Comparing NHL Players' Shots and Goals by Algorithmically
Total Page：16
File Type：pdf, Size：1020Kb
Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Devan Becker, Ph.D. Candidate Department of Statistical and Actuarial Sciences The University of Western Ontario, Canada Collaborators: Charmaine Dean and Douglas Woolford New England Symposium on Statistics in Sports Harvard University September 28, 2019 Devan Becker - [email protected] University of Western Ontario 1 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Introduction Introduction Devan Becker - [email protected] University of Western Ontario 2 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Introduction A Song of Ice Hockey and Forest Fire Devan Becker - [email protected] University of Western Ontario 3 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Introduction Outline • Describe the NMF-LGCP algorithm in general. • Apply it to NHL shot data. • Use these estimates to describe shot efficiency (with a twist). Devan Becker - [email protected] University of Western Ontario 4 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Image Recognition and Spatial Estimates Devan Becker - [email protected] University of Western Ontario 5 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Idea: Convert one big matrix into two smaller matrices X ≈ WH • X is a giant collection of images • In this case, LGCP estimates of the shot locations. • W defines the “bases” • Building blocks for all images • H defines the “coefficients” • How much each image uses each building block Devan Becker - [email protected] University of Western Ontario 6 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Dummy Data and Estimates Normalized soDevan that Becker the - [email protected] total volumeUniversity is 1 so of Western that Ontario we are focused on 7 location, not number of goals. Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Dummy Bases and Coefficients Devan Becker - [email protected] University of Western Ontario 8 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Choosing the number of bases • Too many bases and you’re modelling the noise. • Too few and your residuals are too large. • Balance: Interpretability and accuracy. Devan Becker - [email protected] University of Western Ontario 9 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates NHL Shot Location Data • Scraped from https://statsapi.web.nhl.com/api/v1/. • Incomplete version available from a Kaggle data set. • Always check robots.txt before scraping. • Only looked at players with > 500 shots. • Removed rebounds. • Focused on shot choice, not opportunity. In all, 510,847 shots and goals from 466 players. Devan Becker - [email protected] University of Western Ontario 10 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Estimated Bases for NHL Data Devan Becker - [email protected] University of Western Ontario 11 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Some Select Players Crease LCrease RCrease/LIce Slot Alex Ovechkin (LW−R) P.K. Subban (D−R) Ryan Getzlaf (C−R) Sidney Crosby (C−L) Connor McDavid (C−L) Anze Kopitar (C−L) 0.0 0.5 1.0 1.5 2.0 2.5 RIce RPoint LPoint Alex Ovechkin (LW−R) P.K. Subban (D−R) Ryan Getzlaf (C−R) Sidney Crosby (C−L) Connor McDavid (C−L) Anze Kopitar (C−L) 0.0 0.5 1.0 1.5 2.0 2.50.0 0.5 1.0 1.5 2.0 2.50.0 0.5 1.0 1.5 2.0 2.5 Coefficient (x100) Different position/handedness shoot from different places, but same position/handedness can be different too. Devan Becker - [email protected] University of Western Ontario 12 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Image Recognition and Spatial Estimates Basis Results - Outliers Principal Components of Coefficients C−L C−R D−L D−R 0.03 0.02 Lecavalier Malkin 0.01 Wisniewski Spooner Coburn Subban Niskanen 0.00 Richards Stoll Daley Mitchell Ott Getzlaf −0.01 Rask −0.02 Oduya −0.03 LW−L LW−R RW−L RW−R 0.03 0.02 PCA Dimension 2 Hoffman Yakupov 0.01 Thornton Voracek Pominville Hansen Drouin Jagr Steen Dupuis Pastrnak 0.00 Ovechkin TeravainenBeleskey Kopecky Laine Read Kovalchuk −0.01 Clifford Bickell Iginla −0.02 −0.03 −0.02 −0.01 0.00 0.01 −0.02 −0.01 0.00 0.01 −0.02 −0.01 0.00 0.01 −0.02 −0.01 0.00 0.01 PCA Dimension 1 Devan Becker - [email protected] University of Western Ontario 13 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Shot Quality Devan Becker - [email protected] University of Western Ontario 14 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Basketball versus Hockey For players who played at least half of the 2017-2018 season: Basketball • WORST 3-point percentage: 29.8% (Russell Westbrook) Hockey • BEST shooting percentage: 23.5% (Alexander Kerfoot) Hockey just doesn’t have enough goals per player! Devan Becker - [email protected] University of Western Ontario 15 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Where do Goals Come From? Devan Becker - [email protected] University of Western Ontario 16 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Perfect Players Instead of looking at a player’s goals, we use the goals as a player. • Goals are split up based on position/handedness. • Name of player is replaced by the position/handedness. • These “perfect players” are added to the algortihm. Devan Becker - [email protected] University of Western Ontario 17 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Coefficients for the Perfect Players Coefficients for Perfect Players Crease LCrease RCrease/LIce Slot RW−R RW−L LW−R LW−L D−R D−L C−R C−L 0.0 0.5 1.0 1.5 2.0 RIce RPoint LPoint RW−R RW−L Position/Handedness LW−R LW−L D−R D−L C−R C−L 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 Coefficient (x100) Devan Becker - [email protected] University of Western Ontario 18 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Comparing Stastny, Crosby, and Aho to the Perfect Player Crease LCrease RCrease/LIce Slot C−L Paul Stastny Sidney Crosby Sebastian Aho 0.0 0.5 1.0 1.5 RIce RPoint LPoint Player C−L Paul Stastny Sidney Crosby Sebastian Aho 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 Coefficient (x100) • Stastny: Follows the Perfect Player, “Could stand to shoot the puck more.” - EliteProspects.com • Crosby: Shoots from different place, but scores. • Aho: “Very good passer.” - EliteProspects.com Devan Becker - [email protected] University of Western Ontario 19 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Euclidean distance as shot quality Players "closest" to the perfect player have better shooting percentage Blue line is the linear trend. C−L C−R D−L D−R 0.150 0.08 0.07 0.07 0.12 0.125 0.06 0.06 0.09 0.100 0.05 0.05 0.04 0.06 0.075 0.04 0.03 0.03 0.2 0.3 0.4 0.5 0.6 0.2 0.3 0.4 0.5 0.6 0.7 0.3 0.4 0.5 0.2 0.3 0.4 0.5 0.6 LW−L LW−R RW−L RW−R 0.14 0.150 0.12 0.14 0.12 0.125 0.10 0.12 Shooting Percentage 0.10 0.08 0.100 0.10 0.08 0.06 0.075 0.08 0.06 0.04 0.04 0.2 0.4 0.6 0.3 0.4 0.5 0.6 0.3 0.4 0.5 0.3 0.4 0.5 0.6 0.7 Euclidean Distance from Perfect Player Recall: estimates were normalized to not include the number of shots or goals. Devan Becker - [email protected] University of Western Ontario 20 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Conclusions • Bases provide good foundations for heuristic advice. • “Shoot more when you’re at the point, less from the faceoff circle.” • Coefficients simplify the comparison of player shooting strategy. • Perfect players give us a measure of quality when we don’t have enough data. • Playing like a perfect player does not mean you’ll always score. Devan Becker - [email protected] University of Western Ontario 21 Comparing NHL Players’ Shots and Goals by Algorithmically Decomposing Shot Intensity Surfaces Shot Quality Acknowledgements Devan Becker - [email protected] University of Western Ontario 22.