1

A statistical development of fixed odds

betting rules in soccer




Ian Milliner
1
, Paul White
1
and Don J. Webber
2

1
Department of Mathematics and Statistics, University of the West of England, Bristol, UK
2
Department of Economics, University of the West of England, Bristol, UK




Abstract
Two simple but seemingly profitable betting rules for betting on the away win in association
football are developed. One rule is consistent with avoiding those games in which there is a
clear favourite. The second rule is based directly on modelling bookmaker odds and
assessing the residuals under the fitted model. Contrary to previous research the betting rule
using the residuals suggests avoiding betting on those games where there are large
discrepancies between bookmaker odds and predicted-model odds.













Keywords: Fixed odds betting rules; away win; bookmakers probabilities

JEL Classification: D81

Correpsonding author: Don J Webber, Department of Economics, University of the West of
England, Bristol, UK. E-mail: [email protected]. Tel.: (+44/0) 117 328 2741. Fax:
(+44/0) 117 328 2289.
2
Betting on the outcome of UK association football matches (soccer matches) is big business
with in excess of 3% of the UK adult population regularly placing bets on the football fixed
odds market (Department for Culture, Media and Sport, 2007). It has also been the subject of
recent academic research (see, for example, Archontakis and Osborne, 2007). In the football
fixed odds market bookmakers offer betting odds for each of the three mutually exclusive and
exhaustive outcomes which are for the home team to win or the away team to win or for the
game to end in a draw. In the UK, it is customary practice to quote odds in the form a to
b for the home win, c to d for the away win, and e to f for the draw, where a, b, c,
d, e and f are integers. Thus, for instance, if a bettor wages b pounds on a home win and the
outcome is a home win then the total return would be b a + pounds (profit a pounds) but
otherwise would result in a loss to the bettor of b pounds.
The precise procedures with which bookmakers derive odds are best viewed as a
commercially guarded secret although it is widely known that the odds are essentially
judgement forecasts by panels of experts employed by bookmakers (Sharpe, 1997). The odds
offered are usually made available approximately one-week prior to the game taking place.
Although bookmakers always reserve the right to alter the odds on offer they seldom do so
irrespective of betting volumes and irrespective of new information that may come to light
during the course of the week (Forrest et al., 2005). Typically odds would only alter in the
rare event of incorrect odds being posted through a typographical error. It is in these senses
that the odds are considered fixed.
Bookmakers odds may be converted readily into decimal odds, whereby the
decimalised odds for the home win (
h
d ), away win (
a
d ) and draw (
d
d ) are given by
=
h
d
) ( b a
a
+
;
) ( d c
c
d
a
+
= ;
) ( f e
e
d
d
+
= respectively. The total of these decimalised odds,
d, will always exceed unity and (d 1) 100 , known as the over-round, reflects an
3
anticipated inbuilt profit for the bookmaker to offset cost of running the market assuming
liabilities are evenly spread over the three outcomes. In the UK fixed odds football market
the over-round is currently about 10% per game. The values
d
d
d
h
h B
=
,
;
d
d
d
a
a B
=
,
and
d
d
d
d
d B
=
,
may be thought of as estimated bookmaker probabilities of match outcomes. A
small but precious point noteworthy of further explanation is the terminology estimated
bookmaker probability. The estimated bookmakers probabilities are a simple multiplicative
rescaling of the decimal odds under the assumption that the over-round is spread
proportionally amongst the decimalised odds (which may not necessarily be the case). In
addition different bookmakers are at liberty to offer different odds giving rise to different
estimated probabilities. Bookmakers odds are set with commercial objectives in mind and as
such the derived probabilities may not reflect their best estimates of the probabilities that they
might otherwise derive. Figlewski (1979) and Knight (1965) make the important point that
in games such as Roulette the probability of winning are known in advance and is therefore a
game of risk but with no uncertainty. In the case of betting on a football match outcome the
bettor is presented with both risk and uncertainty since although each team can be thought of
as having a certain chance of winning the true chance will not be known.
There are large individual differences amongst people relating to gambling. Some
will not gamble. Some will be risk-positive being attracted by large odds with the attendant
bragging rights should a long shot pay dividends (Woodland, 1994). Avery and Chevalier
(1999) found evidence of sentimental betting such as betting on well-known teams or on
teams that have been covered in the media recently. Others may be attracted to betting on so-
called certainties irrespective of the odds on offer. Others may bet on the football team that
they support to win out of loyalty, or bet on their team to not win so as to have some financial
4
comfort should their team not win. The irrational behaviour of gamblers cannot be defined
on a simple continuum.
In contrast a prevalent theory amongst statisticians and econometricians is the idea of
developing betting strategies on value bets by systematically identifying football games in
which there has been a perceived incorrect setting of bookmakers odds through an inefficient
use of information available i.e. the market has failed to capture the information used by a
superior analyst Pankoff (1968, p.204). Our modelling approach partly revolves around the
prediction of bookmaker estimated probabilities.
A number of authors have operationalised the idea of a value bet by building
statistical models to predict the probabilities of match outcomes (say p
h
, p
a
, p
d
) and to
compare these estimated model probabilities with estimated bookmakers probabilities by
considering the ratios
h B
h
h
d
p
r
,
= ;
a B
a
a
d
p
r
,
= and
d B
d
d
d
p
r
,
= . Betting strategies then
take the form, for example, to only bet on a home win outcome if
h
r >
*
h
r where
*
h
r is some
determined constant chosen to balance risk and profitability. These strategies have been
successfully implemented by Dixon and Coles (1997) and by Rue and Salvesen (2000)
amongst others. In passing we comment that there are other ways of quantifying the degree
of mismatch.
The statistical models for predicting match outcomes either model the match outcome
directly using discrete choice models such as binary or ordinal logistic regression models or
through models predicting probabilities of match scores and then aggregating to predict
match outcomes. The Poisson distribution or negative Poisson distribution are invariably
used in this latter approach. In an early attempt to model match scores Reep et al. (1971)
concluded that it was difficult to predict match outcomes confidently through this route and
5
that chance dominates the game. Since then others have successfully used more
computationally intensive complicated models (e.g. Dixon and Coles, 1997).
It is our view that working directly with goals scored and conceded may not be the
correct stance. In league football a home win with a four-nil score line will be awarded the
same number of points as a two-nil home win. During a match a team winning by two clear
goals may change tactics or personnel and be content with that margin of victory rather than
aiming to beat the opposition by the greatest possible score line. In other cases a very strong
club may field a comparatively weakened team against some opposition with a view to
resting star players for future games at the expense of a comparatively low winning margin.
The primary aim for teams involved in a football match is the match outcome and (except in
very restricted circumstances) the extent of goal difference is secondary. The primary aim of
betting on the fixed odds market revolves around the match outcome and financial returns are
not linked to the margin of victory. In a comparative study Goddard (2005) found very little
difference in predictive ability between the discrete choice approach and models directly
modelling goals scored and conceded. There is an argument that modelling exact scores may
be more susceptible to the effect of outliers than a model based on match outcomes. For all of
these reasons we have opted to work directly with discrete choice models and to model the
match outcome directly.
Betting odds on the draw outcome invariably fall over a narrow range relative to the
range of odds on offer for the home win or the away win. Pope and Peel (1989) suggest that
this reflects a lack of ability of experts to forecast draws and Archontakis and Osborne (2007)
argue that this could simply reflect a general inability to predict draw outcomes with any
degree of reliability. Taken at face value this suggests that there may be inefficiencies in the
draw odds market. However the prediction of the draw outcome is notoriously difficult.
Prior to the development of fixed odds a common form of gambling amongst UK football
6
fans was the treble chance football pools whereby bettors would try to identify draws from
games to be played; monies staked would go into a pool and dividends paid on relative
performance. Although highly popular this form of gambling was viewed by most as
essentially a lottery (Forrest, 1999). Accordingly, although inefficiencies may exist in the
draw odds, it does not necessarily follow that they may be systematically exploitable and for
these reasons we will not consider the development of a betting strategy for the draw odds in
this note.
In league football there is a well established home advantage effect (Clarke and
Norman, 1995), and Dixon and Coles (1997) report that approximately 46 percent of games
in the English football leagues result in a home win. For these reasons it naturally follows
that the odds offered for a home win are typically lower than those on offer for an away win.
Consequently a betting strategy based on the away win outcome, although occurring less
often than one based on the home win outcome, may have the potential for greater profits
than a betting strategy for the home win. For these reasons our modelling strategy will focus
on the away win only.
Most of the published betting strategies in the statistical literature that have a positive
expected return are based on models with estimated team specific parameters that are
continually updated. This high dimensionally adversely impacts on the development of a
practical betting rule. For these reasons we consider a low dimensional parsimonious model
specification to underpin the betting rule.
Strategies based around betting on long shots or underdogs have been reported to be
troublesome. For instance Thaler and Ziemba (1988) report on the favourite - long shot bias
in racetrack betting whereby favourites tend to be under-backed and long shots tend to be
over-backed. Bird and McCrae (1987) reported the strategy of betting on favourites or on
long shots in horse racing would not yield a positive return. Likewise Woodland and
7
Woodland (1994) report that the favourite long shot bias is reversed in baseball betting and
that no simple favourite or long shot betting strategy would produce a positive return. It is
our contention that betting rules based on the form bet on the home win if
h
r >
*
h
r may be
particularly susceptible to the inclusion of too many matches where there is a clear favourite .
Our approach is based around a modification of this form of betting rule so as to allow the
possibility of avoiding placing bets on matches involving clearly identified long-shot /
favourite pairs.
Section 1 gives an overview of the methodological approach utilised to develop
betting rules for the away win. A brief account of the data used and the rationale for the
variables used in the model is given in Section 2. Derived statistical models and betting rules
are given in Section 3 and the utility of the models is discussed in Section 4.

1. Modelling Approach

Our modelling approach is based around sample data ( i = 1, ..., I ) for deriving a betting rule
and a second sample ( j = 1, ..., J ) for out of sample assessment of the efficacy of the rules
derived. Our first model is to directly estimate the probability of an away win using a
discrete choice model.
Let
i a
p
,
denote a model based estimated probability that a match indexed by i will
result in an away win, and let
i a B
d
, ,
denote the corresponding estimated probability derived
from the bookmakers fixed odds. Let
i a
r
,
=
i a B i a
d p
, , ,
/ . We will consider a betting rule of
the form bet on the away win for match i if, and only if,
l
r

i a
r
,


u
r where
l
r and
u
r

are
constants chosen on criteria such as maximum profit per game or maximum profit. This form
8
of rule is a more general version of the structure used by Dixon and Coles (1997) and Rue
and Salvesen (2000) which use a one-sided limit utilising
l
r

only.
One approach to determine optimum values for
l
r

and
u
r is to consider all possible
values for
l
r and
u
r

(
l
r

u
r ) and to apply these to sample values and choose the estimated
parameters to be those values that maximise within sample profit per game or maximise
within sample profit. This approach however may produce many small seemingly good
profitable intervals but which not may be replicated on unseen data (particularly so if profit
per game is considered). For this reason we consider determining
l
r and
u
r

separately. The
optimum value for
l
r is chosen so that the betting rule bet on the away win for game i if
i a
r
,


l
r results in maximum within sample profit. Similarly,
u
r

is chosen so that the betting
rule bet on the away win if
i a
r
,


u
r


results in maximum profit (and not maximum profit
per game). This approach is intended to produce a betting rule with greater relative
robustness which does not overly capitalise on chance idiosyncratic sample characteristics.
In addition we consider the development of a betting rule of the form bet on the
away win for match i if and only if
l
p
i a
p
,

u
p
. The rationale behind this rule is to
determine whether profitable rules can be developed whereby the away team is a strong
favourite (in which case the rule would default to betting on the away win if
i a
p
,

l
p ), or in
opposing a strong home team (in which case the rule would default to betting on the away
win if
i a
p
,
i

u
p
) or whether it is better to focus betting on the away win when there is no
seemingly clear favourite.
Our second approach is to use ordinary least squares regression with bookmakers
odds as the dependent variable. Residuals (
i
e ) under the model may be used to assess the
relative extent of disagreement between the bookmakers odds for the away win and the
9
predicted bookmakers odds under the model. Following the reasoning given above, we
consider a betting rule of the form bet on the away win for match i , if and only if,
l
e
i
e
u
e .

2. Sample data

Data was recorded on the 194 league football games that took places between the 2
nd
October
2007 and 22
nd
October 2007 from the games played in the top four English football leagues
and the top four Scottish football leagues. The outcome of each game was recorded (home
win, draw, away win) along with fixed odds for each outcome offered by Ladbrokes plc, the
UKs largest bookmaker.
Fixed odds are set with commercial and financial gains in mind and may not
necessarily reflect the best assessment of match outcomes since they may be set with
anticipated betting volumes in mind or indeed set to influence betting volumes. For these
reasons we consider as predictor variables the home and away team performance ratings
published weekly by the Racing and Football Outlook (RFO) which is a weekly newspaper
published by Trinity Mirror plc, dedicated to betting on horseracing and association football.
The RFO index is an index based on the results of the past 60,000 games and provides a form
rating on a scale of 0 to 1000 for each team in the English and Scottish football leagues.
Increasing ratings are intended to reflect increasing ability of a team and the difference in
RFO ratings between two teams is intended to reflect the extent of the degree of mismatch
between the two chosen teams. The RFO produces a separate index for home and away
performance to account for the home advantage effect and the extent of club specific home
advantage effect (the home effect cannot be considered to be of the same influence for all
10
teams). We therefore consider the RFO home rating for the home team and the RFO away
rating for the away team as predictor variables for match outcomes.
Our second approach is to use a good predictor of betting odds which utilises
information that might not be used by bookmakers in deriving odds. For this reason, for each
team in each game, we consider the average proportion of time that the team was winning,
irrespective of goal margin, in their previous three league games as a predictor variable. This
choice of predictor is partly informed by the ready availability of the data and partly informed
by the idea that the margin of victory is not of primary importance but that the percentage of
time winning in previous games will still provide an indication of relative dominance in
recent games against teams from the same league. We therefore consider average measure
of time winning in previous three games and RFO ratings as predictor variables of estimated
bookmaker probabilities.
A second data set comprising all of those matches held in the English and Scottish
divisions (63 games) between 15
th
January 2008 and 21
st
January 2008 was used to assess
independently the out of sample usefulness of the derived betting rules.

3. Derived betting rules

Table 1 summarises the discrete choice complementary log-log model for predicting the
probability of an away win. Overall the model is statistically significant (Log-likelihood chi-
square = 17.50, df = 2, p < 0.001), the individual predictors are statistically significant (p <
0.001) and the direction of effects for the RFO ratings for the home team at home (RFO HH)
and the RFO ratings for the away team playing away (RFO AA) make good conceptual sense.
The model adequately captures the structure in the data (percentage concordant pairs between
model predictions and outcomes is 66.6%) and goodness-of-fit tests using Pearsons residuals
11
(p = 0.246) and deviance residuals (p = 0.106) do not cast doubt on the appropriateness of the
model specification. Inspection of delta beta and delta deviance graphics indicate that model
does not suffer from the presence of overly influential observations. Prior to fitting this
model we did consider a simple logistic specification however application of Browns test
indicated that a model with a non-symmetric link function would be more appropriate.

{Table 1 about here}

Figure 1 is a plot of within sample profit against possible choices for
l
r for the betting
rule bet on the away win in match i if and only if
, a i
r
l
r with
, a i
r estimated for match i in
the data set using the complementary log-log regression equation and with a one pound bet
wagered each time the rule is fired. In this way the optimal value for
l
r was found to be
*
l
r =
1.596. In a similar way the value for the upper bound
*
u
r was determined to be 7.597, which
is the largest observed ratio in the data set. For the within sample data the rule bet on the
away win in match i if, and only if, 1.596
, a i
r 7.597 effectively defaults to bet on the
away win if
, a i
r 1.596 and fired 29 times yielding an absolute profit of 19.43 giving a
67% profit on monies staked. When applied to the test data, the rule fired on 13 occasions
giving an essentially break-even return of 0.80.

{Figure 1 about here}

Applying the same procedure but using the predicted probabilities for an away win
from the complementary log-log model gives the betting rule bet on match i to be an away
win if 0.4470
, a i
p 0.7146. This rule fired on 22 occasions and with 1 staked on each
12
game an overall profit of 16.34 was obtained (i.e. a 74% return). When applied to the test
data the rule fired on five occasions giving an overall percentage profit of 62.8%.
Table 2 summarises the fitted ordinary least squares model with the estimated
bookmaker odds for the away win as the dependent variable. The overall model is
statistically significant (
2
R = 41.9%, F(4, 195) = 34.04, MSE = 1.186, p < 0.001), each
predictor provides a unique statistically significant contribution to the model and the direction
of the effects in the model make good conceptual sense. The model does not suffer with
problems associated with multicolinearity (all variance inflation factors are less than 4). A
visual examination of the residuals under the model suggests that the assumption of
independence of errors has not been grossly violated although there is some evidence of a
small departure from normality (Kolmogorov-Smirnov test statistics for normality has a p-
value of 0.01). Adopting the same procedure as earlier, but using the residuals, gives a
betting rule of the form bet on the away win in match i if and only if 0.1489
i
e 0.2042
where
i
e is the residual for match i. Application of this rule to the sample data gives rise to
placing 28 bets yielding an overall profit of 9.75 (34.8% profit). Applying the rule to the
out of sample data gives a percentage profit of 19.8%.

{Table 2 about here}

4. Discussion and conclusions

The preceding analyses indicate that a profitable betting strategy based on gambling on the
away win may be possible. The results of the discrete choice model indicate that profitability
may be obtained by avoiding those matches where there is a large estimated probability of an
away win or a small estimated probability of an away win. This finding is consistent with
13
previous research cautioning against a betting strategy based on a long-shot or on a clear
favourite. Instead the derived rule suggests that it may be profitable to wager on the away
win outcome on those seemingly difficult to call matches. This may be a reasonable finding
if the extent of the home effect advantage has been incorrectly estimated by the bookmaker.
The results from the value bet approach which considers the ratio of model estimated
probability of the away win to the derived bookmaker probability of an away win as a betting
trigger seem to be less spectacular.
Distinct from other approaches we considered the direct modelling of bookmaker
odds using OLS regression. This analysis supported our prior reasoned hypothesis that
average time winning in previous games is associated with the odds on offer. Distinct from
other approaches we considered the residuals under the regression as quantifying the extent
of mismatch between the bookmaker odds for the away win and the model predicted odds.
The derived betting rule from this approach suggests avoiding betting on the away win when
there is a large discrepancy between predicted values and bookmaker values and this finding
is quite contrary to the usual stance of betting on so called value matches.
The results presented relate to league football only and due to the small sample size
should be treated with caution. However we have only fitted prior reasoned models and have
not undertaken a data dredging exercise which otherwise may have lead to too many false
findings. In deriving and assessing the betting rules we have simply placed a one-unit stake
per game. In practice it might be favourable to vary the stake in some optimal way (e.g.
betting stakes in proportion to perceived risk) and on this basis the percentage returns quoted
might be optimistically considered as an understatement. Likewise in practice a bettor will
be in a position to shop around the different bookmakers for best prices for the away win and
doing so would give a non-trivial positive impact on the percentage returns offered. We
chose to consider average winning times in the past three games a predictor variable although
14
there may be further merit in extending this predictor variable over a different number of
previous games.
A similar strategy could be considered for betting on the home win, however if
betting rules for both the home win and away win are to be considered then some additional
thought would have to be given to the possibility or prevention, of both rules firing on the
same game.

15
References

Archontakis, F. & Osborne, E. (2007). Playing it safe? A Fibonacci strategy for soccer
betting. Journal of Sports Economics, 8(3), 295-308.
Avery, C. & Chevalier, J. (1999). Identifying investor sentiment from price paths: The case
of football betting. Journal of Business, 72(4), 493-521.
Bird, R. & McCrae, M. (1987). Tests of the efficiency of racetrack betting using bookmaker
odds. Management Science, 33, 1552-1562.
Clarke, S. R & Norman, J. M. (1995). Home ground advantage of individual clubs in English
soccer. Statistician, 44, 509-521.
Department of Culture, Media & Sport (2007). Taking Part: The National Survey of Culture,
Leisure and Sport, Chapter 9, Gambling available from
http://www.culture.gov.uk/images/research/TPMay2007_9_Gambling.pdf
Dixon, M. J. & Coles, S. G. (1997). Modelling association football scores and inefficiencies
in the football betting market, Applied Statistics, 46(2), 265-280.
Figlewski, S. (1979). Subjective information and market efficiency in a betting market.
Journal of Political Economy, 87(1), 75-88.
Forrest, D. (1999). The past and the future of British football pools. Journal of Gambling
Studies, 15(2), 161-176.
Forrest, D., Goddard, J. & Simmons, R. (2005). Odds-setters as forecasters: The case of
English football. International Journal of Forecasting, 21, 551-564
Goddard, J. A. (2005). Regression models for forecasting goals and match results in
association football. International Journal of Forecasting, 21, 331-340.
Knight, F. H. (1965). Risk, Uncertainty and Profit. New York: Harper Torchbooks.
16
Pankoff, L. D. (1968). Market efficiency and football betting, The Journal of Business, 41(2),
203-214.
Pope, P. F. & Peel, D. A. (1989). Prices and efficiency in a fixed-odds betting market.
Economica, 56(223), 323-341
Reep, C., Pollard, R. & Benjamin, B. (1971). Skill and chance in ball games. Journal of the
Royal Statistical Society, Series A, 134, 623-629.
Rue, H. & Salvesen, O. (2000). Prediction and retrospective analysis of soccer matches in a
league. The Statistician, 49(3), 399-418.
Sharpe, G. (1997). Gambling on goals: A century of football betting. Mainstream, Edinburgh
Thaler, R. H. & Ziemba, W. T. (1988). Parimutuel betting markets: racetracks and lotteries.
Journal of Economic Perspectives, 2, 161-174.
Woodland, L. M. & Woodland, B. M. (1994). Market efficiency and the favorite-longshot
bias: the baseball betting market. Journal of Finance, 49, 269-279.
Woodland, L. M. (1994). Market efficiency and the favourite-longshot bias: The baseball
betting market. Journal of Finance, 49(1), 269-279
17
Table 1 Complementary log-log model for the probability of the away win
Variable Coefficient
(B)
SE(B) Z p
Constant -1.1897 1.7051 -0.70 0.485
RFO HH -0.0157 0.0038 -4.12 <0.001
RFO AA 0.0165 0.0042 3.90 <0.001


18
Table 2 OLS regression model with bookmaker odds of away win as the dependent
variable
Variable Coefficient
B
SE(B) t P
Constant 1.057 1.0060 1.05 0.295
RFO HH 0.0202 0.0022 9.32 <0.001
RFO AA -0.0194 0.0025 -7.79 <0.001
Time 1+ Home Team 0.0164 0.0047 3.50 0.001
Time 1+ Away Team -0.0129 0.0048 -2.66 0.008

19
8 7 6 5 4 3 2 1 0
20
10
0
-10
-20
-30
-40
-50
r > r*
P
r
o
f
i
t

f
o
r

u
n
i
t

s
t
a
k
e
s
0
1.595

Figure 1: Profit from rule bet on away win if r > r*