Any stat junkies here?

Moderators: greenyellow, Autzenoise, UOducksTK1

Post Reply
maxduck
Senior
Posts: 3769
Joined: Sun Jan 11, 2009 4:37 pm

Any stat junkies here?

Post by maxduck »

Found a reference to this on another board. Obviously click on the Oregon link at the top of the page. Statistically it seems the defense has a long way to go.

https://docs.google.com/spreadsheets/d/ ... RR/pubhtml
wepto
Four Star Recruit
Posts: 818
Joined: Mon Jan 12, 2009 11:07 pm

Re: Any stat junkies here?

Post by wepto »

Statistical models aren't good at predicting college football games.

A 12 game (plus postseason) schedule just doesn't provide enough data to make particularly accurate predictions. Adding to this, a large number of games have a low chance of being competitive for four quarters--this further reduces the amount of useful data. Then, when games aren't close and go into garbage time, there is a lot of variability in the way that teams play with a lead--this reduces the quality of the already limited data; as it isn't really possible to set an objective metric for when a game goes into garbage time.


That being said, it seems to be the best publicly available tool for someone who insists on gambling on college football. It beats the spread between 50-54% of the time. Which means it's better than the average gambler. The exact model used to determine S&P+ is not published. But he publishes his predictions ahead of time.


In terms of 2018 Oregon, S&P+ factors in points per play (disclaimer: it is not the only factor). This is why the 2014 Oregon team was the #1 ranked S&P+ team of all time (at the time, not sure if that's still true); a team that produces explosion plays on offense and a bend but don't break defense will be favored by S&P+. This is also why it rates Cristobal's team below where the human polls have them.


2018 Oregon looks better in terms of points per possession than it does in points per play. But cherry picking stats is a dangerous game and you might end up hiring Don Pellum based on 'yards per play allowed'. (This is a cheap shot at Pellum and Aliotti. Don't read into it as anything more than a joke.)


Long story short, no one really knows because there is a lack of data--or at the very least, if someone does know, they haven't shared it because that knowledge is worth a lot of $$$$.
User avatar
UOducksTK1
Site Admin
Site Admin
Posts: 37590
Joined: Thu Jan 08, 2009 11:28 pm
GM: Boston Celtics GM
Location: Portland, Oregon

Re: Any stat junkies here?

Post by UOducksTK1 »

wepto wrote:Statistical models aren't good at predicting college football games.

A 12 game (plus postseason) schedule just doesn't provide enough data to make particularly accurate predictions. Adding to this, a large number of games have a low chance of being competitive for four quarters--this further reduces the amount of useful data. Then, when games aren't close and go into garbage time, there is a lot of variability in the way that teams play with a lead--this reduces the quality of the already limited data; as it isn't really possible to set an objective metric for when a game goes into garbage time.


That being said, it seems to be the best publicly available tool for someone who insists on gambling on college football. It beats the spread between 50-54% of the time. Which means it's better than the average gambler. The exact model used to determine S&P+ is not published. But he publishes his predictions ahead of time.


In terms of 2018 Oregon, S&P+ factors in points per play (disclaimer: it is not the only factor). This is why the 2014 Oregon team was the #1 ranked S&P+ team of all time (at the time, not sure if that's still true); a team that produces explosion plays on offense and a bend but don't break defense will be favored by S&P+. This is also why it rates Cristobal's team below where the human polls have them.


2018 Oregon looks better in terms of points per possession than it does in points per play. But cherry picking stats is a dangerous game and you might end up hiring Don Pellum based on 'yards per play allowed'. (This is a cheap shot at Pellum and Aliotti. Don't read into it as anything more than a joke.)


Long story short, no one really knows because there is a lack of data--or at the very least, if someone does know, they haven't shared it because that knowledge is worth a lot of $$$$.
Couldn't your model take such information into consideration? Just train the model to exclude garbage time or when starters are not in the game.

Do Not Fear. Isaiah 41:13
wepto
Four Star Recruit
Posts: 818
Joined: Mon Jan 12, 2009 11:07 pm

Re: Any stat junkies here?

Post by wepto »

UOducksTK1 wrote:
wepto wrote:Statistical models aren't good at predicting college football games.

A 12 game (plus postseason) schedule just doesn't provide enough data to make particularly accurate predictions. Adding to this, a large number of games have a low chance of being competitive for four quarters--this further reduces the amount of useful data. Then, when games aren't close and go into garbage time, there is a lot of variability in the way that teams play with a lead--this reduces the quality of the already limited data; as it isn't really possible to set an objective metric for when a game goes into garbage time.


That being said, it seems to be the best publicly available tool for someone who insists on gambling on college football. It beats the spread between 50-54% of the time. Which means it's better than the average gambler. The exact model used to determine S&P+ is not published. But he publishes his predictions ahead of time.


In terms of 2018 Oregon, S&P+ factors in points per play (disclaimer: it is not the only factor). This is why the 2014 Oregon team was the #1 ranked S&P+ team of all time (at the time, not sure if that's still true); a team that produces explosion plays on offense and a bend but don't break defense will be favored by S&P+. This is also why it rates Cristobal's team below where the human polls have them.


2018 Oregon looks better in terms of points per possession than it does in points per play. But cherry picking stats is a dangerous game and you might end up hiring Don Pellum based on 'yards per play allowed'. (This is a cheap shot at Pellum and Aliotti. Don't read into it as anything more than a joke.)


Long story short, no one really knows because there is a lack of data--or at the very least, if someone does know, they haven't shared it because that knowledge is worth a lot of $$$$.
Couldn't your model take such information into consideration? Just train the model to exclude garbage time or when starters are not in the game.
S&P+ has defined rules for what qualifies as garbage time. My points were more that:

1: It reduces available data from an already limited 12 games per team to some amount less than that.

and

2: Not every coach/team is going to see garbage time the same way as S&P+ does, which reduces the quality of data collected.
User avatar
UOducksTK1
Site Admin
Site Admin
Posts: 37590
Joined: Thu Jan 08, 2009 11:28 pm
GM: Boston Celtics GM
Location: Portland, Oregon

Re: Any stat junkies here?

Post by UOducksTK1 »

Yeah nothing you can do about the limited amount of data.

Do Not Fear. Isaiah 41:13
Post Reply