Detecting Match-Fixing in Tennis

Similar documents
A point-based Bayesian hierarchical model to predict the outcome of tennis matches

Swinburne Research Bank

04 SINGLES 04 Champions & Finalists 05 Titles, Match Win Leaders 09 Player History 13 Tournament Results 20 Playing Same Opponent Twice

FEATURED MEN S SINGLES QUARTER-FINALS

FEATURED MEN S SINGLES CHAMPIONSHIP [1] Rafael Nadal (ESP) vs [28] Kevin Anderson (RSA) Nadal Leads 4-0

How the interpretation of match statistics affects player performance

First Server Advantage in Tennis. Michelle Okereke

MEN S SINGLES QUARTER-FINALS (TOP HALF)

Men s Best Shots Poll

UNOFFICIAL WORLD JUNIOR CHAMPION AGE 12

Using Markov Chains to Analyze a Volleyball Rally

Lecture 11 and 12: Probabilistic Ranking

Existence of Nash Equilibria

Improving the Australian Open Extreme Heat Policy. Tristan Barnett

Predicting Tennis Match Outcomes Through Classification Shuyang Fang CS074 - Dartmouth College

INSIDE VOLLEY TENNIS

"TenisRank": A new ranking of tennis players based on PageRank

Grand Slam Tennis Computer Game (Version ) Table of Contents

A Markov Decision Process-based handicap system for. tennis. T.C.Y. Chan 1,2 and R. Singal,1,3. January 3, 2017

Introduction to Ranking

ROSE-HULMAN INSTITUTE OF TECHNOLOGY Department of Mechanical Engineering. Mini-project 3 Tennis ball launcher

1.1 Game Logic. League Mode: 16 Teams Home & away matches 30 match days 8 concurrent matches per match day 240 matches per season

THE BOXES UNDER THE HIGH PATRONAGE OF H.S.H. THE SOVEREIGN PRINCE OF MONACO

Sports Analytics: Designing a Volleyball Game Analysis Decision- Support Tool Using Big Data

Simulating Major League Baseball Games

USTA Junior National Tournament, Ranking, and Sanctioning Regulations

February 12, Winthrop University A MARKOV CHAIN MODEL FOR RUN PRODUCTION IN BASEBALL. Thomas W. Polaski. Introduction.

Last September, just four months before the. Smashing the racket IN DETAIL

Using Markov Chains to Analyze Volleyball Matches

REDUCING THE LIKELIHOOD OF LONG TENNIS MATCHES

Mathematics in Sports

1. OVERVIEW OF METHOD

FIRST TIME QUALIFIERS GOFFIN, DIMITROV BATTLE IN CHAMPIONSHIP

ASSESSING THE RELIABILITY OF FAIL-SAFE STRUCTURES INTRODUCTION

The final set in a tennis match: four years at Wimbledon 1

LEE COUNTY WOMEN S TENNIS LEAGUE

ADVANCED TACTICS: CONCEPTS, FEATURES AND 5 GAME SITUATIONS

Predicting Results of a Match-Play Golf Tournament with Markov Chains

The Role of Olive Trees Distribution and Fruit Bearing in Olive Fruit Fly Infestation

ALL-TIME TOP 10 in EMIRATES ATP RANKINGS (162 players )

The Effect of Pressure on Mixed-Strategy Play in Tennis: The Effect of Court Surface on Service Decisions

Opening up the court (surface) in tennis grand slams

c 2016 Arash Khatibi

FIFA World Ranking. Semester Thesis. Lukas Affolter.

[5] S S Blackman and Casey J W. Development of a rating system for all tennis players. Operations Research, 28: , 1980.

RULES OF THE COURT 2016 SUMMARY OF MODIFICATIONS

Gamblers Favor Skewness, Not Risk: Further Evidence from United States Lottery Games

MATHEMATICAL MODELLING IN HIERARCHICAL GAMES WITH SPECIFIC REFERENCE TO TENNIS

Summarizing tennis data to enhance elite performance. Tristan Barnett PhD University of South Australia

A One-Parameter Markov Chain Model for Baseball Run Production

VIRTUAL TENNIS TOUR SEASON 2014 OFFICIAL RULEBOOK

1 Introduction. 2 Review of tennis prediction models. Stephanie Ann Kovalchik* Searching for the GOAT of tennis win prediction. 2.

FEATURED MEN S SINGLES 4R MATCHES

This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and

The Monte-Carlo Country Club in 1928

SHANGHAI ROLEX MASTERS: DAY 8 MEDIA NOTES Sunday, October 15, 2017

THE FLATMATES Language point: Tennis vocabulary

Homework 2: Relational Algebra and SQL Due at 5pm on Wednesday, July 20, 2016 NO LATE SUBMISSIONS WILL BE ACCEPTED

DOWNLOAD PDF THE FEDERER EXPRESS

Grand Slams are short changing women s tennis

AGL: BASIC RULES... 3 AGL: STANDARD TOURNAMENT RULES... 6 AGL: OPEN TOURNAMENT RULES... 9 AGL: STANDARD LEAGUE RULES... 11

AEGON CHAMPIONSHIPS: DAY 6 MEDIA NOTES Saturday, June 20, 2015

Swinburne Research Bank

Alberta 55 plus Tennis Rules

Good luck to alumni Jared Hiltzik and Aleks Vukic, along with our current Fighting Illini competing this week!

Unit 3 Day 7. Exponential Growth & Decay

Railway collision risk analysis due to obstacles

ANALYSIS OF A BASEBALL SIMULATION GAME USING MARKOV CHAINS

Appendix: Tables. Table XI. Table I. Table II. Table XII. Table III. Table IV

Denise L Seman City of Youngstown

How to Play a Tennis Tiebreaker

arxiv: v1 [stat.ap] 18 Nov 2018

Player Development. Journal

Player Grading System

AGA Swiss McMahon Pairing Protocol Standards

Invincible! C R O A T I A S P E C I A L

On Probabilistic Excitement of Sports Games

THE BEHAVIOR OF GASES

Player Development. Journal

Reston Team Tennis 2015 Rules

MIAMI OPEN PRESENTED BY ITAU: DAY 12 MEDIA NOTES ZVEREV, ISNER EACH SEEK FIRST TITLE OF 2018 IN MIAMI SHOWDOWN

LEE COUNTY WOMEN S TENNIS LEAGUE

21st ECMI Modelling Week Final Report

MEN S SINGLES QUARTER-FINALS (BOTTOM HALF)

Paper Reference FM Paper Reference(s) FM201/01 Edexcel Functional Skills Mathematics Level 2

Welcome to our Tennis Sanctuary

AITA Tennis 10s and Under Competition

Exponential Decay In 1989, the oil tanker Exxon Valdez ran aground in waters near

Player Bio Personal Year By Year

A Game-Theoretic Intelligent Agent for the Board Game Football Strategy

What Causes the Favorite-Longshot Bias? Further Evidence from Tennis

INSTITUTE AND FACULTY OF ACTUARIES. Curriculum 2019 AUDIT TRAIL

ROLEX MONTE-CARLO MASTERS: DAY 8 MEDIA NOTES

FIRST-TIME MASTERS 1000 FINALISTS SEEK CINCY TITLE

Honest Mirror: Quantitative Assessment of Player Performances in an ODI Cricket Match

2018 Junior Know the Rules. Recent changes are highlighted in red.

4/27/2016. Introduction

Should bonus points be included in the Six Nations Championship?

RULES AND REGULATIONS OF FIXED ODDS BETTING GAMES

One of the most-celebrated feats

Transcription:

1 Oliver Hatfield 1 (Supervisor) 1 Lancaster University September 2, 2016

Contents 1 Introduction and Motivation 2 Simulations of Matches Markov Chains Improvements & Conclusions 3 Detecting Changes in Performance Likelihood Ratio Test Finding an Appropriate Threshold Improvements & Conclusions 4 Probabilities of Winning Matches Motivation Methodology 5 Modified ELO Rating System Why need a new rating system? Why Modified ELO? How does ELO work? Further Improvements 6 Overall Conclusions Richard Ings, Former ATP executive vice president.

Introduction and Motivation January 2016, documents released revealing widespread accusations of match-fixing. Reports of players throwing matches in return for large sums of money. Article on Buzzfeed The Tennis Racket 1 1 https://www.buzzfeed.com/heidiblake/the-tennis-racket

Introduction and Motivation Most publicised match-fixing accusation between Davydenko and Arguello. Davydenko was ranked 4th and Arguello 87th in the world. Davydenko won the first set. Surely the betting odds would have Davydenko as favourite? Figure: Davydenko was involved in the most highly publicised match-fixing accusation. Figure: Pilgrim Tennis Club where Davydenko trained.

Introduction and Motivation However the betting data had Davydenko as the underdog. Large sums of money placed on Arguello to win. Davydenko lost the match. So how do we detect match-fixing?

Simulations of Matches Simulate tennis matches to gain data to test upon. Probability of winning a point on serve is p = S ± D/2. S = 0.645. D is the difference in ability between the players. Random numbers, X i, drawn from a binomial distribution X i Bin(p, n). Random numbers, X i, determine who wins the points.

Example Random binomial numbers, 1 0 1 1 0 1, would result in: 1 15-0 1 40-15 0-0 0 1 30-15 0 1 W1 15-15 40-30

Markov Chains Markov Chains Problems with deuce and at end of tiebreaks. Absorbing Markov chains can be used. 40 40, A1 and A2 are transient states. W1 and W2 are absorbing states. p p 1-p 1-p W1 A1 40-40 A2 W2 1-p p

Markov Chains Absorbing Markov Chains Canonical form used for the transition matrix ( Q R ) 0 I Q is for the transient states. R is for the absorbing states. We use equations 1 and 2 to obtain probabilities of reaching the absorbing states. N = (I Q) 1 (1) B = NR (2)

Improvements & Conclusions Simulations of Matches Improvements: Introduce return abilities, creating better estimations of p. Track individual points within deuce and end of tiebreaks. Conclusions Programme that simulates tennis matches based on player s abilities. Creating data that can tested upon.

Likelihood Ratio Test Detecting Changes in Performance Changepoint methods used to detect changes in player s performance. Assume points are i.i.d. Likelihood ratio test used to detect single changepoints. Produces test statistic, λ, which is tested against a threshold, c.

Finding an Appropriate Threshold Finding an Appropriate Threshold Computed many simulations on simulated data with no changepoints. Found the 95% quantile of the test statistics. c = 3.76. Figure: Histogram of test statistics over multiple simulations of matches with 95% quantile marked

Improvements & Conclusions Detecting Changes in Performance Improvements: Different probabilities dependant on who s serving. Online changepoint methods. Conclusions: Detect potential changes in player s performance. Assumption of i.i.d data is unrealistic.

Motivation Probabilities of Winning Matches Need ability to calculate probability of winning matches from different scores. Then compare to the betting odds. If large discrepancies, potential case of match fixing.

Methodology Working out the Probabilities Using the Law of Total Probability: Pr(Win Match) = Pr(Win Match Win Set)Pr(Win Set) + Pr(Win Match Lose Set)Pr(Lose Set). Pr(Win Set) = Pr(Win Set Win Game)Pr(Win Game) + Pr(Win Set Lose Game)Pr(Lose Game). Probabilities calculated using absorbing Markov chains.

Why need a new rating system? Modified ELO Rating System Why do we need a new rating system? Current system only includes previous year s results. Doesn t incorporate strength of the opposition. Long term injuries result in rapid ranking declines. Figure: Del Potro has had many injuries and his rating has been severely affected in ATP system.

Why Modified ELO? Why use Modified ELO? Ratings are accumulated over all previous matches. Strength of opponent accounted for. Inactive players ratings declines slower. Expectation, E, of winning the match is calculated. E used to calculate an estimate for D. Figure: How ratings decay for an inactive player over 2 years.

How does ELO work? How does ELO work? Each player has a ratings R i,t. i is the unique player number. t is time. Expectation for each player to win the game is calculated. E i = Ratings are updated by: 1 1 + 10 (R j,t R i,t )/400. R i,t+1 = R i,t + K(S i E i ), S i is 1 or 0 depending who won the match. K is the maximum possible adjustment. K optimised to be 15.6.

How does ELO work? The Modifications Increased K by 50% for Grand Slam so more points can be gained. Exponential decay in rating points for players who are inactive for 8 weeks or more. Ratings only start to decay when completed a tournament.

How does ELO work? Initial Ratings 100 log(atp Ranking Points+1) Scales ratings to below 1000. +1 to allow for players entering system with 0 rating. Ran data, using ELO, for all matches from 2005-2016. Player ATP ELO Federer 6525 878 Roddick 3655 820 Safin 3360 812 Moya 2520 783 Coria 2400 778 Henman 2360 777 Agassi 2100 765 Nalbandian 1945 757 Table: Rankings in 2005 for ATP and new ratings for ELO.

How does ELO work? Player Ratings Figure: Graph showing changes in players ranking points. Things to note: Ratings don t decay until played first tournament When players retire, ratings decay to zero. Players with low ratings can increase rapidly.

How does ELO work? Up to Date Rankings ELO ATP Player Rating Player Rating Djokovic 920 Djokovic 16790 Federer 749 Murray 8945 Murray 694 Federer 8165 Nadal 595 Wawrinka 6865 Wawrinka 575 Nadal 5230 Nishikori 553 Berdych 4560 Berdych 550 Nishikori 4235 Ferrer 526 Ferrer 4145 Table: Table showing the top 8 players in the ELO and ATP rating systems as of 01-18-2016

Further Improvements Further Improvements Optimize increase in K for Grand Slams. Introduce different K s for different standard tournaments. Stop average ratings from decreasing due to exponential decay.

Conclusion Simulations used to generate data. Compare the probability of the player winning the match to betting odds to check for differences. If discrepancies are detected, check for changes in performance. Estimate the difference in abilities of the players using the ELO rating system s expectations of winning the matches.

Thanks for Listening! Any Questions?