Quick links:
 Latest Team Rankings
 Free Text Alerts
 Member Services
ShopMobileRadio RSSRivals.com Yahoo! Sports


The Big Orange Blazer Club
Register User Options
Site: Forum:


Post New topic Post New Poll Post Reply   Page 1

Previous Topic | Next Topic | Back to Topics
here is how my algorithm has performed so far this season.
Prediction	W-L	   Pct
50-55% 45-38 .542
56-60% 50-40 .556
61-65% 61-30 .670
66-70% 61-27 .693
71-75% 76-23 .768
76-80% 60-13 .822
81-85% 95-25 .792
86-90% 102-16 .864
91-95% 166-8 .954
96+% 224-6 .974
Total 940-226 .806
The prediction column represents the win probability band my system gives to the team I expect to win. Meaning my system gives a percent confidence in each prediction, and the column represents the confidence range a game fell into. The W-L column represents my record in predicting the games in that probability band.

Overall my algorithm has predicted the winner in about 81% of games, but as Ken Pomeroy pointed out in his latest blog post, that is a meaningless statistic since so many of these games represent lopsided competitions. A better auditing statistic is how often a system is able to predict a winner in a certain probability band. For instance, my system should predict the correct winner somewhere between 50-55.9% for games that fall into that probability band. If my system is correctly predicting 65% of those games (or 35%), the system would need tweaking. So far the system has held up pretty well using this type of tracking.

There is also a matter of how well a system predicts the margin and total. Since I do not systematically track my predictions versus the spread, I cannot say how well it is doing there. What I can say though is the average difference between my predicted margin and the actual margin is about 8 points per game, which on the surface does not seem very good. My predicted total is off about 16 points on average from the actual total, with a current slight bias towards under-projecting the total. I am not yet sure whether this is due to a low sample.



Posted on 12/5 11:18 AM | IP: Logged

Just passing through, but that's a damn fine correlation, especially down at the lower confidence intervals.

It would be interesting to plot the sample size of each confidence interval over time (as a %age of the total sample size) - if your tweaked algorithm is improving the prediction capability the distribution would shift towards the higher CIs as time progresses. Of course, since you've obviously spent a lot of time on this you probably have an even better way of monitoring improvement :)



Posted on 12/5 11:59 PM | IP: Logged

I believe you really like math.



Posted on 12/12 11:08 PM | IP: Logged


Originally posted by VolunteerusMaximus:
I believe you really like math.


i do



Posted on 12/16 11:23 PM | IP: Logged

go vols



Posted on 12/16 11:24 PM | IP: Logged

vb



Posted on 12/16 11:42 PM | IP: Logged

The tracking table has been updated through games of December 31st, 2010.

Prediction W-L Pct
50-55% 86-68 .558
56-60% 85-57 .599
61-65% 100-51 .662
66-70% 104-45 .698
71-75% 142-50 .740
76-80% 108-31 .777
81-85% 173-41 .808
86-90% 169-36 .824
91-95% 273-20 .932
96+% 416-6 .986
Total 1656-405 .803




Posted on 1/2 11:57 AM | IP: Logged

Previous Topic | Next Topic | Back to Topics

Post New topic Post New Poll Post Reply Page 1

LATEST NEWS




Rivals.com is your source for: College Football | Football Recruiting | College Basketball | Basketball Recruiting | College Baseball | High School | College Merchandise
Site-specific editorial/photos © VolQuest.com. All rights reserved. This website is an officially and independently operated source of news and information not affiliated with any school or team.
About | Advertise with Us | Contact | Privacy Policy | About our Ads | Terms of Service | Copyright/IP policy | Yahoo! Sports - NBC Sports Network

Statistical information ©2007 STATS LLC All Rights Reserved.