Mad-Brackets Demo

Transcript

1. Mad-Brackets! Reducing bracket risk in your March Madness pool Gabe Shaughnessy
6. Web App Demo: http://mad-brackets.com Web app build with:
7. Team Statistics
8. Team Statistics Scraped last 13 years of tournament games and team data - 19 team features for each team
9. Team Statistics Scraped last 13 years of tournament games and team data - 19 team features for each team
10. Team Statistics Scraped last 13 years of tournament games and team data - 19 team features for each team Logistic regression with L1 regularization for feature selection
11. Team Statistics Scraped last 13 years of tournament games and team data - 19 team features for each team 10-fold cross-validation Logistic regression with L1 regularization for feature selection Improves on the traditional metric used by several popular sports prediction sources (Pythagorean win percentage)
12. Important team features
13. Important team features Most impactful features: - Great offensive / defensive team
 - Tournament seed
 - Luck Rating - Prior opponent’s offense - strong teams play other strong teams
14. Model the pool: Incorporate selection psychology
15. Model the pool: Incorporate selection psychology Compare: Teams advancing in ESPN pool vs. Model prediction
16. Model the pool: Incorporate selection psychology Compare: Teams advancing in ESPN pool vs. Model prediction Underdog bias High variance In close games
17. Model the pool: Incorporate selection psychology Compare: Teams advancing in ESPN pool vs. Model prediction Introduce logistic transformation Underdog bias High variance In close games
18. Model the pool: Incorporate selection psychology Compare: Teams advancing in ESPN pool vs. Model prediction Introduce logistic transformation
19. Pool of brackets Matched to ESPN pool distribution
20. Pool of brackets Matched to ESPN pool distribution M y awesom e first bracket Favorites of the model
21. Pool of brackets Matched to ESPN pool distribution M y awesom e first bracket Favorites of the model M y awesom e second bracket Monte Carlo selection from LogReg model with logistic probability transform
22. Pool of brackets Matched to ESPN pool distribution M y awesom e first bracket Favorites of the model M y awesom e second bracket Monte Carlo selection from LogReg model with logistic probability transform Determine placement vs. other brackets in pool and assume top 3 payout
23. Monte Carlo with 10000 bracket vs pool comparisons
24. Monte Carlo with 10000 bracket vs pool comparisons First bracket
25. Monte Carlo with 10000 bracket vs pool comparisons Bust! First bracket
26. Monte Carlo with 10000 bracket vs pool comparisons Bust! Second bracket First bracket
27. Web app build with:
28. First bracket: - Picked Arizona Web app build with:
29. First bracket: - Picked Arizona Web app build with: Second bracket: - Picked UConn
32. Ph.D. in Theoretical Particle Physics University of Wisconsin Gabe Shaughnessy Research: - Particle physics - Medical physics & Atmospheric Science Hobbies: - Astrophotography, nature photography - IOT devices - Basketball and billiards
33. Tournament simulation Pool size=25 participants Entry fee=$1 Payout=70%,20%,10% of total Pool scoring=1,2,4,8,16,32 pts for subsequent rounds
34. What is Luck Rating? - Measure of how well a team actually did vs. their expected performance - If a team has a lot of close games - shouldn’t win/lose a majority of them
35. Other examples of bracket pair’s performance