Yudkin's Defense Variability Article (1 Viewer)

Tuomi said:

Are your top/bottom 11 in order?

Click to expand...

wdcrob said:

Seems like no matter what small edge I think I have, the FBGs publish an article on it. Which is why we come here I guess.

Anyhow, loved Yudkin's analysis on Total Defense and thought I'd add a few things (now that my most important draft is finished) since I did more or less the same thing over the summer. My work hasn't been vetted and probably has a few holes (big believer in "good enough" when no lives or money is at stake), but my overall findings are so similar with the stuff he published I figured there's no obvious reason to think I screwed the pooch on the rest of it.

The major findings:

--As Yudkin says, takeaways, sacks and turnover margin are all pretty much random from year to year. There are a few slight exceptions, but they don't seem to matter much when predicting FF defenses.

--Yards/game, yards/play and points/game seem to be the only NFL-stats that predict fantasy points, but are very highly correlated with each other, so using more than one of them just confuses things. I forget why I settled on Pts/game, but there's probably some reason (see "good enough" above).

--Because roughly half of a Defense's performance in the previous year is due to more or less random events (mostly turnover margin), there's a very strong tendency for teams to regress to the mean the following year. i.e. because the really good and really bad teams had extreme turnover margins (good or bad) the previous year they are more likely than not to drift towards the middle again the next season. The Bears probably won't replicate last year's maulings (though they are still a very good defense) and the Patriots should be much improved FF-wise (because they are also a very good defense -- see below).

Combining the two ideas above, you get a model that explains the non-random half of a defense's upcoming fantasy performance as follows:

--Last Year's FF points scored plus 236

minus

--3 NFL Seasons Average Points Surrendered * 5.22 (t-stat -2.02)

(so a team averaging 20 points/game over the last three seasons would subtract ~104 points from 236)

minus

--FF points last year * .94 (t-stat -5.36 !!)

(this means teams that were poor the previous year lose fewer points than those that were good)

For the statsy sorts, the model above has an R-squared of 49%, which means it explains about half of why defenses are better or worse than average. More or less. And as is true for any regression-based model, it bunches the predicted points together more tightly than will actually happen (since the randomness is removed).

But there are also a couple of somewhat subjective twists that seem to make the predictions much better:

--All eight teams that surrendered less than 18 points/game for two consecutive seasons between 2001 and 2005 blew up in the third year. That's an "eyeball" analysis (literally looking at a color-coded chart of every team's defense for the past five years) and so could be random. But eight for eight is probably significant. I'm not sure what the real-world explanation is for this, but it's not hard to come up with a few hypotheses: NFL parity catches up with everyone over three years; superstar defenses hang onto players too long, etc.

--It is almost impossible for a team that gave up 24 points/game or more the previous season to turn it around in a single year. So a team that was great in the recent past, but had a really bad season the previous year is very unlikely to be a good fantasy defense.

Applied to the 2005 season the model predicted that:

--Carolina, Tampa, New York Giants, Jacksonville, Washington, Chicago, Pittsburgh and Denver would be the top eight teams. (Six finished in the top eight as predicted, seven were in the top nine and only one finished outside the Top 10, Washington @ #17).

--Baltimore, New England, Buffalo and Philly would fall from the upper eschelon. (They finished #10, #21, #27 and #30).

--Cleveland, Green Bay, San Diego, Tennessee, Minnesota, San Francisco, Atlanta, New Orleans, Kansas City and Oakland would be the bottom eleven teams. (None finished higher than #15 and eight of them finished in the bottom eleven, as predicted).

Since there's such a large random element to all of this, half of it, I don't bother trying to rank teams within each of my three clusters. I leave it at good, bad and everyone else. Overall, last year 14 of 19 teams predicted to finish in the good and bad categories did so, and NONE of the good/bad teams finished on the wrong end of the spectrum.

And since none of the teams in the "good" cluster finished worse than #17 and none of the teams in the "bad" cluster finished higher than #15 -- it really did a fantastic job of separating the teams you might want from the teams you definitely didn't.

But that's only one year. The subjective bits could be coincidence. So, I'll do a public test for 2006 and we can watch the results. Hopefully if I'm wrong you'll be kindly and generous.

--Good teams (11): New England, Tampa Bay, Baltimore, Denver, Chicago, Carolina, Washington,

Miami, Indianapolis, Seattle, Dallas

--Teams due for a fall: Pittsburgh and Jacksonville

--Bad teams (11): Detroit, Minnesota, Kansas City, Atlanta, Cincinnati, New Orleans, St. Louis, Arizona, Tennessee, Oakland, San Francisco

For comparison, here are the FBGs top and bottom eleven:

--Good: Carolina, Chicago, Pittsburgh, Baltimore, Tampa Bay, Washington, Jacksonville, Indianapolis, Green Bay, Miami, Seattle

--Bad: Detroit, Cincinnati, Tennessee, Cleveland, Kansas City, Buffalo, New Orleans, St. Louis, San Francisco, New England, Arizona, Oakland (Houston omitted).

So there are six test cases of the models where predictions differ...

--Good: New England, Denver and Dallas (me) vs Pittsburgh, Jacksonville and Green Bay (FBG)

and

--Bad: Minnesota, Atlanta and Arizona (me) v Cleveland, Buffalo and New England (FBG)

New England is the only team of the 31 where the models differ across the good/bad divide. I've got them in the top eleven, FBGs have them in the bottom eleven. So they'll be especially interesting to watch.

On the one hand, I wonder why I bothered doing all this if FBGs was going to publish an article saying much the same thing a few months later. And the models are obviously pretty similar. On the other, I had no idea there were even sites like this around six months ago and maybe there's a bit of useful info in there somewhere that will help them refine their predictions for 2006.

Along those same lines, I'm also saving my model and the FBG strength--of schedule data to see if that provides any additional insight for next year.

EDIT: NOTE THAT THE STATS MODEL WAS NOT BASED ON FBGs DATA!! My shame knows no limits, but last year was the first year I played and I was using data from Yahoo!. So the ideas will be good, but the specific #s will not work if you actually plug them in using historical data from this site.

Click to expand...