Tournament Selection Efficiency: An Analysis of the PGA TOUR’s
FedExCup1
Robert A. Connolly and Richard J. Rendleman, Jr.
October 10, 2012
1Robert A. Connolly is Associate Professor, Kenan-Flagler Business School, University of North Carolina, Chapel Hill. Richard J. Rendleman, Jr. is Visiting Professor, Tuck School of Business at Dartmouth and Professor Emeritus, Kenan-Flagler Business School, University of North Carolina, Chapel Hill. The authors thank the PGA TOUR for providing the data used in connection with this study, Pranab Sen, Nicholas Hall and Dmitry Ryvkin for helpful comments on an earlier version of the paper and Ken Lovell for providing com- ments on the present version. Please address comments to Robert Connolly (email: robert [email protected]; phone: (919) 962-0053) or to Richard J. Rendleman, Jr. (e-mail: richard [email protected]; phone: (919) 962-3188). Tournament Selection Efficiency: An Analysis of the PGA TOUR’s
FedExCup
Abstract
Analytical descriptions of tournament selection efficiency properties can be elusive for realistic tournament structures. Combining a Monte Carlo simulation with a statistical model of player skill and random variation in scoring, we estimate the selection efficiency of the PGA TOUR’s
FedExCup, a very complex multi-stage golf competition, which distributes $35 million in prize money, including $10 million to the winner. Our assessments of efficiency are based on traditional selection efficiency measures. We also introduce three new measures of efficiency which focus on the ability of a given tournament structure to identify properly the relative skills of all tournament participants and to distribute efficiently all of the tournament’s prize money. We find that reason- able deviations from the present FedExCup structure do not yield large differences in the various measures of efficiency. 1 Introduction
In this study, we analyze the selection efficiency of the PGA TOUR’s FedExCup, a large-scale athletic competition involving a regular season followed by a series of playoff rounds and a “finals” event, where an overall champion is crowned. FedEx- Cup competition began in 2007. Each year, at the completion of the competition, a total of $35 million in prize money is distributed to 150 players, with those in the top three finishing positions earning $10 million, $3 million and $2 million, respectively.1 Research into selection efficiency highlights the importance of the criterion for assessing tournament properties.2 Most who study tournament competition em- phasize the probability that the best player will be declared the winner (“predictive power”) as the critical measure of tournament selection efficiency. Largely main- taining the focus of the selection efficiency literature on a single player, Ryvkin and Ortmann (2008) and Ryvkin (2010) introduce two additional selection efficiency measures, the expected skill level of the tournament winner and the expected skill ranking of the winner. They develop the properties of these selection efficiency measures in simulated tournament competition. While we use these efficiency measures in our work, we also develop three new measures of selection efficiency that evaluate the overall efficiency of a tourna- ment structure, not just the the mean skill and mean skill rank of the first-place fin- isher and the expected finishing position of the most highly-skilled player. Much of the existing literature (e.g., Ryvkin (2010), Ryvkin and Ortmann (2008)) assumes a specific set of distributions (e.g., normal, Pareto, and exponential) to describe com- petitor skill and random variation in performance. In this paper, we integrate an empirical model of skill and random variation in performance with a detailed tour- nament simulation to explore the selection efficiency of FedExCup competition. We do not specify the matrix of winning probabilities as in some studies; instead, it is generated naturally from the underlying estimated distributions of competitor skill and random variation and the tournament structure itself. In the next section of the paper we describe the characteristics of FedExCup competition. We develop tournament selection efficiency measures in Section 3. We present an overview of the statistical foundations of our work in Section 4, describe our simulation methods in Section 5, and present results and a discussion of practical implications of our work in Section 6. We summarize our findings in the final section. Appendix A describes the details of our simulation.
1See http://www.pgatour.com/r/stats/info/?02396. 2See Ryvkin and Ortmann (2008) for an excellent recap of existing work along these lines. 2 Characteristics of FedExCup Competition
2.1 Structure of FedExCup Competition
Under current FedExCup rules, similar in structure to NASCAR’s Sprint Cup points system, PGA TOUR members accumulate FedExCup points during the 35-week regular PGA TOUR season.3 As shown in the “Regular Season Points” portion of Table 1, points are awarded in each regular season PGA TOUR-sanctioned event to those who make cuts using a non-linear points distribution schedule, with the greatest number of points given to top finishers relative to those finishing near the bottom. At the end of the regular season, PGA TOUR members who rank 1 - 125 in FedExCup points are eligible to participate in the FedExCup Playoffs, a series of four regular 72-hole stroke play events, beginning in late August. In the Playoffs, points continue to be accumulated, but at a rate equal to five times that of regular season events. The field of FedExCup participants is reduced to 100 after the first round of the Playoffs (The Barclays), reduced again to 70 after the second Playoffs round (the Deutsche Bank Championship), and reduced again to 30 after the third round (the BMW Championship). At the conclusion of the third round, FedExCup points for the final 30 players are reset according to a predetermined schedule, with the FedExCup Finals being conducted in connection with THE TOUR Championship. The player who has accumulated the greatest number of FedExCup points after THE TOUR Championship wins the FedExCup.4
2.2 FedExCup Competition Objectives
It is clear that the objectives of FedExCup competition are multidimensional and complex. From the November 25, 2008 interview with PGA TOUR Commissioner Tim Finchem (PGA TOUR (2008)), it is possible to identify a number of these dimensions.
1. The points system should identify and reward players who have performed exceptionally well throughout the regular season and Playoffs. As such, among those who qualify for the Playoffs, performance during the regular season should have a bearing on final FedExCup standings.
3The rules associated with FedExCup competition have been changed twice. Detail about the revisions is presented in Hall and Potts (2010). 4A primer on the structure and point accumulation and reset rules may also be found at http://www.pgatour.com/fedexcup/playoffs-primer/index.html. 2. The Playoffs should build toward a climactic finish, creating a “playoff-type feel,” holding fan interest and generating significant TV revenue throughout the Playoffs. 3. The points system should be structured so that the FedExCup winner is not determined prior to the Finals. (In 2008, Vijay Singh only needed to “show up” at the Finals to win the FedExCup. This led to significant changes in the points structure at the end of the 2008 PGA TOUR season.) 4. The points system should give each participant in the Finals a mathematical chance of winning. We note that Bill Haas, the 2011 FedExCup winner and lowest-seeded player to ever win, was seeded 25th among the 30 players who competed in the Finals.5 5. The points system should be easy to understand. Under the current system, any player among the top five going into the Finals who wins the final event (THE TOUR Championship) also wins the FedExCup. Otherwise, under- standing the system, especially during the heat of competition, can be very difficult.
We do not attempt to quantify the PGA TOUR’s objectives, as summarized above. Instead, we evaluate the optimal selection efficiency of FedExCup compe- tition based on two decision variables. The first is the Playoffs points multiple. Presently, Playoffs points are five times regular season points. This has a poten- tial impact on Commissioner Finchem’s objective points 1 and 2 above. Talking with PGA TOUR officials, we understand that the TOUR reassesses the FedExCup points structure at the end of every season and that this multiple is an important part of the discussion. Reflecting these discussions, we vary the multiple between 1 and 5 in integer increments. Our second decision variable is whether or not to reset accumulated FedExCup points at the end of the third Playoffs round. The present reset system is structured to satisfy objectives 3 and 4 and guarantee that any player among the top five going into the Finals who wins the final event will win the FedExCup (objective 5, at least in part). Although we are able to identify optimal competition structures evaluated in terms of our six efficiency measures, we find that the cost of deviating from optimal structure appears to be small. This finding suggests that the costs of the implicit constraints associated with the objectives listed above may not be high.
5Although confusing, we adopt the convention used throughout sports competition that a “low” seeding or finishing position is a higher number than a “high” position. For example, in a 10-player competition, the “highest” seed is seeding position 1, while the lowest seed is position 10. 3 Measures of Efficiency
In order to measure the selection efficiency of various FedExCup competition struc- tures, we simulate entire seasons of regular PGA TOUR competition followed by four Playoffs rounds. In each simulation trial, we begin with a set of “true” player skills, or expected 18-hole scores. Throughout the regular season and Playoffs com- petition, each simulated score for a given player equals his expected score, as given by his true skill level, plus a residual random noise component. As the season pro- gresses, and throughout the Playoffs, each player accumulates FedExCup points according to a defined set of rules as described in Section 4.1. We then estimate the efficiency of the FedExCup points system using the criteria described below.
3.1 Ryvkin/Ortmann Selection Efficiency Measures
We use the following three measures of tournament selection efficiency, examined in detail by Ryvkin and Ortmann (2008) and Ryvkin (2010).
1. The winning (%) rate of the most highly-skilled player, also known as “pre- dictive power.” 2. The mean skill level (expected 18-hole score) of the tournament winner. 3. The mean skill ranking of the tournament winner.
Note that these three criteria focus on a single player, either the most highly- skilled player (predictive power) or the tournament winner. No weight is placed on the finishing positions of other players other than through their effect on the finishing position of the most highly-skilled player or the mean skill ranking or skill level of the tournament winner. We propose three new measures of selection efficiency that capture the abil- ity of a given tournament format to properly classify all tournament participants according to their true skill levels, not just the player who is the most highly skilled, and to properly allocate tournament prize money. Even if the most highly-skilled player in FedExCup competition wins most of the time, the FedExCup would surely lose credibility if the worst players in the competition could frequently finish near the top and win a significant portion of the prize money. Ideally, the FedExCup design would not only identify the single best player in the competition with high probability but would also place players in finishing positions relatively close to their true skill rankings. As such, tournament prize money would generally be the highest for the most highly skilled and lowest for the lowest skilled and, there- fore, players would be rewarded in relation to their true skill levels. Our final three measures of selection efficiency take the form of loss functions that reflect these tradeoffs.
3.2 Mean Squared Rank Error (LRE)
Consider a tournament of N players, i = 1,2,...,N, ordered by true skill (or expected score) µi, with µ1 < µ2,... < µN. Let j(i) denote the tournament finishing position of player i. For example, if the most highly-skilled player finishes the tournament in 5th position, j(1) = 5. Then µ j(i) is the inverse transformation of true skill implied by player i’s tournament finishing position, j(i), which henceforth, we refer to as “implied skill.” Finally, let Mj(i) denote the monetary prize to player i if he finishes the tournament in position j(i), with M1 > M2,... > MN. Thus, Mi denotes what player i’s prize would have been if his tournament finishing position had equalled his true skill ranking and Mj(i) denotes player i’s actual prize. Our first loss function, the mean squared ranking error, LRE, measures the extent to which the tournament errs in identifying the true skill rankings of the N tournament participants.
N 1 2 LRE = N ∑ (i − j (i)) i=1 2 = 2σR (1 − ρ), (1) 2 2 where σR = N − 1 /12 is the variance of the ranking positions, i = 1,2,...,N, and ρ is the Spearman rank order correlation of the true skill ranks, i, and tournament finishing positions, j(i). Thus, a tournament scheme that maximizes the Spearman 6 rank, ρ, will minimize the mean squared ranking error, LRE. We note that LRE weights all ranking errors equally, regardless of the actual skill differences of the players who have been miss-ranked. Our final two efficiency measures reflect these differences.
3.3 Mean Squared Skill Error (LSE)
The mean squared skill error is defined as follows:
N 1 2 LSE = N ∑ µi − µ j(i) i=1
6We note that Spearman’s footrule, another measure of ranking error, is equivalent to minimizing the sum of absolute ranking errors rather than squared ranking errors. 2 = 2σµ 1 − βµ . (2)
2 Here, σµ is the variance of true player skill, and βµ is the OLS slope coefficient associated with a regression of true player skill µi on implied player skill, µ j(i), or vice versa. When true skill rankings and tournament finishing positions are per- fectly aligned, βµ = 1, and LSE = 0. Note that if µ is linear in skill rank, LSE = LRE. The mean squared skill error takes the form of a quadratic loss function, equivalent to the loss function underlying OLS regression and Taguchi’s (2005) loss function used in quality control.
3.4 Mean Money-Weighted Squared Skill Error (LWSE)