2018 Pre-Season Probability Poll Discussion Thread

coit

coit

Bullseye
Joined
Nov 29, 2007
Messages
93,617
Ok folks, I've kicked off the polls for the year. We are officially 1 month away from the start of the season!

Remember that I will be tracking everyone's performance through the season and handing out kudos at the end for top performers.

Any questions?


EDIT: Based on a recommendation from TIA, we will track individual performance using a sum of each voter's probability for each win and 1 - probability for each loss. So, you get more points for a higher probability for the games we win, and more points for voting a lower probability for the games we lose.

EDIT2: We seem to have settled into a final scoring system in the comments below. Here's a summary:

The @ThisIsAtlanta / @GoGATech Scoring System

TIA Score Weighting:

For a Win, your score equals your win probability.
For a Loss, your score equals 1 minus your win probability.
  • If we win a game you predicted as 0.9 you get 0.9
  • If we lose a game you predicted as 0.9 you get 0.1
  • If we lose a game you predicted as 0.1 you get 0.9
  • If we win a game you predicted as 0.1 you get 0.1
This rewards those who realistically vote low probabilities for the games we end up losing and those who vote high probabilities for the games we end up winning.


GoGATech / 18in32 Scoring Modification:

Score starts with the TIA scoring described above. Each game is given a result value according to the following index:
  • 0.4 = three possession (>16 pt) win
  • 0.3 = two TD (>11 pt) win
  • 0.2 = two possession (>8 pt) win
  • 0.1 = TD (>3 pt) win
  • 0.0 = win or loss by < 3 pts
  • -0.1 = TD (>3 pt) loss
  • -0.2 = two possession (>8 pt) loss
  • -0.3 = two TD (>11 pt) loss
  • -0.4 = three possession (>16 pt) loss
The modified score is calculated by multiplying the TIA score by 1 minus the absolute value of each vote probability minus 0.5 minus the game result value above.

GGTScore = TIAScore * (1 - abs(VoteProb - 0.5 - GameResultValue)​

This rewards vote probabilities that are closer to final outcome by giving them a higher weight than results that are further from the actual final outcome.

----

The final result of these weights produces a score that reasonably reflects how close to the actual result a voter's predictions are on a game by game and seasonal basis.

Your mom.
 
Last edited:
coit said:
Ok folks, I've kicked off the polls for the year. We are officially 1 month away from the start of the season!

Remember that I will be tracking everyone's performance through the season and handing out kudos at the end for top performers.

Any questions?
Click to expand...

none, thanks for running this Mr. Colt.
 
Guys I'm gonna be posting the remainder of the polls every day or so through the weekend so that I'll have everything wrapped up on Friday before the first game.

Make sure you are voting in each poll, looks like folks might be missing some because there are several in play right now. Also if you missed a vote and want to add it just let me know here.
 
coit said:
Remember that I will be tracking everyone's performance through the season and handing out kudos at the end for top performers.
Click to expand...
I don't understand how this works. If the question is, "What is the probability we beat Team X in 2018?," how do you track everyone's performance? How do you determine the probability for each game?
 
coit said:
Here's the tracking thread from last season. Look at the second tab of the spreadsheet and scroll all of the way down.

https://www.stingtalk.com/board/threads/2017-preseason-probability-poll-results.94145/
Click to expand...
That seems weird to me. So let's say font A predicts 50% win likelihood for every game except Clemson, which he predicts at 90%. That would give him (on your analysis, not mine) a yearlong prediction of (11 x 0.5) + (1 x 0.9) = 6.4.

And let's say font B predicts 50% win likelihood game for all games except Clemson, which he predicts at 60%. That would give him a yearlong prediction of 6.1. Now let's say that we win all our games except Clemson, yielding a record of 11-1.

On your analysis, font A will win (since 6.4 is closer than 6.1 to our actual win total of 11), even though he and font B agreed on everything except Clemson, and font B was more correct about Clemson than font A.

What am I missing?
 
18in32 said:
That seems weird to me. So let's say font A predicts 50% win likelihood for every game except Clemson, which he predicts at 90%. That would give him (on your analysis, not mine) a yearlong prediction of (11 x 0.5) + (1 x 0.9) = 6.4.

And let's say font B predicts 50% win likelihood game for all games except Clemson, which he predicts at 60%. That would give him a yearlong prediction of 6.1. Now let's say that we win all our games except Clemson, yielding a record of 11-1.

On your analysis, font A will win (since 6.4 is closer than 6.1 to our actual win total of 11), even though he and font B agreed on everything except Clemson, and font B was more correct about Clemson than font A.

What am I missing?
Click to expand...

It isn't rocket science. Just an attempt to predict the season record. I am only tracking individual performance on a W/L basis where anything .6 or higher means you think we will win and .5 or less means you think we will lose.

I think the most interesting thing about this is how the probabilities versus different opponents change year by year.
 
coit said:
It isn't rocket science. Just an attempt to predict the season record. I am only tracking individual performance on a W/L basis where anything .6 or higher means you think we will win and .5 or less means you think we will lose.

I think the most interesting thing about this is how the probabilities versus different opponents change year by year.
Click to expand...

You should invert the probabilities for each loss, then sum the seasons probabilities to gauge individual performance.
 
ThisIsAtlanta said:
You should invert the probabilities for each loss, then sum the seasons probabilities to gauge individual performance.
Click to expand...

I like where you are going, but inverting their probability for a loss would reward the idiots that vote 0.1 every game. If we went 11-1, a voter that votes 0.1 every time would end up with a higher score than a voter that voted 0.9 every time.

If we did a 1-prob for each loss, that would be a better indicator don't you think?
 
coit said:
I like where you are going, but inverting their probability for a loss would reward the idiots that vote 0.1 every game. If we went 11-1, a voter that votes 0.1 every time would end up with a higher score than a voter that voted 0.9 every time.

If we did a 1-prob for each loss, that would be a better indicator don't you think?
Click to expand...

Yes, that's what I was getting after. Invert was the wrong word in retrospect. :lol:
 
I hate to say but I still don't like it. Say we are in a nail-biter game that we "should" easily win, and lose very closely or at the last second by some fluke. Someone that gave us a 0.1 probability (probably just being smartass) to start with will gain the most points.
 
To be the most fair, probability of winning should reflect on final score. Any game we have a 0.9 probability to win should be somewhere around a +20 margin of victory and vice-versa.
 
GoGATech said:
To be the most fair, probability of winning should reflect on final score. Any game we have a 0.9 probability to win should be somewhere around a +20 margin of victory and vice-versa.
Click to expand...

I like that one as well. Anyone have ideas on how to implement? How about this?

0.9 = >20 point victory
0.8 = >15 point victory
0.7 = >10 point victory
0.6 = >5 point victory
0.5 = toss up
0.4 = >5 point loss
0.3 = >10 point loss
0.2 = >15 point loss
0.1 = >20 point loss


Each game would be assigned one of those values. Then your score for that game would be your probability divided by that score.

Need some other sort of factor though based on win and loss I think.
 
coit said:
I like that one as well. Anyone have ideas on how to implement? How about this?

0.9 = >20 point victory
0.8 = >15 point victory
0.7 = >10 point victory
0.6 = >5 point victory
0.5 = toss up
0.4 = >5 point loss
0.3 = >10 point loss
0.2 = >15 point loss
0.1 = >20 point loss


Each game would be assigned one of those values. Then your score for that game would be your probability divided by that score.

Need some other sort of factor though based on win and loss I think.
Click to expand...

Could probably derive the point/prob intervals based on scoring data from last season. Or if you're doing it at the end of this season, you could use this season's data.
 
ThisIsAtlanta said:
Could probably derive the point/prob intervals based on scoring data from last season. Or if you're doing it at the end of this season, you could use this season's data.
Click to expand...


I have updated the spreadsheet to have the @GoGATech modification as follows. I take the @ThisIsAtlanta scoring and I assign the result of each game as follows:

0.4 = >20 point victory
0.3 = >15 point victory
0.2 = >10 point victory
0.1 = >5 point victory
0.0 = toss up
-0.1 = >5 point loss
-0.2 = >10 point loss
-0.3 = >15 point loss
-0.4 = >20 point loss

I then take each voter's probability, subtract 0.5 to normalize it with the values above, and take 1 - abs(vote-result). That produces a weighting factor between 0.2 and 1 depending on how close your vote was to the result. Then I multiply that weighting factor by the TIA factor.

As for expectations going in, I could use the ESPN FPI numbers to determine a predicted point differential, factoring in 3 point home field advantage. But I wonder if using the results for the weighting factor wouldn't be better? Otherwise a voter would be penalized for predicting a high probability in a game we are heavily favored to lose. Right?
 
You must log in or register to reply here.
Back
Top