SECFans Podcast on Michigan/Ohio State

Submitted by The Man Down T… on November 19th, 2018 at 2:56 PM

Their model tightened a touch but still has Michigan winning comfortably.  They provided a decent explanation on the Michigan/Penn State miss on the model.  35 minutes.  Fun listen.

 

 

Comments

Mgoczar

November 19th, 2018 at 3:05 PM ^

So their model missed and they are explaining it away. 

Committed murder, let me explain that away. 

Lesson is that no model is perfect. Own it. That being said, do like that they do outside of SEC games and are good natured (no curse words and not belligerent when hosting/posting etc).

secfans

November 19th, 2018 at 3:36 PM ^

It's really coming off the heels of our model nailing better teams and bigger games really well. Our model had 31-17 last year OSU, ended up 31-20. It nailed OSU/PSU last year, UM/Wisky this year. 

But accounting for a back up QB coming in and throwing a pick six to break a game open that's 14-0 going into the 4th that ends up 42-7 is a lot to ask for a robot. ;)

secfans

November 19th, 2018 at 3:52 PM ^

My point is that a game that's 14-0 with 1 minute left in the third, ends up 42-7 and people were killing us for a model that didn't predict a blowout. Which is a little unreasonable.

We weren't saying that PSU was a threat in the game (we stressed this multiple times in explaining the model discrepancy), but if PSU goes into a shell instead of a high variability mode trying to come back with a backup QB that game may end up something like 21-7. 

It was really just an interesting opportunity to talk about how data isn't perfect, but can still provide insight.

secfans

November 19th, 2018 at 3:56 PM ^

Better way of putting it: halfway through the 3rd quarter, with the game at 14-0, would you have guessed Michigan would score 42? That was our point, that if the model result seems in any way reasonable halfway through the 3rd quarter, it's done its job. The massive blowout that actually happened was a result of Penn State's high risk-high reward offense blowing up in their face. We talked quite a bit in our preview that week that it was a possibility, but it wasn't something we could really predict or model for.

It's also worth noting that our model predicted 5.36 yards per play for Michigan and they ended up with 5.84, which is pretty dang good. Predicted 4.33 for PSU and they were right at 5.0 when McSorely was in.

The PSU discussion (about variability) was a small point in a video about tOSU-UM. If you guys want to actually discuss The Game, let us know ;)

Njia

November 20th, 2018 at 2:11 AM ^

It might be tough to do with some of the statistics you’re using in your model (which seems to be based on average points allowed versus points scored), and I’m sure you’d need much more sophisticated, play by play data, but to the point about in-game coaching decisions by coaches like Frames Janklin, it should be possible to get more insight into offenses like Penn State’s that are “highly variable.” I don’t think “randomness” and “chance” are as much the issue as the effects associated with incomprehensible decisions. 

It seems to me that Field Goal Franklin costs his team, on average, about 7 points per game. The last offensive play call against OSU is a perfect example. 

That's not to say that other coaches don’t make bad calls. But I would bet that in some - maybe many - cases, the negative impacts of those decisions over time are quantifiable.

By the way, I completely agree that, at best, models like these are useful insights into how a game might go. 

secfans

November 19th, 2018 at 4:01 PM ^

OSU scores a lot of points, but they aren't efficient scorers relative to yardage. Neither team has a good red zone TD%, but it seems like UM is able to do more with less in terms of yardage.

 

I suspect UM will give up more yards, drives, and 3rd down conversions than UM fans are comfortable with - but Michigan's offense will find things much easier going than it's been even against average opponents.

This is statistically the worst defense in the Meyer era by a significant, significant margin. Their run defense is worse than Ole Miss' adjusted for opponent averages. And Ole Miss has one of the worst defenses I've ever seen.

But OSU can score, and that can keep them in any game.

Goblue228

November 19th, 2018 at 4:09 PM ^

It's strange they keep mentioning OSU offense scoring 26 against MSU.  They scored 17 (with 7 in garbage time).  Do they not account for if it's actually the offense putting up the points?

secfans

November 19th, 2018 at 4:15 PM ^

Our point was more that Haskins had 5.8 YPA vs Michigan St and with all the extra possessions that were gifted them with Lewerke being out, they still couldn't take advantage. We weren't impressed with OSU's performance vs Michigan St, in spite of us being very high on the MSU defense.

Goblue228

November 19th, 2018 at 4:26 PM ^

I just listened to this part again and where you say the MSU game gave them more points at 4.31 ypp with 26 points vs purdue 5.6 ypp and only 20 points bumped up the floor for how many points they can get per ypp.  But that doesn't make any sense if you are factoring defensive scores into offensive ypp as a measure of offensive production.  Is that how it's factored into the model?

Chitown Kev

November 19th, 2018 at 4:26 PM ^

I still want SECfans to come to the board abt. their model for the UM-Penn State game...

EDIT- Ah, I see they're here, already!...still love you guys, though.

Synful

November 19th, 2018 at 4:31 PM ^

In digging into the MSU comparisons...  sprinkle in that all of that bonus field position that OSU enjoyed they won't with Michigan as Will Hart is a monster this year punting the ball.  M has a very good shot at the 40s while OSU, they'll be excelling if they get the 24 the model suggests.  

 

chunkums

November 21st, 2018 at 10:57 AM ^

And that's somewhat propped up by earlier performances that occurred before they had some major injuries. Lewerke is a pretty decent quarterback when he's healthy but he's been lame for a while. LJ Scott is a pretty decent back but he's out for the season. When they were scoring touchdowns (something they haven't done in three weeks) they had Lewerke chucking it up to Felton Davis, who is also out for the season.

They had a bad offense before they had key injuries. With those injuries, they're probably worse than S&P suggests.  

Hail Harbo

November 19th, 2018 at 5:09 PM ^

Not mentioned is that OSU had to work Haskins into the rush offense to beat both Sparty and Maryland. Haskins doesn't run, the Bugeyes lose to the Turtles.  Don Brown sends his thanks.  

MGK10

November 19th, 2018 at 6:32 PM ^

Two things:

What team has pressured Haskins the most so far (causing hurry up throws, throw always, or being sacked)

Is The Don strongly preparing for their back up, Martell (pen state had a back up we needed to account for as well)

b618

November 19th, 2018 at 7:21 PM ^

Speaking as a guy who has built plenty of models in other realms (both professional and academic), I would like to say that, unless the thing you are modelling is very simple (which precludes football), a model isn't going to be good on every prediction.

To measure the quality of a model, you look at how its predictions do over a larger set.

It sounds like the SECfans guys have said their model performs pretty well on average.  If so, that is a significant accomplishment and maybe good enough to make them some money in sports betting.

Alas, sports betting is illegal in the state where I currently reside.

secfans

November 19th, 2018 at 10:23 PM ^

Last few years it was a really good model, it started off something like 15-0-1 ATS last season. This year it's been haywire as has all of college football. The fact that there's just a few teams at the top and the rest is a middling pile of mediocrity makes projections tough.

There's really not much difference between the 7th best team and the 20th best team. And that's a recipe for disaster in modeling.

But with good teams, or good vs average teams, or big time match ups, it tends to do well because you're relying less on variability for a win, and more on proven repeatable results.