Beating the bookies with their own numbers - and how the

online sports betting market is rigged

Lisandro Kaunitz , Shenjun Zhong3* and Javier Kreiner4*
1,2*

1- Research Center for Advanced Science and Technology, The University of Tokyo, Tokyo,

Japan.

2- Progress Technologies Inc., Tokyo, Japan.

3- Monash Biomedical Imaging Center, Monash University, Melbourne, Australia.

4- Data Science department, CargoX, Sao Paulo, Brazil

Corresponding authors: Lisandro Kaunitz / Javier Kreiner
E-mail: [email protected] / [email protected]

 Abstract

The onlinesportsgamblingindustryemploysteamsofdataanalyststobuildforecastmodelsthat
turn the odds at sports games in their favour. While several betting strategies have been
proposed to beat bookmakers, from expert prediction models and arbitrage strategies to odds
bias exploitation, their returns have been inconsistent and it remains to be shown thatabetting
strategy can outperform theonlinesportsbettingmarket.Wedesignedastrategytobeatfootball
bookmakers with their own numbers. Instead of building a forecasting model to compete with
bookmakers predictions, we exploited the probability information implicit in the odds publicly
available in the marketplacetofindbetswithmispricedodds.Ourstrategyprovedprofitableina
10-year historical simulation using closing odds, a 6-month historicalsimulationusingminuteto
minute odds, and a 5-month period during which we staked real money with the bookmakers1.
Our results demonstrate that the football betting market is inefficient bookmakers can be
consistently beaten across thousands of games in both simulated environments and real-life
betting. We provide a detailed description of our betting experience to illustrate how thesports
gambling industry compensates these market inefficiencies with discriminatory practices against
successful clients.

1
Code, data and models are publicly available: https://github.com/Lisandro79/BeatTheBookie
1
Introduction
In the midst of chaos, there is also opportunity.

Sun Tzu, The Art of War

In recent years, the emergence of web technologies, product platforms andTVbroadcastrights

transformed the online gambling industry into a worldwide $452billionbusiness2.Clientsofthe

online sports betting industry dream of beating the bookies and, most often, find in the

adrenaline and excitement of their risky gambling activities an escape from the boredom of

everyday life (Blaszczynski, McConaghy, & Frankova, 1990; Lee, Chae, Lee, & Kim, 2007;

Loroz, 2004; Platz & Millar, 2001). To maximize profit, bookmakers employ teams of data

scientists to analyzedecadesofsportsdataanddevelophighlyaccuratemodelsforpredictingthe

outcome of sports events (Cantinotti, Ladouceur, & Jacques, 2004; Garca, Prez, &Rodrguez,

2016). Although several strategies have been proposed to compete with bookmakers models,

from expert predictions (Forrest, Goddard, & Simmons, 2005), probability models based on

Power scores, Elo ratings and/or Maher-Poisson approaches (Dixon & Coles, 1997; Maher,

1982; Vlastakis, Dotsis, & Markellos, 2008) and prediction markets (Spann & Skiera, 2009) to

arbitrage strategies and odds bias exploitation (Ashiya, 2015; A. C. Constantinou, Fenton, &

Neil, 2013; A. Constantinou & Fenton, 2013; Franck, Verbeek, & Nesch, 2009), to our

knowledge there is no precedent in the scientific literature thattheyconsistentlyoutperformthe

market and show sustained profit over years and across football leagues around the world (A.

Constantinou & Fenton, 2013; Deschamps & Gergaud, 2012; Kain & Logan, 2014; Spann &

Skiera, 2009; Vlastakis et al., 2008; Vlastakis, Dotsis, & Markellos, 2009).

2
www.statista.com
2


Can a betting strategy outperform the sports betting market? Although bookmakers profitable

business (along with theirmodellingadvantageandcontrolofoddspricing)seemstosuggestthe

opposite, we aimed to demonstrate thatbookmakerscanbebeatenwiththeirownnumbers.We

developed a betting strategy for the football market that exploited the implicit information

contained in the bookmakers aggregate odds (Kuypers 2000; Cortis 2016; Cortis, Hales, and

Bezzina 2013) to systematically take advantage of mispriced events. Our bettingsystemdiffered

from previous betting strategies in that, instead of trying to build a model to compete with

bookmakers forecasting expertise, we used their publicly available odds as a proxy of the true

probability of agameoutcome.Withtheseproxieswesearchedformispricingopportunities,i.e.,

games with odds offered above the estimated fair value (see glossary in Box 1). Our strategy

returned sustained profits over years of simulated betting with historical data, and months of

paper trading and betting with actual money. These results suggest that the football betting

market is inefficient. Bookmakers, however, deploy aspecialsetofpracticalrulestocompensate

for these inefficiencies. A few weeks after we started trading with actual money some

bookmakers began to severely limit our accounts, forcing us to stop our betting strategy. We

thus demonstrate that (i) bookmakers can be beaten consistently over months/years of betting

with a single strategy, in both simulated environments and in real-life betting situations and (ii)

the online sports betting system is rigged against successful bettors through discriminatory

practices.

Methods

Who wishes to fight must first count the cost.

Sun Tzu, The Art of War

Betting strategy

For a bet to be fair, i.e., for the expected value of a bet to be zero, the odds paid by the

bookmaker must be the inverse of the underlying probability of the result. Once bookmakers

build an accurate model that estimates the underlying probability of the result of a game, they

offer odds that are below the fair value. The mechanism operates similarly totherouletteatthe

casino. For example, when a customer places a bet on red in an American roulette, there is a

18/38 chance of doubling the wager (18 green numbers, 18 red numbers, plus 0 and 00, which

are green). Under these conditions, the fair value for the bet is 2.111 but the house pays only2

and, therefore, the house pays below fair value. This is the tax or commission charged by the

bookmaker, in this case, for every dollar bet at the roulette, the house expects to earn $(2/38),

or 5.3c.

In order to calculate the odds that, statistically, will allow bookmakers to earn a desired

percentage of the total money bet at sport games, they need accurate models to estimate the

probability of eachevent.Therearemanydifferentfactorsthatcanbeincorporatedintoamodel

to predict the probabilityoftheoutcomeofafootballgame,forinstance:theresultsofthelastn

games for the two teams, the record of successful games at home or away for those teams, the

number of goals scored and conceded by each team during the previous games, player injuries

before the game and even the expected weather conditions on the day of the match (Dixon &

4
Coles, 1997; Langseth, 2013; Maher, 1982). If we consider the scope of these variables the task

of developing accurate models to predict the outcome of thousands of games across football

leagues around the world becomes an extremely complex challenge. In recent years, however,

teams of professional analysts have improved the outcomes of their prediction models with

increasingly sophisticated statistical analysis and large amounts of data in variety of forms

(Gandar, Zuber, & Lamb, 2001).

To quantify the predictive power of bookmakers models we conducted a historical analysis of

football game outcomes.WecollectedthehistoricalclosingoddsoffootballgamesfromJanuary

2005 to June 2015 from online sports portalsavailableontheInternet.Forthisanalysisweused

data from 32 bookmakers: 'Interwetten', 'bwin', 'bet-at-home', 'Unibet', 'Stan James', 'Expekt',

'10Bet', 'William Hill', 'bet365', 'Pinnacle Sports', 'DOXXbet', 'Betsafe', 'Betway', '888sport',

'Ladbrokes', 'Betclic', 'Sportingbet', 'myBet','Betsson','188BET','Jetbull','PaddyPower','Tipico',

'Coral', 'SBOBET', 'BetVictor', '12BET', 'Titanbet', 'Youwin', 'ComeOn', 'Betadonis', and

'Betfair'. In total, we analyzed 479,440 games from 818 leagues and divisions across the world.

To measure the accuracy of bookmakers in estimating a games final result probabilities we

calculated the consensus probability as follows:

pcons = 1/ mean() (Eq. 1)

where is a set containing the odds across bookmakers for a given event and a given game

result (home team win, draw, away team win), whenever there were more than 3 odds available

forthatresult(insomegamesonlyasubsetofbookmakersofferedodds;thenumberofoddswe

employed for analysis varied from a minimum of 3 to a maximum of 32). In this way, we

calculated the consensus probability of a home team win, a draw, or an awayteamwinforeach

5
of the 479,440 games (in total 3 x 479,440 consensus probabilities). Then, we binned the data

according to the consensus probability from 0 to 1 instepsof0.0125(i.e.,80bins).Withineach

bin we calculated: 1) the mean consensus probabilities across games at closing time (the final

odds provided by bookmakers before the start of the match); and 2) the mean accuracy in the

predictionofthefootballgameresult(i.e.,theproportionofgamesendinginhometeamvictory,

draworawayteamvictoryforthatbin;seeFigure1).Weusedaminimumof100gamesforeach

bin. With these data we ran a preliminaryanalysisandobservedthattheconsensusprobabilityis

a good predictor of the underlying probability of an outcome (see Results section). Based on

these results, we decided to build our betting strategy on this evidence that bookmakersalready

possess highly accurate models to predict the results of football games.

A strategy intendedtobeatthebookmakersatpredictingtheoutcomeofsportsgamesrequiresa

more accurate model than the ones bookmakers have developed over many years of data

collection and analysis. Instead of trying to create such a model, we decided to use the

bookmakers own probability estimates of the outcomes to find mispricing opportunities. More

specifically, we searched for opportunities where some odds offered were abovetheirestimated

fair value. Sometimesbookmakersofferoddsabovefairvalueeithertocompetetoattractclients

or to maintain a balanced book to avoid getting overly exposed to risk. For example, when too

many clients bet on an outcome (e.g.hometeamvictory)bookmakerscanincreasetheoddsfor

the corresponding counterpart ( e.g. away teamvictory),inordertoattractmoregamblerstobet

on it and decrease their exposure to the overbooked outcome. This means that bookmakers

might offer odds with a lower implied probability than the actual probability of a result. This is

the key factor that we exploited in our strategy.

6
We based our strategy on the estimated payoff of each bet. The expected payoff of betting $1 is:

E () = preal (outcome materializes) payof f (outcome materializes)

+ P real (outcome does not materialize) payof f (outcome does not materialize)

(Eq. 2)

E () = preal ( 1) + (1 preal ) ( 1) = preal 1 (Eq. 3)

Where is the payoff of the bet (a random variable), preal istheactualunderlyingprobability

that the outcome materializes, and are the odds paid by the bookmaker in case that the

outcome comes about.

We performed a preliminary data analysis and found that:

preal pcons (Eq. 4)

Where pcons is the consensus probability as calculated above and is an adjustment term that

allows us toincludetheinterceptweestimatedinaregressionanalysisonoutcomesofgamesfor

Home, Draw and Away. The estimated was 0.034, 0.057 and 0.037 forhomevictory,

draw and away victory, respectively (see results section).

Then:

E () (pcons ) 1 (Eq. 5)

Under these conditions, we should place a bet when the expected payoff is greater than 0, i.e.,

when:

> 1/ (pcons ) (Eq. 6)

We followed this line of reasoning to define our betting strategy, and decided to place a bet

whenever the maximum odds offered for a given result fulfilled the following inequality:

max() > 1 / (pcons 0.05) (Eq. 7)

The expected value of each bet increases with the parameter, while the number of games

available for betting decreases. This occurs because the condition becomes more stringent and

less bookmakers offer odds with such high margins. To select an appropriate value for the

parameter we analyzed the performance of the simulation strategy by varying the value of

from 0.01 to 0.1. We found that an of 0.05 produced the optimal payoff with the largest

amount of games (an of 0.06, forexample,wasequallyprofitablebutwedecidedtousean

of 0.05 because it provided twice as many games to bet in, which might be useful in a strategy

that increases the amount staked as a function of the earnings).

In summary, we based our betting strategy on the assumption that odds published by

bookmakers allow us to obtain a highly accurate estimate of the actual probability of the

outcome of an event (by taking the inverse of the mean odds across bookmakers minus a

constant that allows for the bookmakers commission). Thus, our betting strategy consisted of

placing bets whenever theoddsofferedbysomebookmakerdeviatedfromtheaverageandwere

above fair value, i.e., when the expected payoff of placing the betwaspositive.Importantly,the

task of identifying the odds that satisfied the threshold in (Eq. 7) did not require a model with

higher accuracy than the bookmakers models.

Strategy implementation

Our betting strategy was implemented as a real time system, and deployed on a virtualmachine

hosted on the cloud. The system continuously collected data from online sports betting portals

and providedthewebservicethatmadeadashboardavailable,wheretherecommendedbetsand

the betting history were shown (SupplementaryFigure1).Thevirtualmachinewasusedtoruna

program that searched for the odds of everyfootballgamefrom5hoursbeforetheonsetofthe

game. For each game, the program continuously collected odds across 32 bookmakers and

calculatedwhetherthemaximumofferedoddscompliedwithourstrategysconditionforplacing

a bet - i.e., maximum odds fulfilling, Eq. (7). Whenever the program found a situationinwhich

this happened, it displayed the information about the game, bookmaker and odds on the

dashboard (Supplementary Figure 1), so that the users (including us) could see the list of bets

recommended by the system and place a bet of fixed amount with the bookmaker.Tokeepthe

amount of money placed on each independent game constant,onceabetwasplacedforagame

at some bookmaker, that game was not considered for further analysis.

Results
Victorious warriors win first and then go to war . The greatest victory is that which requires no battle.

Sun Tzu, The Art of War

Analysis to define the betting strategy

To select the appropriate strategy we first performed a descriptive statistical analysis of the

relationship between the bookmakers predictions and the actual probability of the outcome of

football games. A linear regression analysis showed a strong correlation between the

9
bookmakers consensus probability and the results of the game for home victory (R2= 0.999),

draw (R2= 0.995) and away victory (R2= 0.998). The slopes and intercepts oftheregressionline

were 1.003and-0.034forahomevictory,1.081and-0.057fordraw,and1.012and-0.037foran

away victory, respectively. These results suggest that the consensus probability is an extremely

accurate proxy (up to a constant intercept) of the actual probability ofoccurrenceofeachevent

(home victory, dray, away victory;notethattheslopesofthethreeregressionlinesareveryclose

to 1). Based on these results, we decided to build our betting strategy on the evidence that

bookmakers already possess highly accurate models to predict the results of football games.

Strategy Outcome

Wetestedourbettingstrategybyanalyzingtheoddsandresultsof479,440footballgamesplayed

in 818 leagues during a ten-year period, from 2005 to 2015. We began our analysis by applying

our betting strategy to the closing odds of each game (i.e., the odds values offered by

bookmakers at the start of the game3). We simulated placing bets when the closing odds of a

bookmaker complied with Eq. (7) at the closing time of the odds. With this approach, our

betting strategy reached anaccuracyof44.4%andyieldeda3.5%returnovertheanalysisperiod.

For example, for an imaginary stake of $50 per bet, this corresponds to an equivalent profit of

$98,865 across 56,435 bets (Table 1, Figure 2A). We performed a bootstrap analysis to assess

whether our returns were above chance level. We repeatedly simulated a strategy thatchoosesa

random sample of games and, for each game in the sample, randomly selects to bet for home,

draw or away (with a priorprobabilityequaltothatofourstrategy,seebelow)andplaceswagers

3
In practice, closing odds are a particular case of odds because they reflect the latest odds that were
available for clients to bet. We note, however, that these odds values are not a perfect estimate of a
system that could be used in real life because bookmakers close their odds shortly before the start of the
game.

10
at the maximum oddsofferedacrossbookmakers.IneachrunofthesimulationweA)randomly

sampled 56,435 games (the same amount of games that were selected by our betting strategy)

from the complete set of games in the historical series,B)selectedarandomoutcomeonwhich

to bet with a probability of 0.595 for home victory, 0.021 for draw and 0.384 for away victory

(these aretheproportionsofhome,drawandawaygamesthatwereselectedbyourstrategy)and

C) calculated thereturnofplacingthebet.Werepeatedtheprocedure2000times(samplingwith

replacement) to obtain a distribution of returns (Figure 2A). The random strategy yielded an

accuracy of 38.9%, an average return of -3.32% and an averagelossof$93,563(STD=$17,778),

further confirming that Eq. 7 successfully selects bets with a positive expected payoff above

chance level. The return of our strategy was 10.82standarddeviationsabovethemeanreturnof

the random betstrategy.Theprobabilityofobtainingareturngreaterthanorequalto$98,865in

56,435 bets using a random bet strategy is less than 1 in a billion.

We observed that the final accuracywashigherforourstrategy(44.4%)thanfortherandombet

strategy (38.9%). Correspondingly, our strategy selected odds with a mean value of 2.30

(STD=0.99) and the random bet strategy selected odds with a mean valueof3.10(STD=2.42).

The discrepancy in the accuracy between strategies originated from the selection of events: our

strategy picked up games with lower odds values and higher probability of occurrence than the

games selected by the random bet strategy. We confirmed this finding with an analysis of the

mean closingoddsacrossbookmakersforeachstrategy.Asshownabove,themeanclosingodds

across bookmakers is a precise estimateofthetrueprobabilityofoccurrenceofanevent(Figure

1). The expected accuracy (as predicted by the inverse of the mean closing odds across

bookmakers) precisely estimates the final accuracy in each strategy. We calculated the expected

11
accuracy of the strategy using Eq. 4. For each bet of the strategy we calculate

i
pcons = 1/ mean() (Eq. 8)

where is equal to 0.034, 0.057 and 0.037 for home win, draw and away bets respectively (and

where the intercept comes from the regression analysis performed in the first paragraph of

the Results section), and i indexesthebet.Wethencalculatedtheaverageestimatedprobability:

i
E (accuracy) mean( pcons ). The expectedaccuracyforourstrategywas45.9%,andtheactual

accuracy was 44.4%, while the expected accuracy for the random betstrategywas38.9%andits

actual accuracy 38.9%. Although the mean closing odds values differed between strategies, the

final accuraciesofbothstrategiescloselymatchedtheexpectedaccuracyaccordingtoEq.8.This

confirms that the the probability information implicit in the mean closing odds across

bookmakersrepresentsapowerfulpredictorforthetrueoutcomeoffootballgames(asshownin

our historical analysis).

Following the success of our initial analysis, and considering that in real life individuals cannot

place betsattheclosingtimeofodds,wedecidedtoconductamorerealisticsimulationinwhich

we placed bets at odds available from 1 to 5 hours before the beginning of each game. To this

end, we wrote scripts to continuously collect odds from multiple sourcesontheInternet.While

the historical closing odds for football games can be easily retrieved online, we could not find

any source of data containing the time series of odds movements before the beginning of each

game. To obtain these times series we wrote a new set of scripts to gather information in real

time for upcoming games as they became available online. In total, we were able to obtain data

from 31,074 games, from the 1st of September 2015 to the 29th of February 2016.Usingthese

times series data, we placed bets according to our betting strategy at any time starting 5 hours

before the beginning of a game until 1 hour beforethestartofthegame.Underthesesimulated

12
conditions,ourstrategyselectedoddswithameanvalueof2.32(STD=0.99.),hadanaccuracyof

47.6% and yielded a 9.9% return; i.e., if every bet placed was $50 our strategy would have

generated $34,932inprofitacross6,994bets(Table1,Figure2B).Incontrast,thedistributionof

returns of the randombetstrategyselectedoddswithameanvalueof3.29(STD=2.96),hadand

accuracy of 38.4% and would have generated, for bets of $50, a return of 0.2% and an average

profit of $825 (STD=$7,106).Similarlyasshownabove,theexpectedaccuracywa46.5%forour

strategy and 37.7% for the random bet strategy, which closely matched the actual accuracies of

both strategies. The return of our strategy was 4.80 standard deviations above the mean of the

random bet strategy. The probability of obtaining a profit greater than or equal to $34,932 in

6,994 bets with a random bet strategy is less than 1 in a million.

Once we determined that ourbettingstrategywassuccessfulwiththehistoricalclosingoddsand

with the analysis of odds series movements from 5 hours to 1 hour before the game start, we

decided to test our betting strategy under more realistic betting conditions. To this end we

employed a technique called paper trading, a simulated trading process in which bettors can

practice placing bets without committing real money. We used the information displayed on

the dashboard to check the bookmakers accounts, verify that the possibility to lay a bet at the

advantageous odds was available, and subsequentlymarkthebetaslaidonthedashboard.Paper

trading allowed ustoempiricallycheckwhethertheoddswereavailableatthebookmakersatthe

time of placing a bet. We had to test the discrepancy between theinformationthatbookmakers

showed on theirwebsitesandtheinformationthatwasdisplayedonourdashboard.Often,there

was a time delay between the moment when bookmakers made their odds available online and

the time it took for our scripts to show that information on the dashboard. We observed that

around 30% of the odds that were displayed on the dashboard had already been changedatthe

13
bookmakers sites. The delay in the update of the odds created a sample bias in the games we

were betting on: in contrast to previous analysis in which every game was used for the

simulation, now a subset of these games was not included at the time of placing bets. To test

how this delay could affect our results, we ran again our strategy simulation, now randomly

discarding 30% of the games. We observed that, despite the missing bets, thestrategyremained

profitable. We decided to continue with our betting strategy, and after three months of paper

trading our strategy obtained an accuracy of 44.4% and a return of 5.5%, earning $1,128.50

across 407 bets for the case of $50 bets (Table 1, Figure 3).

At this point we decided to place bets with real money. All the procedureswereidenticaltothe

papertradingexercise,withtheexceptionthatthehumanoperatoractuallyplaced$50betsatthe

bookmakers online platforms after checking the odds on the dashboard.Ourfinalresultsshow

the profit we obtainedin5monthsofbettingmoneyforreal.Duringthatperiodweobtainedan

accuracy of 47.% and a profit of $957.50 across 265 bets, equivalent to a 8.5% return (Table 1,

Figure3).Combined,papertradingandrealbettinghadanaccuracyof45.5%andyieldedaprofit

of $2,086 in 672 bets, equivalent to a return of 6.2%. We compared the results of our strategy

with the results of a random bet strategy, identical to that employed for the time series odds

(figure 2B) but this time considering games from April 2015 to July 2015 (the period used for

paper trading and real betting). The random strategy yielded an accuracy of 38.7%, an average

return of -0.7% and an average loss of $670 (STD=$2047). The return of ourstrategyafter672

games was 1.34 standard deviations above the mean of the random bet strategy and the

probability of obtaining a profitof$2,086orhigherin672betswitharandombettingstrategyis

1 in 11. This probability corresponds to a p value of 0.089, under the null hypothesis that the

return of our strategy comes from a distribution of final returns obtained with a random bet

14
strategy. A p-value of 0.05 is often considered as the standard threshold for statistical

significance. The p-value we obtained from the analysis of the return of our strategy was

expected given the evolution of the returns obtained in our historical simulations: with an

increase in the number of games our strategy increases its return and separation from the

distribution of returns of the random bet strategy (as seen with the historical analysisofclosing

odds and odds movements series). The reader might noticethatduringasimilartimeperiodthe

simulated strategy bet onapproximatelytentimesmoregames.Thereasonforthisisthatwedid

not have adedicatedoperatorbettingonallavailableopportunities24hoursadayandasaresult

wemissedmanyofthebetsthatappearedonthedashboard.Nevertheless,ourpapertradingand

actual betting activity confirmed the profitability of the strategy.

Although we played according to the sports betting industry rules, afewmonthsafterwebegan

to place bets withactualmoneybookmakersstartedtoseverelylimitouraccounts.Wehadsome

of our bets limited in the stake amount we could lay and bookmakers sometimes required

manual inspection of our wagers beforeacceptingthem.Inmostcases,bookmakersdeniedus

the opportunity to bet or suggested a value lower than our fixed bet of $50 (Figure 4). Under

these circumstances we could not continue with our betting strategy. The limits imposed by

bookmakersnotonlyshrunkourpotentialprofitbutalsocreatedasamplingbiasinthechoiceof

games which was not taken into account in our previous analysis. In our simulations, when we

analyzed the effects of randomly discarding a proportion of the games, the returns were not

affected. However, the selection of games where bookmakers limited our stakes was unlikelyto

be purely random, which could negatively impact the strategys performance. Forthesereasons,

15
and because bookmakers restrictions turned the betting experience increasingly difficult, we

decided to end our betting experiment4.

Discussion

One may know how to conquer without being able to do it.

Sun Tzu, The Art of War

We developed a betting strategy for the online betting football market. In contrast to strategies

that build prediction models to compete with the forecastsofbookmakersmodels,ourstrategy

was developed under the assumption that theaverageoftheoddsacrossbookmakersreflectsan

accurate estimate of the probability of the outcome of a game. Instead of competing against

bookmakers forecasting models, we used the prediction information contained intheaggregate

odds to bet on mispriced events. Our strategy proved successful and returned profit with

historical data,papertradingandrealbettingovermonthsandacrossfootballleaguesaroundthe

world.

Betting strategies based onexpertortipsteranalysisattempttobeatbookmakersbyconstructing

more accurate forecasting models than those of bookmakers (Boulier, Stekler, & Amundson,

2006; A. Constantinou & Fenton, 2013; Daunhawer, Schoch, & Kosub, 2017; Deschamps &

Gergaud, 2012;Vlastakisetal.,2009).Ouranalysisshows,however,thattheimplicitinformation

containedintheaverageoddsacrossbookmakersprovidesahighlyaccuratemodeltopredictthe

outcomes of football games (Boulier et al., 2006; Forrest et al., 2005; Spann & Skiera, 2003).

4
As of the date of writing this paper (August 2017), one of the bookmakers we had accounts with,
Doxxbet, closed its website to clients. We are not able to withdraw the money (90 euro) from them.
Their support teams do not respond to our emails.

16
There are many cases where the aggregate predictions of a group of individuals produce more

accurate predictions than those of each individualseparately,aphenomenonoftenreferredtoas

the wisdom of crowds (Navajas, Niella, Garbulsky, Bahrami, & Sigman, 2017). This idea is often

applied in practice, for example in applications such as ensemble learning in machine learning

algorithms (Gron, 2017). Similarly, in the footballmarket,eachbookmakercanbeconsidereda

predictor,andtheaverageoddsastheaggregateinformationacrosspredictors.Thesepredictions

also include the preferences and opinions of the punters regarding the probability of the

outcome, because they exert pressure on the price of the odds through their collective betting

(bookmakers often alter odds based on demand level to keep a balanced book, e.g. when they

increase theoddsforafavouritewhenadisproportionateamountofpuntersplacemoneyonthe

underdog). As bookmakers already posses excellent predictive models to estimate theoutcomes

of football games, competing with them at forecasting game outcomes becomes a challenging

task. Not surprisingly, previous attempts to beat the football market with expert strategies

showed inconsistent returns (Boulier et al., 2006; A. Constantinou & Fenton, 2013; Daunhawer

et al., 2017; Deschamps & Gergaud, 2012; Forrest et al.,2005;Kain&Logan,2014;Vlastakiset

al., 2008, 2009). In comparison, our strategyshowedpositiveandsustainedreturnsoveryearsof

betting with historical data and over months of betting actual money across leagues in the

football market.

Through our experiments we demonstrated the existence of a betting strategy that consistently

generates profit. Some scholars consider that the existence of one such strategy is inconsistent

with the putative efficiency of the betting market (A. Constantinou & Fenton, 2013;

Deschamps & Gergaud, 2012; Vlastakis et al., 2009). If, on the contrary, a strategy like ours

generates profit consistently either by outperforming bookmakers predictions or by exploiting

17
market flaws then the betting market is necessarily inefficient. Our results suggest that the

online football betting market is inefficient because our strategy was able to obtain sustained

profits over years with historical data and over months of paper trading and actual betting. In

practice, however, the inefficiency of the football betting market was compensated by the

bookmakers restrictive practices. A few months after we began placing bets with real money

bookmakers limited our accounts, which forcedustostopourbettingcompletely.Althoughour

betting activities were legal and were conducted according to the bookmakers rules, our bet

stakes were nevertheless restricted. Our case illustrates some of the discriminatory practices of

the online sports betting market the sports betting industry has the freedom to publicize and

offer odds to their clients, but those clients are expected to lose and, iftheyaresuccessful,they

can be restricted from betting. In comparison, the limits to the accounts imposed in the online

gambling industry constitute illegal practices in other industries, or are evenunlawfulingeneral.

For example, advertising goods or services with intent not to sell them as advertised, or

advertising goods or services with no intent to supply reasonably expectable demand but with

the intention toluretheclienttobuyanotherproduct(apractice,oftencalledbaitorbaitand

switch advertising) is considered false advertising and carries pecuniary penalties in the UK,

Australia and the United States of America5. Most countries have laws regulating advertising in

the gambling industry, but some of these laws have been relaxed in recent years (e.g. the

Gambling Act 2005 in the UK allowed the sports gambling industry to start advertising online

and on TV) and they vary from country to country. Our study sets a precedent of the

discriminatory practices against successful bettors in the online sports gambling industry: the

5
Consumer Protection from Unfair Trading Regulations (2008) Guidance, Interim: Guidance on the UK Implementation
of the Unfair Commercial Practices Directive
(https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/284442/oft1008.pdf)
Australia: COMPETITION AND CONSUMER ACT 2010 - SCHEDULE 2
(http://www.austlii.edu.au/au/legis/cth/consol_act/caca2010265/sch2.html)
US: FTC Guides Against Bait Advertising, Section 238.
18
online football market is rigged because bookmakers discriminate against successful clients. We

advocate forgovernmentstotakeactionintofurtherregulatingthesportsbettingindustry,either

by forcing bookmakers to publicly admit that successful clients will be banned from betting or

by denying bookmakers the chance to discriminate against them.

Acknowledgements

We are very grateful to Katia Giacomozzi, who performed the bulkofthetaskofplacingactual

bets and dealing with bookmakers accounts. This work could not have been possible without

her hard work and enthusiastic contribution. We are also very grateful to Ben Fulcher, Adrian

Carter and Alessio Fracasso for their comments on an earlier version of the manuscript.

Lisandro Kaunitz wassupportedbyafellowshipfromtheJapaneseSocietyforthePromotionof

Science (JSPS P15048).

19
Figures and Tables

Table 1. Results obtained with historical data, paper trading conditions and real betting.

Dates Profit Return Number of Accuracy

(U$D) bets

Historical 01/01/2005 98,865 3.5 % 56,435 44.4%

Closing to
odds 30/06/2015

Continuous 01/09/2015 34,932 9.9 % 6,994 47.6%

odds to
29/02/2016

Paper 06/03/2016 1,128.5 5.5 % 407 44.5 %

trading to
19/04/2016

Actual 23/04/2016 957.5 8.5 % 265 47.2 %

betting to
18/09/2016

20


Figure 1. Bookmakers prediction power. A historical analysis of 10 years of football games

shows the tight relationship between the bookmakers predictions and the actual outcome of
football games. The probability estimated by bookmakers (as reflected by the inverse of the
mean closing odds across bookmakers) is highly correlated with the true probability of the
outcomes of football games for home team victory (black dots), draw (magenta dots) and away
team victory (blue dots). We analyzed 479,440 games from 818 leagues and divisions acrossthe
world during the period 2005-2015. Data was binned according to the estimated probabilities,
from 0 to 1 in steps of 0.0125, and with a minimum of 100 observations per bin.

21
A

22
Figures 2. Two analysis with historical data demonstrate the effectiveness of our betting
strategy. A- We applied our strategy to the closing oddsof479,440gamesandobtainedareturn
of 3.5% in 56,435 bets. To assess the probability of obtaining a return greater than or equal to
3.5% by chance we performed a bootstrap analysis to estimate the distribution of returns for a
Random Bet Strategy. By placing bets at the highest offered odds at random games the
Random bet Strategy yielded, onaverage,areturnof-3.32%.Incomparison,thereturnofour
strategy was 10.82 standard deviations above the mean of the distribution of returns of the
random bet strategy. The probability of obtaining a return greater than or equal to ours with a
randombetstrategyacross56,435gamesislessthan1inabillion.Datainthispanelcomesfrom
a 10-year database (2005-2015) of football games. The figure shows the potential total return
assuming a constant $50 stake per bet. B) We applied the same bootstrap analysis as in A), but
now to the time series ofoddsmovementsduringtheperiod[-5-1]hoursbeforethestartofthe
games. The random bet strategy yielded an averagereturnof0.2%.Incomparison,thereturnof
our strategy was 9.9%, 4.80 standard deviations abovethemeanofthedistributionofreturnsof
the random bet strategy. The probability of obtaining areturngreaterthanorequaltoourswith
arandomstrategythatbetsonthemaximumoddsacross31,074gamesislessthan1inamillion.
Data in this panel comes from a 6-month database of football games (September 2015 -March
2016).

Figure 3. Our betting strategy generated profit with paper trading and in a real-life betting
(placing real stakes with bookmakers). We obtained a return of 5.5% for paper trading (blue
23
line) and a return of 8.5% for real betting (see Table 1 for a detailed analysis) over a 5-month
period of betting. Considering bothpapertradingandrealbettingwemadeaprofitof$2,086 in
672 bets, a return of 6.2%. This was achieved by placing $50 on each bet.

Figure 4. Bookmakers discriminate against successful clients. Betting limits set on some of our
stakes. The figure shows examples ofsuchlimitsimposedonouraccountsbyfourBookmakers.
A- William Hill betting slip showing (www.sports.williamhill.com/bet/en-gb) a 2,428.33 yen
limit on our bet (at thetimethisbettookplace5000yenwereequivalentto$50).B-Interwetten
(www.interwetten.com) imposing a maximum bet of $11.11 C- Sportingbet
(http://www.sportingbet.com/) setting a maximum limit of $1.25 and D- Betway
(www.betway.com) limiting our stakes to $10.45.

24

References

Ashiya, M. (2015). Lock! Risk-Free Arbitrage in the Japanese Racetrack Betting Market. Journal of

Sports Economics, 16(3), 322330.

Blaszczynski, A., McConaghy, N., & Frankova, A. (1990). Boredom proneness in pathological

gambling. Psychological Reports, 67(1), 3542.

Boulier, B. L., Stekler, H. O., & Amundson, S. (2006). Testing the efficiency of the National

Football League betting market. Applied Economics, 38(3), 279284.

Cantinotti, M., Ladouceur, R., & Jacques, C. (2004). Sports betting: can gamblers beat

randomness? Psychology of Addictive Behaviors: Journal of the Society of Psychologists in Addictive

Behaviors, 18(2), 143147.

Constantinou, A. C., Fenton, N. E., & Neil, M. (2013). Profiting from an inefficient association

football gambling market: Prediction, risk and uncertainty using Bayesian networks.

Knowledge-Based Systems, 50, 6086.

Constantinou, A., & Fenton, N. (2013). Profiting from arbitrage and odds biases of the

European football gambling market. The Journal of Gambling Business and Economics, 7(2),

4170.

Daunhawer, I., Schoch, D., & Kosub, S. (2017). Biases in the Football Betting Market. Available

at SSRN: https://ssrn.com/abstract=2977118.

Deschamps, B., & Gergaud, O. (2012). Efficiency in betting markets: evidence from English

football. The Journal of Prediction Markets, 1(1), 6173.

Dixon, M. J., & Coles, S. G. (1997). Modelling Association Football Scores and Inefficiencies in

the Football Betting Market. Journal of the Royal Statistical Society. Series C, Applied Statistics,

46(2), 265280.

25
Forrest, D., Goddard, J., & Simmons, R. (2005). Odds-setters as forecasters: The case of English

football. International Journal of Forecasting, 21(3), 551564.

Franck, E., Verbeek, E., & Nesch, S. (2009). Inter-market arbitrage in sports betting (NCER

Working Paper Series, Sportometrics Program No. 48). National Centre for Econometric

Research.

Gandar, J. M., Zuber, R. A., & Lamb, R. P. (2001). The home field advantage revisited: a search

for the bias in other sports betting markets. Journal of Economics and Business, 53(4), 439453.

Garca, J., Prez, L., & Rodrguez, P. (2016). Forecasting football match results: Are the many

smarter than the few? Munich Personal RePEc Archive. Retrieved from

https://mpra.ub.uni-muenchen.de/69687/

Gron, A. (2017). Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and

Techniques to Build Intelligent Systems. OReilly Media, Inc.

Kain, K. J., & Logan, T. D. (2014). Are sports betting markets prediction markets? Evidence

from a new test. Journal of Sports Economics, 15(1), 4563.

Langseth, H. (2013). Beating the bookie: A look at statistical models for prediction of football

matches. In SCAI (pp. 165174).

Lee, H.-P., Chae, P. K., Lee, H.-S., & Kim, Y.-K. (2007). The five-factor gambling motivation

model. Psychiatry Research, 150(1), 2132.

Loroz, P. S. (2004). Golden-age gambling: Psychological benefits and self-concept dynamics in

aging consumers consumption experiences. Psychology & Marketing, 21(5), 323349.

Maher, M. J. (1982). Modelling association football scores. Statistica Neerlandica, 36(3), 109118.

Navajas, J., Niella, T., Garbulsky, G., Bahrami, B., & Sigman, M. (2017). Deliberation increases the

wisdom of crowds. arXiv [cs.SI]. Retrieved from http://arxiv.org/abs/1703.00045

Platz, L., & Millar, M. (2001). Gambling in the context of other recreation activity: A quantitative

26
 comparison of casual and pathological student gamblers. Journal Of Leisure Research, 33(4),

383.

Spann, M., & Skiera, B. (2003). Internet-Based Virtual Stock Markets for Business Forecasting.

Management Science, 49(10), 13101326.

Spann, M., & Skiera, B. (2009). Sports forecasting: a comparison of the forecast accuracy of

prediction markets, betting odds and tipsters. Journal of Forecasting, 28(1), 5572.

Vlastakis, N., Dotsis, G., & Markellos, R. N. (2008). Nonlinear modelling of European football

scores using support vector machines. Applied Economics, 40(1), 111118.

Vlastakis, N., Dotsis, G., & Markellos, R. N. (2009). How efficient is the European football

betting market? Evidence from arbitrage and trading strategies. Journal of Forecasting, 28(5),

426444.

27
Supplementary Materials

Box 1. Glossary

Bookmaker The bookmaker (or bookie), refers to the company that provides an
odds market for betting and offers to pay a price for each possible
outcome of a sporting event.

Event This term denotes a specific match between two teams or individuals.
For example: Ath. Bilbao vs Barcelona, Thursday, January 5, 21:15
GMT.

Market A betting market is a type of betting proposition with two or more

possible outcomes. The result of the match (home win, away win, or
draw), the number ofgoalsscored(twoorlessgoals,threeormore),or
the time of the first goal are a few examples of different markets fora
single sporting event.

Stake The amount of money wagered in a single bet.

Odds The odds of a result refer to the payoff to be received if the chosen
result materializes. In this paper we use the European notation, where
the odds are equal to the currency units to be received for each
currency unit wagered. For example, if an outcome offers odds of 2
means that for each dollar wagered the house will pay 2 back, giving a

profit of 1 dollar per dollar invested.

Fair odds Fair odds for an outcome are the ones that result in a zero expected
payoff. For example, if the probability of the outcome is , the fair
odds would be 2, because E(payoff) = x (2 - 1) + x (-1) = 0. In
general for the odds of an outcome to be fair, they should be the
inverse of the probability of the outcome.

28
 Result/Outcome The actual outcome of an event. E.g. in a 1x2 soccer bet, local win,
draw, and away win are the three possible results. If the result that
comes about coincides with the chosen result of a bet, the gambler

wins the odds times the stake, otherwise he loses the whole stake.

Profit The amount ofadditionalmoneythebettorreceivesontopofhisstake

if he chooses the result that actually happens. For example,iftheodds
are 2 and the stake is 10dollars,thegamblerreceives20dollarsintotal

from the bookie, and the profit is 10 dollars.

Yield A measure of the profitability of a series of bets, it is calculated as the

sum of the profits made from all thebetsplaceddividedbythesumof
the money staked in all bets, usually expressed as a percentage. For
example, if after 10 bets of $1 each there is a net profit of $1.50, the
yield is (1.5/10) = 0.15=15%.

29


Supplementary Figure 1.A)Screenshotoftheonlinedashboarddisplayingthegamesthatwereselected

for betting. The dashboard displayed the names of both teams, league name, bet value atwhichtoplace
the bet for the strategy toworkandthetimeremaininguntilthestartofthegame.B)Asecondtabinthe
Dashboard was used to keep track of the bets list. There the dashboard displayed the names of the
participating teams, football league, the result that was backed by the bet (1: home team to win, 2:away
team to win), final result of the game, odds value for the bet, thebookmakerthatwasusedforeachbet,
the amount of money placed on eachbet(weemployedU$D50throughout),theresultofthebet(1:bet
won; -1: bet lost) and the profit obtained from each bet. Some of the games used for paper trading are
displayed in this figure.

30