edge right?
11/
The model is designed to best at predicting the "next" game, does it to that okay?
It& #39;s not great, but it can describe 3-5% of the variation in xFP.
Not much, but that& #39;s a small edge right?
11/
Wrong.
What happened? Where did my 5% edge go?
12/
What happened? Where did my 5% edge go?
12/
It& #39;s dwarfed by offensive data... even though offensive data only reaches about 0.15-0.22 rsq accuracy on a per-game basis in my model!
13/
13/
But we& #39;re just getting started:
The offensive model is swallowed up by betting lines, but interestingly... the total stays at 0.15-0.22 rsq on a per-game basis.
That means my offensive model is about as accurate as betting information.
But, we& #39;re not done yet.
14/
The offensive model is swallowed up by betting lines, but interestingly... the total stays at 0.15-0.22 rsq on a per-game basis.
That means my offensive model is about as accurate as betting information.
But, we& #39;re not done yet.
14/
Enter the final player model.
Its baseline is ~0.3 rsq per game... not to bad considering it doesn& #39;t know things like injuries, snap %, depth charts, and slots in new players with a "flex" level expectation.
Here& #39;s the chart for every RB in every game since 2010:
15/
Its baseline is ~0.3 rsq per game... not to bad considering it doesn& #39;t know things like injuries, snap %, depth charts, and slots in new players with a "flex" level expectation.
Here& #39;s the chart for every RB in every game since 2010:
15/
We have one lever left to pull.
Does adding defensive, offensive, or betting data to individual player projections improve them?
16/
Does adding defensive, offensive, or betting data to individual player projections improve them?
16/
No.
The adj-rsq only raises by 0.05 if we include even the most meaningful data... and opposing defensive strength data adds zero value.
17/
The adj-rsq only raises by 0.05 if we include even the most meaningful data... and opposing defensive strength data adds zero value.
17/
What does that mean?
Well, if you are using "good" projections and you adjust them further by considering team defensive or offensive data (and even betting lines to an extent) you& #39;re likely making them worse!
You may be right, but it& #39;s that& #39;s just luck.
18/
Well, if you are using "good" projections and you adjust them further by considering team defensive or offensive data (and even betting lines to an extent) you& #39;re likely making them worse!
You may be right, but it& #39;s that& #39;s just luck.
18/
FIN.
Okay, there are a few caveats.
While I trust my data & analysis, by no means have I completed an exhaustive search on every defensive stats/metrics/models and their ability to predict fantasy points.
I& #39;m looking for ANY signal, if you find one, I& #39;d be thrilled.
fin/ +1
Okay, there are a few caveats.
While I trust my data & analysis, by no means have I completed an exhaustive search on every defensive stats/metrics/models and their ability to predict fantasy points.
I& #39;m looking for ANY signal, if you find one, I& #39;d be thrilled.
fin/ +1
FYI, it& #39;s also difficult predicting plays-per-game and passing-to-rush rates too... even *in* season, so careful with that data too.
Read more about it here:
https://twitter.com/rbkeeney/status/1274111436508315649?s=20
fin/">https://twitter.com/rbkeeney/... +2
Read more about it here:
https://twitter.com/rbkeeney/status/1274111436508315649?s=20
fin/">https://twitter.com/rbkeeney/... +2
Want more? Check out my @FF_Astronauts Analytics show where I break down the key trends going into week 3 and discuss how defensive stats can be a trap! https://youtu.be/Zg2SazJLxFc ">https://youtu.be/Zg2SazJLx...
