Saturday, November 08, 2014

ESPO - Introduction

ESPO - "Expected Shot Probability Outcome"


  • Chess players are rated
  • Compare ratings to generate probability of winning
  • Ratings are adjusted after a tournament or a match

Elo Rating System

  • Value used in determining score adjustment is constant

Glicko Rating System

  • Value used in determining score adjustment evolves with player

ESPO - Applying Elo

Why Is Elo an Appropriate System?

  • Many parallels between Chess and Ice Hockey
  • Two opposing parties playing to win a match
  • Evolving skill levels
  • Binary outcomes
  • Adding effects (black-white effect is similar to home-ice effect)

ESPO - Impact

Why Is Team Evaluation Important and What Can ESPO Predict?

  • Tracking which teams are "truly" the best, if the standings don't show it
  • Predicting outcomes of games, shot ratios, and goal ratios

Why Is This Model Interesting?

  • ESPO ratings evolve, so as a team's skill changes, the ratings should change
  • Since the ratings evolve, it can be used to predict an outcome at any specified point in time.
  • Effects can be added, such as home ice, score effects, etc.

ESPO - Calculating Probabilities

  • Ratings are initialized at 1500, Value used in determining score adjustment was optimized through testing
  • Expected Outcome for Team A: \(\frac{1}{1+10^{(Rating_B - Rating_A)/400}}\)
  • 80% chance of event occurring requires 240 point differential
  • 70% chance of event occurring requires 149 point differential
  • 60% chance of event occurring requires 72 point differential
  • Update ratings after each outcome: \(R_A' = R_A + K* (Outcome - Expected Outcome)\)
  • An underdog team gains more points from an upset than a favorite that wins

ESPO - Findings


  • The best teams of each season had the highest ratings (in recent years, CHI, LAK)


  • The team that shoots the puck next can be predicted with 54% accuracy
  • The team that scores the next goal can be predicted with 55% accuracy
  • Game outcomes can be predicted with 58% accuracy
  • Room for improvement, perhaps use other systems

ESPO - Sample Progression Season

plot of chunk unnamed-chunk-1

ESPO - Conclusions & Further Development


  • There is room for improvement, especially in determining the uncertainty associated with the rating itself

Further Development

  • Kalman filters use to rate teams in NBA (Poropudas, 2011)
  • Elo ratings in the NFL (Nate Silver,