In Search of Sure Things An Analysis of Market Efficiency in the NFL Betting Market

Similar documents
Using Actual Betting Percentages to Analyze Sportsbook Behavior: The Canadian and Arena Football Leagues

Sportsbook pricing and the behavioral biases of bettors in the NHL

Market Efficiency and Behavioral Biases in the WNBA Betting Market

Hot Hand or Invisible Hand: Which Grips the Football Betting Market?

Working Paper No

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

Determinants of NFL Spread Pricing:

Information effects in major league baseball betting markets

STUDY BACKGROUND. Trends in NCAA Student-Athlete Gambling Behaviors and Attitudes. Executive Summary

Revisiting the Hot Hand Theory with Free Throw Data in a Multivariate Framework

All the Mail Pay instructions will be clearly printed on the back of your winning ticket!

CONTENTS SPORTS BETTING 101 TYPES OF SPORTS WAGERS WELCOME TO IMPORTANT TERMS GLOSSARY FOOTBALL BASKETBALL BASEBALL HOCKEY SOCCER BOXING/MMA

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

Peer Reviewed. Abstract

Predicting the Total Number of Points Scored in NFL Games

Has the NFL s Rooney Rule Efforts Leveled the Field for African American Head Coach Candidates?

PREDICTING the outcomes of sporting events

GAMBLING PROBLEM? CALL GAMBLER.

Is lung capacity affected by smoking, sport, height or gender. Table of contents

Case 2:18-cv WJM-MF Document 1-1 Filed 10/23/18 Page 1 of 17 PageID: 13 EXHIBIT A

Compression Study: City, State. City Convention & Visitors Bureau. Prepared for

WELCOME. Within this brochure, you will. find information to assist you with. your wagering on a variety of sports. events. If you have any questions

The Project The project involved developing a simulation model that determines outcome probabilities in professional golf tournaments.

Sport Hedge Millionaire s Guide to a growing portfolio. Sports Hedge

1. Answer this student s question: Is a random sample of 5% of the students at my school large enough, or should I use 10%?

Psychology - Mr. Callaway/Mundy s Mill HS Unit Research Methods - Statistics

Staking plans in sports betting under unknown true probabilities of the event

WELCOME TO FORM LAB MAX

WELCOME CONTENTS SPORTS BETTING 101 TYPES OF SPORTS WAGERS. Within this brochure, you will find. information to assist you with your

Racial Bias in the NBA: Implications in Betting Markets

Overreaction in Football Wagers

Individual Behavior and Beliefs in Parimutuel Betting Markets

Why are Gambling Markets Organized so Differently than Financial Markets? *

Stats 2002: Probabilities for Wins and Losses of Online Gambling

Kelsey Schroeder and Roberto Argüello June 3, 2016 MCS 100 Final Project Paper Predicting the Winner of The Masters Abstract This paper presents a

B. AA228/CS238 Component

Power-law distribution in Japanese racetrack betting

Team 1. Lars Eller vs. Montreal Canadiens. Submissions on behalf of Montreal Canadiens (Team Side)

Analyses of the Scoring of Writing Essays For the Pennsylvania System of Student Assessment

Mathematics of Pari-Mutuel Wagering

Surprisals Revisited. Outcome Surprisal

How Effective is Change of Pace Bowling in Cricket?

Building an NFL performance metric

Does the Hot Hand Drive the Market? Evidence from Betting Markets

A Failure of the No-Arbitrage Principle

Deciding When to Quit: Reference-Dependence over Slot Machine Outcomes

Modeling Fantasy Football Quarterbacks

How To Bet On Football

Contingent Valuation Methods

Opleiding Informatica

Machine Learning an American Pastime

SPORTS WAGERING BASICS

a) List and define all assumptions for multiple OLS regression. These are all listed in section 6.5

Running head: DATA ANALYSIS AND INTERPRETATION 1

Predictors for Winning in Men s Professional Tennis

A SURVEY OF 1997 COLORADO ANGLERS AND THEIR WILLINGNESS TO PAY INCREASED LICENSE FEES

Evaluating The Best. Exploring the Relationship between Tom Brady s True and Observed Talent

By Rob Friday

ELO MECHanICS EXPLaInED & FAQ

College Teaching Methods & Styles Journal First Quarter 2007 Volume 3, Number 1

1. OVERVIEW OF METHOD

POWER Quantifying Correction Curve Uncertainty Through Empirical Methods

Quantitative Methods for Economics Tutorial 6. Katherine Eyal

LEE COUNTY WOMEN S TENNIS LEAGUE

Should bonus points be included in the Six Nations Championship?

Stat 139 Homework 3 Solutions, Spring 2015

Analysis of recent swim performances at the 2013 FINA World Championship: Counsilman Center, Dept. Kinesiology, Indiana University

NFL Overtime-Is an Onside Kick Worth It?

Online Companion to Using Simulation to Help Manage the Pace of Play in Golf

2010 TRAVEL TIME REPORT

University of Nevada, Reno. The Effects of Changes in Major League Baseball Playoff Format: End of Season Attendance

Our Shining Moment: Hierarchical Clustering to Determine NCAA Tournament Seeding

Evaluating the Influence of R3 Treatments on Fishing License Sales in Pennsylvania

Clutch Hitters Revisited Pete Palmer and Dick Cramer National SABR Convention June 30, 2008

STANDARD SCORES AND THE NORMAL DISTRIBUTION

Matt Halper 12/10/14 Stats 50. The Batting Pitcher:

Whitepaper. V2 August BetKings

Efficiency Wages in Major League Baseball Starting. Pitchers Greg Madonia

An Analysis of Factors Contributing to Wins in the National Hockey League

c 2016 Arash Khatibi

Using Markov Chains to Analyze a Volleyball Rally

Applying Occam s Razor to the Prediction of the Final NCAA Men s Basketball Poll

Review of A Detailed Investigation of Crash Risk Reduction Resulting from Red Light Cameras in Small Urban Areas by M. Burkey and K.

Determining Occurrence in FMEA Using Hazard Function

American Football Playoffs: Does the Math Matter? that one of the most prominent American sports is football. Originally an Ivy League sport,

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

ANALYSIS OF SIGNIFICANT FACTORS IN DIVISION I MEN S COLLEGE BASKETBALL AND DEVELOPMENT OF A PREDICTIVE MODEL

Extreme Shooters in the NBA

Business Cycles. Chris Edmond NYU Stern. Spring 2007

Most hit keno numbers ohio

Learning about Banks Net Worth and the Slow Recovery after the Financial Crisis

Does Momentum Exist in Competitive Volleyball?

LEE COUNTY WOMEN S TENNIS LEAGUE

The sentiment bias in the market for tennis betting

Analysis of performance at the 2007 Cricket World Cup

Two Machine Learning Approaches to Understand the NBA Data

Average Runs per inning,

Department of Economics Working Paper Series

Why We Should Use the Bullpen Differently

Online Read One Ball Handicap Football Betting System how to get books for free on google play

Transcription:

In Search of Sure Things An Analysis of Market Efficiency in the NFL Betting Market By: Rick Johnson Advisor: Professor Eric Schulz Submitted for Honors in Mathematical Methods in the Social Sciences Spring, 2007

Abstract The market for betting on National Football League games is well established and can be observed on a weekly basis for four months every year. Despite the visibility of the market, past research has tested its efficiency from several angles and reported a variety of inefficient areas. This thesis uses more recent data and a system of probit models to test if some of these inefficiencies continue to exist and if they can be consistently exploited to earn above average returns. While a model that incorporates measures of overall team strength and recent performance can be used to reliably predict in-sample results, its ability to predict out-of-sample games is significantly weaker and it can only be used to generate excess profits if the gambler observes an extreme cut-point strategy. Furthermore, this model is also applied using a varying number of years to predict outcomes during the following season. The results suggest using a mid-range number of years, such as three, may actually generate the most successful model as inefficiencies evolve over time and a model that bases its predictions on more years may not capture recent trends as well.

Acknowledgements I would first like to thank my advisor, Professor Eric Schulz, for his help and guidance throughout the MMSS thesis process. I walked into his office with broad ideas and he helped me focus them and made me think deeper about the results I found. In addition, I would like to thank my technical advisors, Kisa Watanabe and Scott Payseur, for providing additional advice on the Econometric methods I use. Furthermore, I would like to thank all my colleagues at Deloitte who not only were willing to work around the unique constraints placed on me while finishing my thesis, but also provided an immense amount of support along the way. From a more personal standpoint, I would like to thank my parents, siblings, and the rest of my family, who have always supported my academics and pushed me to succeed. Additionally, I would like to thank my friends and girlfriend for all the good times they have provided me with over the past four years.

Table of Contents I. INTRODUCTION...1 II. THE BETTING MARKET FOR FOOTBALL...3 III. REVIEW OF EXISTING LITERATURE...6 IV. DATA...15 V. DATA ANALYSIS AND RESULTS...18 VI. DISCUSSION...34 VII. CONCLUSION...37 VIII. WORKS CITED...38

Johnson 1 I. Introduction Every year an increasing amount of money is gambled on sports in America. While one could argue the market used to have significant barriers to entry, that case can no longer be made. Sports gambling used to only be legally available in select locations in the United States, such as Las Vegas and Atlantic City, or through use of illegal channels, such as a bookie a person who takes bets. However, since 1995, reputable internet gambling sites have grown and become popular; so popular, in fact, that it has become a major focus of United States legislative bodies. Although this has caused the industry to adapt and find loopholes in the law, internet gambling is still readily available to anyone with a credit card and internet connection. While internet gambling provides lower barriers to entry, a relatively small amount of money is wagered online an estimated $11.9 billion, including all forms of gambling such as poker and other casino games. On the other hand, $2.43 billion was wagered in Nevada s sportsbooks in 2006 and the National Gambling Impact Study Commission estimates as much as $380 billion was wagered illegally on sports 1. This shows that although the industry is easy to access through the internet, the majority still prefer more traditional, yet illegal, forums. With an industry so widely accessible, one must wonder if the market truly is efficient or if there are arbitrage opportunities that persist due to inherent patterns of irrationality on the part of bettors or, investors. Therefore, I have attempted to find answers to the questions: Are there market inefficiencies apparent in historical data that can be systematically exploited to earn a profit? If these inefficiencies exist, are they similar to most other arbitrage opportunities that disappear over time, or can they actually be applied to out-of-sample data to achieve similar results? 1 http://www.americangaming.org/

Johnson 2 In order to explore these questions, I employed a variety of econometric methods and tested the applicability of other papers findings to present day data and found some of the inefficiencies reported still exist, although they have weakened over time. In addition, I tested the predictive power of models generated using historical results and found the only way these inefficiencies can be exploited is by observing extreme cut-point strategies which would limit the number of betting opportunities a gambler would have over the course of a season.

Johnson 3 II. The Betting Market for Football Betting on football is unlike other forms of betting. Rather than betting on a single event, or a combination of events, like most other forms of betting, the bettor chooses a team to win against the spread. In theory, the spread (S) is a value that if added to the underdog s number of points at the end of the game (U) would equal the favorite s number of points (F). Or, in mathematical terms, F = U + S Therefore, in betting on the favorite to beat the spread, the bettor is predicting: F > U + S And the opposite is true if the bettor predicts the underdog will beat the spread. This method of betting is different from most other sports, which use odds betting meaning, given a fixed wager, one event will pay a different amount than another if it is the correct outcome. For example, if football used the odds betting system and the San Francisco 49ers were favored over the Green Bay Packers, the 49ers might have odds of -310 and the Packers might have odds of +255. This means if you bet on the correct team to win outright, you would have to bet $310 on the 49ers to win $100 ($100 would win $32.26) or $100 on the Packers to win $255. However, in a spread system, a correct bet on one team to cover the spread would typically earn the same amount as if you had bet correctly on the other team. While the spread is also being used more commonly in basketball betting, it is not widely used in low-scoring sports, such as baseball and soccer; sports with multiple competitors, such as races or golf; or sports where defining a margin of victory is difficult, such as boxing. Another interesting characteristic of betting markets is the role played by the bookie, or the house in cases where betting in a sportsbook a legal forum for placing sports bets.

Johnson 4 Rather than betting against the bettor, the bookie can be better characterized as a conduit for betting against other bettors, and the bookie charges a commission for his services. Typically, in a game in which you are betting on the spread, the payout for both teams will be -110 meaning whichever team you bet on, you must wager $110 to win $100. In the world of betting, this is commonly referred to as the 11 for 10 rule and the commission the bookie is earning is referred to as the vigorish. This means if you give the bookie a bet of $110 and win, he will give you $210 back; however, if you lose he will give you nothing back. Therefore, we can define WP* as the winning percentage necessary to break even and calculate accordingly: (WP* x 210) + [(1 WP*) x 0] = 110 WP* = 52.4% So, a bettor must predict 52.4 percent of games correctly in order to break even. In this situation, the bookie is essentially placing an even money bet between two people of $105 and charging each person $5 to place the bet a commission of 4.8 percent per bet. Therefore, as long as the bookie can keep the amount of money wagered equal on each team, he earns a 4.8 percent commission, or 4.5 percent of the total money wagered, with absolutely no risk. While the bookie s task may sound easy, they must be sure to keep an equal amount of money on each side in order not to be gambling themselves. Therefore, lines for an upcoming game are typically released early in the week, Tuesday or Wednesday for a Sunday game, and allowed to move if there is too much money wagered on one side. However, this can create a unique problem in which the bookie may get middled. For example, if the opening line in the 49ers and Packers game is the 49ers by seven points and the majority of the public feels they will win by more than this and bet accordingly, the line will move upwards. However, if the line moves too far, to ten for example, 49ers supporters may stop betting and Packers supporters

Johnson 5 may start betting on them. Thus, the majority of bets will be on either the 49ers at -7 or the Packers at +10, and there will be very few on the 49ers at -10 and the Packers at +7. This is the worst possible situation for a bookie to be in because if the actual game spread is 8 or 9, the bookie will lose the vast majority of the bets he accepted. Therefore, the opening lines are determined by experience oddsmakers and are reluctantly changed especially by a substantial margin. One final characteristic of spread betting worth mentioning is importance of the halfpoint. While it may not make sense to say a team is favored by ten and a half points, the majority of lines use this terminology. Thus, instead of the 49ers being favored by ten points, they might be favored by nine and a half points. In these cases, if the 49ers win by ten, a bettor who bet on them to cover would win; however, if they only win by nine, a bettor who bet on the Packers to cover would win. The party that derives the biggest advantage from this convention is the bookie. Traditionally, a game in which the actual spread equals the betting line results in a tie, or no action. In these cases, the bookie refunds both bettors money and earns nothing. However, if the line could have been set at nine and a half points instead of ten without dramatically changing the level of action on each side, the bookie would earn his commission on all bets. Therefore, it is in the bookie s best interest to use a line with a half-point whenever possible.

Johnson 6 III. Review of Existing Literature The pool of existing literature on the football betting market, and the betting market as a whole, is fairly limited. In addition, the majority of empirical research that has been done is not current and there is no guarantee certain trends still exist due to changes in the market structure. Therefore, one of the primary goals of this paper is to examine models and results from existing literature, see if their results are still applicable in the present day, and attempt to improve on their theories. Spread vs. Odds One important feature of the football betting market, previously mentioned, is its reliance on spread betting compared to the more conventional odds betting. As this is a major characteristic of the market, multiple authors have investigated why this is the case and, although these models will not be critiqued or improved upon in this paper, their results are worth noting. Prior to these analyses, the conventional wisdom was spreads were the dominant medium for betting due to bettor sentiment by making games more interesting to watch. For example, in football there are often heavy favorites and points come in large chunks, generally three or seven, rather than one at a time. A byproduct of this is games can exhibit large point swings in a short amount of time. Therefore, if a bettor is watching a game with a point spread, he is essentially watching two games in one the competition to be the absolute winner and the competition to be the winner against the spread which is likely to generate more interest. Spread betting also makes one-sided games more interesting because even if the favorite is winning and the losing team does not have a realistic chance of winning the game, they have a better chance of beating the spread and, therefore, the game might still come down to the last minute in the eyes of the bettor even though the outcome has been determined.

Johnson 7 While Bassett does not reject these theories of preference towards spread betting, and even believes they do have influence, he presents a model where the bookie is allowed to set the spread and odds for a given game. He demonstrates that, given unimodal and symmetric belief functions for the bettors, presenting the bettor with a spread, rather than odds, is the profitmaximizing behavior of a bookie. However, Woodland and Woodland challenge Bassett s assumption of risk neutrality due to his assumption bettors do not gamble a significant proportion of their wealth. Woodland and Woodland argue this unnecessarily restricts the size of the bet a bettor can choose in Bassett s model. They present a model that demonstrates the spread system is indeed the profit-maximizing choice for a bookie, but primarily due to the risk aversion in bettors. They show in games where neither team is thought to have a significant advantage, the amounts wagered will be similar under spread and odds betting. However, if one team is thought to have an inherent advantage, less will be wagered by a risk adverse bettor under the odds system than the spread system. Therefore, because bookies, in theory, make a percentage of total bets placed, they will make more under the spread system. Predicting the Winner of a Game Another focus of literature on the betting market has been developing ways to predict the absolute winner of a given game. While these papers are not direct tests of market efficiency, they are relevant to research as they demonstrate the predictive value of the spread. Harville uses a complex statistical algorithm to demonstrate he can use public information to predict the winner of a game. When applied to actual results, Harville predicts the absolute winner of a game 70.3 percent of the time, based on 1,320 games from 1971 to 1977. However, Harville s success rate is overshadowed by the fact that the team favored by the spread is the absolute winner in 72.1 percent of games in the same sample.

Johnson 8 Stern also attempts to model a team s probability of winning a given game; however, he uses far simpler methods. Using data from the 1981 to 1984 seasons, Stern calculates the margin of victory over the spread defined as the favorite s actual score minus the underdog s actual score minus the spread for each game. This statistic for Stern s sample is normally distributed with a standard deviation of 13.86 points and a mean of.07, which he rounds to zero for simplicity. Stern conducts a Chi-squared test to show a normal distribution with a mean of zero and standard deviation of 13.86 can not be rejected. Therefore, the probability of a favorite (underdog) winning a game is the area under the normal curve centered at the point spread, and with standard deviation of 13.86, and below (above) zero. To test his results Stern simulates the 1984 regular season ten thousand times and is able to correctly predict five of six division winners. However, Stern does not attempt to apply his simulation to out-of-sample games. Perhaps the most relevant piece of Stern s work is that he is not able to reject a normal distribution centered on zero for the margin of victory over the spread. Efficiency of the NFL Betting Market Turning to the focus of this paper, there is a substantial amount of literature on the efficiency of the NFL betting market. On one hand, there is research that claims to identify systematic errors in the betting market that can be exploited, even when looking at out-of-sample data. However, there are just as many pieces of literature either commenting on these papers, or presenting data and models of their own, in support of market efficiency. In the middle, there are papers that identify large errors in the market but argue they cannot be systematically exploited. One such paper is by Pankoff, and is widely regarded as the first piece of literature to address the efficiency of the betting market. To test the efficiency of the market, Pankoff relies on an analysis of the mean square error of the market forecasts from 1963 to 1965. He attempts

Johnson 9 to adjust the unadjusted mean square error by subtracting a mean error and a variance error. The mean error is intended to account for home-field advantage not being accurately valued and the variance error which incorporates constants from regressing actual spreads on forecasted spreads is intended to account for the market s inability to accurately judge the teams average abilities to score points. However, subtracting these error terms from the mean square error only reduces the mean square error by one percent. Therefore, Pankoff concludes market forecasts will not be amenable to much, if any, statistical improvement (Pankoff). He elaborates on this conclusion by arguing the large mean square error implies the market is not successful in forecasting game outcomes, however there are not patterns that would allow this inefficiency to be exploited. Although Pankoff does not believe these errors can be statistically exploited, he does not rule out the possibility of superior analysts being able to consistently beat the market. In fact, he presents an example where the predicted spreads of five different analysts, and the median of their predictions, were compared to the betting spread. A strategy of placing a bet if the absolute value of the analyst s prediction minus the spread is greater than a cut-point, which varies, is employed. Pankoff finds that as the cut-point is increased, so does the probability an analyst will place a correct bet and, at the highest level of the cut-point, four of the five analysts and the median of them earn a profit. Therefore, while Pankoff does not believe a statistical model can outperform the market, he does not eliminate the possibility that some bettors may be able to outperform it. Another early paper is by Vergin and Scriabin, who look for statistical significance in a variety of common betting methods in games from 1969 to 1974. Some best bet of the week methods, such as placing a bet on the team which outscored its opponent by the most points in

Johnson 10 the last n weeks or betting on the team that beat the spread by the most points in the past n weeks, showed some statistically significant positive returns, depending on what parameters are used. However, a similar method of betting on the team that won by the largest margin in the previous week s game actually provides negative, albeit not significant, results. Perhaps the most consistent and applicable method they analyze is that of betting on the underdog. They segment games by the value of the spread into categories of [0, 5], (5, 10], (10, 15], and >15, and find betting on the favorite tended to yield positive returns only when it is a small favorite, with a significance level of.098. On the other hand, betting on underdogs in the (5, 10], (10, 15], and >15 categories yields positive returns with significance levels of.085,.378, and.049, respectively. However, upon further examination, they conclude it is logical to group all games with a spread >5 together, as these are all games with a heavy favorite. Betting on any underdog given at least five points yields a winning percentage of 54.6 percent, with a significance level of.017, which is enough to earn a profit. Another test of efficiency is done on a weekly basis by Zuber, Gandar, and Bowers. Using data from the 1983 season, they regress the actual spread on the betting spread and test the null of efficiency the constant equal to zero and the coefficient on betting spread equal to one. They are unable to reject the null for 13 of the 16 weeks, which might imply an efficient market. However, they also use the same model to test against the null of both the constant and the coefficient being equal to zero which would suggest no correlation between the actual game result and the predicted spread. In this test they are unable to reject the null in 15 of the 16 weeks, which suggests an inefficient market and, thus, they search for a different way to test efficiency. To do this, they regress the actual spread on several explanatory variables which are the differences of the home team s value minus the visitor s value for a variety of statistics, such

Johnson 11 as yards rushed, yards passed, number of wins, etc., for games through the first eight weeks of the season. Then, for the next eight weeks, they use the generated model and collected statistics to compute their projected point spread for each game. If the difference between their point spread and the betting spread is greater than a value, λ, they place a bet. Although they float λ to higher levels, at its lowest level of.5, the model generates 102 bets; 60 of which are correct. If one were to assume the probability of picking a winner is.5, there is less than a five percent chance of picking 60 out of 102 games correctly (Zuber, et al). Although their model generates results, Zuber, at al, concede they do not prove market inefficiency because their data is limited, there is a significant cost to collecting and analyzing the data which would have to be included, and, as gambling was illegal in 1983, there might be additional transaction costs that are not included. While Zuber, et al, admit limitations of their model, Sauer, Brajer, Ferris, and Marr wrote a comment on their paper in defense of the efficient markets hypothesis. They criticize the fact that Zuber, et al, attempt to reject the null hypotheses on a weekly basis because the small sample size within each week would make it very hard to reject a null hypothesis. Therefore, Sauer, et al, run the same regressions using data from the entire 1983 season as one sample. They are not able to reject the null of the efficient market hypothesis, but are able to reject the null that both the constant and coefficient are zero, which demonstrates there is a correlation between actual outcome and the betting spread. In addition, Sauer, et al, criticize the construction of Zuber s, et al s, model. They argue the efficient markets hypothesis states the betting line incorporates all public information and cannot be improved. Therefore, the spread should be included in the model in addition to the explanatory variables, and if any coefficient on another explanatory variable is significantly

Johnson 12 different from zero the efficient markets hypothesis can be rejected. Sauer, et al, run this regression and are not able to reject the null that the coefficient is equal to zero for any of the additional explanatory variables, thus further supporting the efficient markets hypothesis. Finally, Sauer, et al, assume Zuber s, et al s, model for 1983 and run a similar test for 1984. They calculate the coefficients based on the first eight weeks of 1984, and apply their model to the second half. However, unlike Zuber s, et al s, simulation, Sauer, et al, only predict 39 of 101 bets correctly, thus concluding Zuber, et al, have not constructed a model that can consistently produce profitable gambling opportunities (Sauer, et al). Another inefficiency that has been reported, but will not be addressed by this paper, is the movement in the line from when it opens to game time. Gandar, Zuber, O Brien, and Russo examine the patterns of lines moving to see if the general public moved the spread in the correct direction or away from the eventual result. They find a simple betting strategy of betting against the public betting on a favorite that became less of a favorite or betting on an underdog that became a bigger underdog was successful 54.9 percent of the time from 1980 to 1985, which has a significance value of.002. Avery and Chevalier conduct a similar study of games from 1984 to 1994. They calculate a model to predict the movement of a line over the course of the week and bet against the predicted public movement. Using this strategy, they are able to win 50.5 percent of bets when betting on all games, but if they segment the sample and only bet on larger predicted swings, their winning percentage increases. However, they also run a test similar to Gandar, et al, and bet against the actual movement in the spread and are not able to generate a profit. Therefore, this is yet another speculative area in which the research can not reach a consensus.

Johnson 13 As for the focus of this paper, it will mainly expand on the work done by Gray and Gray in their paper, Testing Market Efficiency: Evidence from the NFL Sports Betting Market. Gray and Gray do not concern themselves with the typical explanatory models to test market efficiency that are primarily rooted in the OLS framework. Rather, they develop a variety of probit models to predict the probability a given team will cover the point spread. Their primary model is specified by regressing a binomial variable, 1 if the home team wins and 0 if they lose, on the season winning percentage against the spread of both teams, the number of wins against the spread each team has in the past four games, and a dummy variable that is 1 if the home team is the favorite. The logic behind using both overall winning percentages and recent winning percentages is analogous to the stock market theory that a stock which is overall valuable, but has performed poorly recently, will be underpriced and a stock that has recently done well, but is not actually valuable, will be overpriced. Although the coefficients of all variables are not independently significant the home winning percentage and the away wins in last four games do not reach significance Gray and Gray argue the two winning percentage variables should be considered jointly, as well as the two streak variables. The logic for this is: a team that has been performing well (relative to expectations), over the course of a season to date is more likely to beat the spread than a team that has been performing poorly. This is consistent with the idea that the market is slow to realize that a particular team is having a particularly good (or bad) year. As for the rationalization of the joint test for the streak variables, Gray and Gray argue: a team that has been performing well (relative to expectations), over the past four games is less likely to beat the spread than a team that has been performing poorly over the last

Johnson 14 four games. This is consistent with the idea that the market overreacts to recent form, discounting the performance of a team over the season as a whole. In addition to these sets of variables being jointly significant, the coefficient on the dummy variable for the home team being the favorite is negative and statistically significant which implies if a home team is the underdog it is more likely to beat the spread. The coefficients for this model are specified using data from the 1976 to 1994 season which is by far the largest sample analyzed and the model is able to correctly predict the winner of 54.6 percent of the games within the sample. However, when only specifying the model using data from 1976 to 1992, and predicting the outcomes for 1993 and 1994 the model predicts the correct winner 56.0 percent of the time, which has a significance level of.013 when compared to 50.0 percent and.091 when compared to 52.4 percent the minimum to earn real profits. In addition, Gray and Gray run similar tests using different cut-point strategies. While returns tend to increase, and be more significant, for the in-sample data, the results were mixed for out-of-sample games.

Johnson 15 IV. Data The primary source of data for this paper is The Gold Sheet a popular resource for sports bettors. The Gold Sheet archives provide complete scores, closing lines, dates, and location for every regular season and playoff game from 1993 to 2005. However, due to the fact that playoff games may be atypical due to certain intangible factors, these games are excluded from the sample, which leaves a sample of 3,176 games. In addition to eliminating playoff games from the samples, games which the market predicted correctly and the actual spread equaled the betting spread ( no-action games ) have to be given special consideration. Due to the fact that the focus of this paper, as is Gray and Gray s, is developing probit models to predict the probability a team will win a game, these games can not be reliably incorporated into the sample. However, because most explanatory variables are based on what has already happened that season whether it be a season-long variable like total winning percentage or a streak variable like the number of wins in the last four games deciding whether or not to include these games can influence the result. While Gray and Gray do not specify how they treat this issue, I calculate historical statistics before excluding these no-action games in order to capture the true influence of the variables. After these calculations, eliminating these no-action games cuts the sample by 86 games, or 2.7 percent, to 3,090 games. Table 1 contains summary statistics of the primary data sample used for this analysis.

Johnson 16 Table 1 Summary statistics of the 3,090 NFL games analyzed from 1993 to 2005. Mean Standard Deviation Lower Bound Lower Quartile Median Upper Quartile Upper Bound Winning Score 26.51 8.77 6 20 26 31 59 Losing Score 14.99 7.94 0 10 14 20 48 Margin of Victory 11.51 9.02 0 4 9 17 49 As point spreads are a focus of this paper, it is also important to examine them. However, as pointed out by Gray and Gray, calculating any statistics of a point spread also presents a unique problem as it is defined relative to one team and is not an absolute. For example, when looking at a game with an away favorite of ten points, summary statistics can be computed by using the absolute value of the spread (10), which is in effect defining it from the perspective of the favorite; by using the number of points the home team is favored by (-10); or by choosing the Team of Record (Gray and Gray) randomly. Gray and Gray believe the most accurate way of looking at the summary statistics is taking the mean and standard deviation of the betting spread, actual spread, and difference by determining the Team of Record randomly. However, while using a random Team of Record might provide the most accurate standard deviation, the mean is close to zero, and therefore does not convey much information. Therefore, examining the spread by using the favorite as the Team of Record will provide the most meaningful mean an average spread of 5.2 points as it gives a true sense of what the spread is in a randomly chosen game. Finally, using the home team as the Team of Record may not provide better summary statistics, but it should provide an estimate of how many points home-field advantage is worth, both in bettors minds (2.47) and in actuality (2.81). Therefore, summary statistics for the point spread computed using all three methods are presented in Table 2.

Johnson 17 Table 2 Summary statistics of the betting and actual point spreads of the 3,090 NFL games analyzed, varying based on the team of record. Favorite Home Random Betting Spread Mean 5.20-2.47-0.09 Standard Deviation 3.39 5.70 6.21 Lower Quartile 3-6 -4 Median 4-3 0 Upper Quartile 7 2 4 Actual Spread Mean 5.33-2.81-0.02 Standard Deviation 13.62 14.34 14.62 Lower Quartile -3-12 -9 Median 5-3 -1 Upper Quartile 14 6 9 Difference Mean 0.13-0.34 0.07 Standard Deviation 13.13 13.12 13.14 Lower Quartile -8-9 -9 Median 0 0 0 Upper Quartile 9 8 8 In addition to data on of past football games, this analysis also incorporates metropolitan population as a proxy for fan support. Therefore, population estimates for each year, 1993 to 2005, were obtained from the online Census database for each metropolitan area with an NFL team. Similarly, team revenue was collected as an alternative measure of fan support. This data was gathered from Forbes.com s annual valuation of NFL franchises.

Johnson 18 V. Data Analysis and Results This paper attempted to examine the efficiency of the NFL betting market from multiple angles. Most namely, the data was developed to update and test Gray and Gray s results. However, this data can also be used to conduct other widespread tests of efficiency, as well as incorporating new variables in the search for a systematic inefficiency that can be exploited. General Test of Market Efficiency The most general test of market efficiency in the NFL betting market is to see whether, on an aggregate basis, the actual spreads are highly correlated to the predicted spreads and do not show a systematic bias. This model takes the basic OLS form of: A i = b 0 + b 1 S i + ε i This OLS regression simply regresses the actual game spread (A) on the predicted spread (S) and estimates the values of b 0, a constant, and b 1, a coefficient. If the spreads truly are accurate forecasts of actual game spreads, b 0 should take on a value of zero as a positive (or negative) value would suggest actual spreads are consistently larger (or smaller) on an aggregate basis, and b 1 should take a value of one, in order to show perfect correlation to the actual spreads. Therefore, the hypotheses for this test of efficiency are: H 0 : b 0 = 0 and b 1 = 1 H A : b 0 0 or b 1 1 This model was run using the entire data set, including the games that were exactly predicted, and the results are presented in Table 3. From the results, neither condition of efficiency can be rejected and, therefore, the market appears to be efficient on an aggregate level. However, this does not necessarily imply systematic inefficiencies do not exist on a lower level and, therefore, this paper will continue to examine other existing models and variables for exploitable trends.

Johnson 19 Table 3 Basic test of market efficiency estimating A i = b 0 + b 1 S i + ε i using the complete sample of 3,176 games from 1993 to 2005. Coefficient b 0 b 1 Estimate 0.52 (0.23) 1.03 (0.04) Gray and Gray s Probit Models The goal of probit models, like those developed by Gray and Gray, are to take exogenous variables and estimate the probability of a binary variable taking the desired value. The first model developed by Gray and Gray is a simple probit model of the form: Win i = b 0 + b 1 Home i + b 2 Fav i + ε i The explanatory variables, Home and Fav, are dummy variables that take a value of one if the team of record is the home team or the favorite, respectively, for a given observation. Similarly, as the model is designed to predict the probability of a win against the spread and is not concerned with margin of victory, the dependent dummy variable is one if the team of record beats the spread and zero if it does not. This model was run using the sample of 3,090 games and the results are presented in Table 4. From the probit regression, the coefficient on Home is positive and the coefficient on Fav is negative, meaning a home team is more likely to beat the spread and a favorite is less likely to win against the spread. According to these results, a home underdog is the most likely to beat

Johnson 20 Table 4 Simple Probit test of market efficiency estimating Win i = b 0 + b 1 Home i + b 2 Fav i + ε i using the sample of 3,090 games from 1993 to 2005. Coefficient b 0 b 1 b 2 Estimate 0.02 (0.04) 0.03 (0.05) -0.09 (0.05) the spread and an away favorite is most likely to lose. Although Gray and Gray conclude from their model the coefficient on the Fav variable is significant, and the two variables are jointly significant, the coefficients resulting from the updated data are neither individually, nor jointly, significant even though they share the same signs as those generated by Gray and Gray s regression. In addition, while Gray and Gray s model correctly predicts 52.51% of in-sample games, which is enough to earn a slim profit, the updated model only predicts 51.52% of insample games correctly, which would earn a loss once the bookie s commission is taken into account. Although Gray and Gray conclude the simple probit model exhibits systematic inefficiencies, they also develop what they term an augmented probit model to incorporate additional variables and increase prediction accuracy. As previously mentioned, this model attempts to model commonly speculated stock market inefficiencies that good stocks with poor recent performance are undervalued and bad stocks with good recent performance are overvalued. In other words, momentum is overvalued in bettors minds and outweighs the actual strength of a team. In order to test this, Gray and Gray s model takes the following form:

Johnson 21 Win i = b 0 + b 1 HWP i + b 2 AWP i + b 3 HWP4 i + b 4 AWP4 i + b 5 HFav i + ε i Here, the home team is always the team of record, so the dependent binomial variable, Win, is defined as it was previously except it equals one of the home team beats the spread and zero if it does not. The HWP and AWP variables are the season-long winning percentages of the home and away teams, respectively, which are intended to serve as proxy variables for a team s actual strength. On the other hand, HWP4 and AWP4 represent the home and away teams winning percentages in the past four games, respectively, which are intended to represent what type of momentum a team is exhibiting. Because HWP4 and AWP4 rely on each team s results from the previous four games, the first four games of each season were eliminated, thereby cutting the sample to 2,321 games. The probit model was run using this sample and the results are presented in Table 5. Although Gray and Gray find b 2, b 4, and b 5 are individually significant, the results from the probit estimation using updated data only shows b 2 is individually significant. In addition, Gray and Gray find HWP and AWP are jointly significant, and the same is true for HWP4 and AWP4. In this sample, the likelihood ratio test (LRT) statistic for HWP and AWP s joint significance is 12.11 (χ 2 2 under the null), which is significant at any reasonable level, and the LRT statistic for HWP4 and AWP4 s joint significance is 5.33 (χ 2 2 under the null), which is significant at the 90 percent level; for purposes of comparison, Gray and Gray s LRT statistics are 6.14 and 12.27, respectively. While this estimation seems to imply less significant results, the most interesting aspect is the coefficients b 1, b 2, b 3, and b 4 have opposite signs as those estimated by Gray and Gray. The coefficients on HWP and AWP are negative and positive, respectively. This can be interpreted that a team that has performed well over the course of a season is overestimated when at home and underestimated when on the road perhaps implying

Johnson 22 Table 5 Augmented Probit test of market efficiency estimating Win i = b 0 + b 1 HWP i + b 2 AWP i + b 3 HWP4 i + b 4 AWP4 i + b 5 HFav i + ε i using the sample, excluding the first four games, of 2,321 games from 1993 to 2005. Coefficient b 0 b 1 b 2 b 3 b 4 b 5 Estimate -0.08 (0.12) -0.26 (0.22) 0.71 (0.22) 0.14 (0.14) -0.28 (0.14) -0.08 (0.06) home field advantage is given too much weight. Conversely, the coefficients on HWP4 and AWP4 are positive and negative, respectively, which can be interpreted that the market is actually slow to react to recent performance and resists changing its opinion on a team s strength despite recent evidence to the contrary. Therefore, these results contradict Gray and Gray s; however, they also have reasonable explanations. While one can infer from Gray and Gray s results that the market lacks long-term memory and over-reacts to recent events, these results suggest the market actually clings to its beliefs about a team s strength even though a team may develop or deteriorate over the course of a season. Another way to compare Gray and Gray s model with the model generated by updated data is by examining the actual success rates generated by the models. In their paper, Gray and Gray show their model predicts the correct winner 54.6 percent of the time which provides an

Johnson 23 excess return of 2.2 percent after accounting for the bookie s vigorish. However, as expected given the lower values of significance, the updated model only predicts 53.8 percent of games correctly an excess return of 1.4 percent. Out-of-Sample Results While both Gray and Gray s model and my model show results that indicate it would have been possible to systematically earn a profit over the time span of the data samples, this, in itself, does not imply a bettor can use these models to earn a profit in the future. In other words, the predictive value of out-of-sample results is much more important than the success rate for insample results. With this in mind, Gray and Gray conduct a simple test of out-of-sample success. They use their data from 1976 to 1992 to estimate the augmented model and apply the results to games from 1993 and 1994. In this test, Gray and Gray find their model predicts 56.0 percent of games correctly an excess return of 3.6 percent, which is even greater than their in-sample results. In order to similarly test the updated results, I estimated the model using ten years of data and two time spans 1993 to 2002 and 1994 to 2003 and applied the estimates, as applicable, to three sample periods 2003 to 2004, 2004 to 2005, and 2003 to 2005. The results of the success rates are shown in Table 6. Unlike Gray and Gray s, these out-of-sample tests show success rates of less than 50 percent in two cases, and the prediction of 2003 and 2004 results was successful 52.5 percent of the time which is only marginally, and not statistically, greater than the breakeven point of 52.4 percent, and is not significantly different from 50 percent when tested under a binomial distribution. Therefore, I cannot conclude the model accurately predicts out of sample results.

Johnson 24 Table 6 Success of out-of-sample predictions using games from the specified years to predict outcomes of games in the specified years. Years Predicted Estimated Using Bets Wins Win % 2004-2005 1994-2003 377 178 47.21% 2003-2004 1993-2002 375 197 52.53% 2003-2005 1993-2002 562 277 49.29% Dynamic Predictive Models Due to the contradiction between Gray and Gray s out-of-sample tests and mine, I was concerned only running the tests using one or two estimations and relatively small sample sizes may not be an accurate representation of my model s success. In addition, as market sentiment may change over time, inefficiencies may shift or change as was witnessed with the signs of coefficients in the augmented probit results. For example, if the market develops an opinion that home favorites are generally too big of favorites (i.e. the coefficient is negative), money will shift to the away teams. However, given data on these subjects is not readily available, there may be an information lag and the lines will shift too much in the direction of road underdogs (i.e. the coefficient on HFav would be positive). Therefore, I analyzed the data set using a variety of dynamic models which took data from the past one, three, five, and ten years to predict game results for the next season. In addition to varying the number of seasons the model was estimated using, I also varied the cut-point for placing bets. In other words, using a cut-point of c would mean the bettor would place a bet on the home team if the model assigned a probability of at least c to the home team winning, and would bet on the away team if the model predicted

Johnson 25 the home team would win with a probability less than 1-c. This system of models was run and the results are presented in Table 7. An analysis of the total number of bets placed and bets won under each strategy provides some interesting results. First, the sample using three years to estimate the models provides the best results, and the sample using ten years of data provides the worst. In addition, the sample using three years is the only sample that significantly predicts more than 50 percent of games correctly. However, upon further examination, this trend might be attributed to the fact that there are only three years of data to predict using the ten year models and 2005 seems to have been a difficult year to predict across the board. In fact, when looking at all of the models predictive value over the 2003 to 2005 time span, there are a few data points slightly above 50 percent and the rest are below. Therefore, no conclusive results can be reached on the optimal number of years results to use to estimate the model.

DRAFT Johnson 26 Table 7 Summary of dynamic modeling success rates using varying cut-point strategies. Cut-Point 0.50 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 2,152 1,098 51.02% 0.166 0.896 48.30% 3 10 1,808 948 52.43% 0.018 0.479 49.55% 5 8 1,459 726 49.76% 0.562 0.977 47.41% 10 3 559 275 49.19% 0.632 0.930 49.19% Cut-Point 0.55 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 1,961 1,001 51.05% 0.171 0.881 48.13% 3 10 1,647 855 51.91% 0.057 0.645 48.91% 5 8 1,325 673 50.79% 0.273 0.874 48.59% 10 3 509 251 49.31% 0.605 0.911 49.31% Cut-Point 0.60 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 1,762 903 51.25% 0.142 0.827 48.04% 3 10 1,489 778 52.25% 0.039 0.536 49.13% 5 8 1,176 600 51.02% 0.233 0.821 49.11% 10 3 458 226 49.34% 0.592 0.897 49.34%

DRAFT Johnson 27 Table 7 (cont.) Summary of dynamic modeling success rates using varying cut-point strategies. Cut-Point 0.65 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 1,557 807 51.83% 0.071 0.665 48.29% 3 10 1,293 667 51.59% 0.121 0.712 48.21% 5 8 1,045 531 50.81% 0.289 0.840 49.63% 10 3 408 205 50.25% 0.441 0.795 50.25% Cut-Point 0.70 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 1,327 684 51.54% 0.124 0.725 47.67% 3 10 1,102 571 51.81% 0.108 0.640 48.65% 5 8 892 455 51.01% 0.262 0.788 49.14% 10 3 358 178 49.72% 0.521 0.832 49.72% Cut-Point 0.75 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 1,119 584 52.19% 0.067 0.544 46.29% 3 10 912 470 51.54% 0.168 0.688 47.87% 5 8 749 380 50.73% 0.331 0.810 49.31% 10 3 312 157 50.32% 0.433 0.751 50.32%

DRAFT Johnson 28 Table 7 (cont.) Summary of dynamic modeling success rates using varying cut-point strategies. Cut-Point 0.80 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 890 474 53.26% 0.024 0.293 49.77% 3 10 737 382 51.83% 0.151 0.607 46.93% 5 8 607 317 52.22% 0.128 0.519 49.57% 10 3 256 125 48.83% 0.623 0.860 48.83% Cut-Point 0.85 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 642 346 53.89% 0.022 0.213 48.70% 3 10 562 303 53.91% 0.029 0.223 49.13% 5 8 435 233 53.56% 0.062 0.297 49.70% 10 3 196 95 48.47% 0.639 0.849 48.47% Cut-Point 0.90 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 421 230 54.63% 0.026 0.167 52.69% 3 10 372 211 56.72% 0.004 0.042 49.07% 5 8 278 158 56.83% 0.010 0.061 52.25% 10 3 116 60 51.72% 0.321 0.522 51.72%

DRAFT Johnson 29 Table 7 (cont.) Summary of dynamic modeling success rates using varying cut-point strategies. Cut-Point 0.95 Number of Years of Data for Regression Number of Years Predicted by Model Total Bets Placed Total Bets Won Success Rate P -value (>50%) P -value (>52.4%) 2003-2005 Success Rate 1 12 199 111 55.78% 0.044 0.153 51.22% 3 10 158 93 58.86% 0.010 0.044 45.65% 5 8 112 66 58.93% 0.023 0.069 53.33% 10 3 43 26 60.47% 0.063 0.112 60.47%

Johnson 30 Another interesting trend in the data pertains to using different cut-point strategies. Gray and Gray examine using cut-point strategies near.5 and do not reach conclusive results. Therefore, I examined similar cut-points, as well as the results when more extreme values are used, and the results are found in the same tables. As the tables show, for the various models, the success rate stays fairly constant across cut-points for most reasonable values. In fact, it is not until we used.85 and above that we start to observe changes in the success rate, and success rates that are significantly greater than 50 percent. Furthermore, using cut-points of.90 and.95 actually provide success rates that are significantly greater than the break-even point of 52.4 percent in the samples that rely on three and five year results to estimate the models (significant at a 5 percent level for three years and a 7.5 percent level for five years). However, observing an extreme cut-point strategy drastically limits the number of bets a gambler can place each year, which would inevitably increase the variance in returns the better would realize. Therefore, it is possible these profits can be seen as compensation for this additional risk incurred by using an extreme cut-point. The Effects of Team Popularity on Probit Models Another commonly speculated area of inefficiency in the betting market is the issue of team popularity. The belief is teams with a larger fan base will have a disproportionate amount of money on them as bettors tend to bet on their own team. The increased amount of money on a team would have the effect of moving the line in that team s direction, thereby making them either a greater favorite or lesser underdog. If this is the case, it would be beneficial to bet on the team that has less fan support in order to capture the inefficiency. For the purposes of testing this theory, this paper attempts to use the two variables previously described in Section IV: population and team revenue.