A Profitable Model For Predicting the

Over/Under Market in Football

Edward Wheatcroft
London School of Economics and Political Science, Houghton Street, London, United
Kingdom, WC2A 2AE.

Abstract
The over/under 2.5 goals betting market allows gamblers to bet
on whether the total number of goals in a football match will exceed
2.5. In this paper, a set of ratings, named ‘Generalised Attacking
Performance’ (GAP) ratings, are defined which measure the attacking
and defensive performance of each team in a league. GAP ratings are
used to forecast matches in ten European football leagues and their
profitability is tested in the over/under market using two value betting
strategies. GAP ratings with match statistics such as shots and shots
on target as inputs are shown to yield better predictive value than
the number of goals. An average profit of around 0.8 percent per bet
taken is demonstrated over twelve years when using only shots and
corners (and not goals) as inputs. The betting strategy is shown to
be robust by comparing it to a random betting strategy.

1 Introduction
Interest in forecasting the outcomes of sporting events has grown a great
deal in recent years. This is perhaps partly as a result of greater access to
data (The Economist (2017)), greater computational power and an increase
in the number of available betting markets. It is now commonplace for media
companies to include estimates of the probabilities of different outcomes of

1
sporting events in their coverage and consumers appear to be largely comfort-
able with this. This practice is perhaps most common in the United States
where probabilistic forecasts of National Football League (NFL), National
Basketball Association (NBA) and Major League Baseball (MLB) events are
routinely disseminated to the public and have been for some years (ESPN
(2018)). Probabilistic forecasting for the purpose of gambling has also taken
off in recent years. Increased competition in the bookmaking industry has
led to smaller profit margins and hence greater opportunities for gamblers.
Betting exchanges allow gamblers to bet against each other with one party
‘laying’ a bet for another and the exchange taking a small cut of the return.
Crucially, under this arrangement, there is no incentive for the company to
exclude those customers making consistent profits and thus, if high quality
forecast models can be built, opportunities for consistent profit making exist.
The purpose of this paper is twofold. First, a set of ratings, named Gener-
alised Attacking Performance (GAP) ratings, are introduced, which measure
the attacking and defensive ability of each team within a sports league. GAP
ratings can take as inputs any measure of attacking performance which could,
for example, be match statistics such as shots or shots on target or might
simply be goals. Second, it is demonstrated how a profit can be made by us-
ing GAP ratings to form probabilistic forecasts for the over/under 2.5 goals
market (which allows gamblers to bet on whether the total number of goals in
a match will exceed 2.5 or not) in association football (hereafter, just referred
to as ‘football’). It is found that, assuming that the maximum odds over all
bookmakers on the BetBrain website were available, using forecasts formed
using GAP ratings with the number of shots and corners as inputs alongside
a level stakes value betting strategy would have provided an average profit
margin of around 0.8 percent over the last 12 years over a total of 68,672
bets. It is subsequently demonstrated that the probability of achieving such
a return by chance (i.e. by betting randomly with the same frequency) is
extremely small (so small that, over one million trials, we were unable to find
such a high return by randomly choosing bets at this frequency) and thus
that the strategy appears to be robust in finding effective betting opportu-
nities. A second strategy based on the Kelly criterion is tested and shown to
be capable of yielding a larger profit by varying the stake according to the
degree of ‘value’ in the odds. The analysis is repeated using average odds and
it is found that a loss is made under both betting strategies. The implication
of this in terms of the efficiency of the market is then discussed.
Whilst a vast number of betting markets are now available, obtaining odds

2
data from a large number of games for these markets is generally difficult.
A large archive of historical odds from the over/under 2.5 goals and match
outcome markets are available, however, and this justifies the choice of the
former in this case. A future paper will consider the use of GAP ratings in
the match outcome market.
A large number of academic papers have focused on producing ratings
systems for sets of players or teams in a sporting context. This has been
done both with specific sports in mind such as football, and in a more general
context with the aim of providing ratings systems that translate between
sports. Almost certainly the most well known approach to producing sports
ratings is the Elo rating system which has a long history in sport and forms an
important basis for a range of ratings systems. Elo ratings were first designed
to produce rankings for chess players and the system was implemented by
the United States Chess Federation in 1960 (Elo (1978)). The Elo system
assigns ratings to each player and, from those ratings, probabilities of the
outcome of a game between two players or teams is estimated. The two
ratings are then updated to reflect the outcome once the game is finished.
The system was initially designed for the case in which outcomes are binary
(i.e. there are no ties), but, more recently, has been extended to account
for draws, making it applicable to sports such as football, in which draws
are common. At the updating stage (once a game has been played), the
system computes the difference between the estimated probabilities and the
outcome (assigned a one, a zero, or 0.5 for a draw). As such, the system in its
original form does not account for the size of a win; for example, in football,
there would be no additional increase in a team’s ranking from winning by
several goals over winning by one goal. Elo ratings have been demonstrated
in a football context and found to perform favourably with respect to six
other rating systems (Hvattum & Arntzen (2010)). In 2018, Fifa switched
to an Elo rating system to produce its international football world rankings
(Fifa (2018)). Elo ratings have also been applied to other sports such as, for
example, Rugby League (Carbone et al. (2016)) and video games (Suznjevic
et al. (2015)), whilst fivethirtyeight.com produce probabilities for NFL
(FiveThirtyEight (2019a)) and NBA (FiveThirtyEight (2019b)) based on Elo
ratings.
The Elo rating system provides a single rating for each participating team
relating to the overall ability of that player or team. This approach has been
taken for other ratings systems. The pi-rating system assigns a home and an
away overall rating to each team in a league, both of which are updated after

3
each match in which a team is involved. The change in rating is dependent on
the winning margin of each team but is tapered such that additional goals on
top of already large winning margins are assigned less weight (Constantinou
& Fenton (2013)). The approach taken in this paper is to assign attacking
and defensive ratings to each team and to use those ratings to calculate
match probabilities. This approach has been taken by a large number of
authors. Maher (1982) used fixed ratings for each team in combination with
a Poisson model to estimate the number of goals scored by each team (but not
match probabilities). This approach was extended by Dixon & Coles (1997)
who used similar attack and defence ratings in a Poisson regression model
to estimate match probabilities. They combined these with a value betting
strategy (described and used later in this paper) to demonstrate a significant
profit for cases in which the model suggested a large discrepancy between the
estimated probabilities and the probabilities implied by the odds. Whilst the
approach of Dixon and Coles assumed a fixed attack and defensive rating over
time for each team, this was extended by Rue & Salvesen (2000) who defined
a Bayesian model to allow the ratings to vary over time. Dixon & Pope
(2004) used a slightly modified version of the model proposed by Dixon and
Coles to demonstrate a profit using a wider range of published bookmaker
odds. Other examples of the use of attacking and defensive ratings systems
include Karlis & Ntzoufras (2003), Lee (1997) and Baker & McHale (2015).
The use of machine learning in prediction has become more and more pop-
ular in recent years. A 2017 special issue in the journal ‘Machine Learning’
presents the results of a ‘Soccer Prediction Challenge’ in which participants
were provided with results from 216,743 past football matches from around
the world and were asked to make predictions regarding 206 future matches
(Berrar et al. (2019)). A number of the entrants achieved very encouraging
results. Hubáček et al. (2019) won the competition using gradient boosted
trees whilst Constantinou (2019) finished second using a model called Do-
lores which combines dynamic ratings with Hybrid Bayesian Networks. Other
past uses of machine learning to predict football results include O’Donoghue
et al. (2004), Van Haaren & Van den Broeck (2015), Joseph et al. (2006) and
Hucaljuk & Rakipović (2011).
A number of other approaches to football prediction have been taken over
the years. These include Bayesian generalised linear models (Rue & Salvesen
(2000)), Bayesian Networks (Constantinou et al. (2012)) and neural networks
(Huang & Chang (2010)).
The question of how to translate team ratings into probabilistic forecasts

4
was addressed by Goddard (2005). Two approaches are generally in use. The
first approach uses ratings as inputs to Poisson regression models, simulating
the number of goals scored by each team and calculating the probabilities
of each outcome accordingly. The other approach predicts the probability of
each result directly using methods such as ordered probit regression. It was
found that there is little difference in the performance of the two approaches.
In this paper, the latter approach is used with the probability of exceeding
2.5 goals modelled directly using logistic regression.
In recent years, the idea that other match statistics might provide a better
indication of the relative performance of the teams than the actual number of
goals scored has been investigated. The reasoning behind this is that there is
a significant element of chance involved in scoring a goal and that the number
of goals actually scored might be indicative of the relative performances of
each team. More effort has thus been invested in attempting to use data to
understand the relative performance of the two teams in the hope that this
will provide a better indication of future performance (Cintia et al. (2015)).
The concept of ‘expected goals’ has also taken off in recent years (Rathke
(2017), Opta (2018)). The aim here is simply to attempt to calculate the
number of goals a team would be ‘expected’ to score given the location and
nature of all its shots taken during the match (Eggels (2016)). The use of shot
data is conceptually very similar to that of expected goals. The difference,
however, is that, in the former case, a constant weighting is placed on each
shot, since the nature and location of each one is not taken into account.
The use of match statistics such as shots, shots on target and corners has
become more feasible in recent years due to the greater availability of these
data. Free data on overall match statistics such as shots, corners, fouls, red
and yellow cards as well as bookmakers’ odds are freely available in an easy-
to-use format at the popular and widely used Football-data.co.uk website
(Football Data (2018)). It is from this source that the data used in this
paper are taken. Other sources collect a vast amount of data for matches in
top European leagues including shot locations, shot methods (headed, right
footed, left footed etc), the locations of and outcomes of passes, changes
of possession and much more (Opta Football (2018)). The cost of obtaining
these data is often prohibitively expensive for individuals, however, and these
data are not considered in this paper.
This paper is organised as follows. First, in section 2, background infor-
mation is given regarding bookmakers’ odds and how these can be interpreted
as implied probabilities. In section 3, the data utilised in this paper is de-

5
scribed. In section 4, GAP ratings are defined. In section 5, the approach
taken in this paper to producing probabilistic forecasts from GAP ratings
is described. Approaches to forecast evaluation and parameter selection are
described in section 6. The approach taken to dealing with promotions and
relegations is described in section 7. The two betting strategies used in this
paper are described in section 8. The experimental design is defined in sec-
tion 9 and the results are described in section 10. Section 11 is used for
discussion.

2 Background
This paper makes extensive use of betting odds, using them both as inputs
with which to produce probabilistic forecasts and as a basis with which to
demonstrate the performance of the forecasts as a potential money making
tool. Before proceeding, however, it is useful to define exactly what is meant
when betting odds are referred to. The format of betting odds varies some-
what in different regions of the world. In the context of this paper, it is
convenient to consider ‘decimal’, or ‘European style’, betting odds (to which
other formats can easily be converted). Decimal odds simply indicate by how
much the stake is multiplied in the event of that bet being successful. For
example, if the odds offered on an event are 3, a unit stake would generate
a return of 3. The ‘fractional’, or ‘British style’ odds, in this case would be
given as ‘2/1’ indicating a profit of 2 for a stake of 1. Hereafter, in this paper,
betting odds are given in decimal format.
Another useful concept is that of the ‘odds implied’ probability. Define
the decimal odds for the ith event to be Oi . The ‘odds implied’ probability
is simply defined as the inverse, i.e. ri = O1i . For example, if the odds on an
event are Oi = 3 then ri = 31 . Whilst, conventionally, probabilities over a set
of exhaustive events are required to add up to one, this is not the case for odds
implied probabilities and, in fact, the sum of odds implied probabilities for an
event will usually exceed one; the excess representing the bookmaker’s profit
margin or ‘overround’. The overround represents a significant challenge to
gamblers because it means that the bookmakers shorten their odds in order
to make a profit.

6
3 Data
This paper utilises the repository of past football match data available at
www.football-data.co.uk which supplies free-to-access match-by-match data
for a range of European Leagues dating back, in some cases, as far back as
the 1993/1994 season. Whilst early years of the data set feature only the
final result of each match, in more recent years, the information provided
has grown more rich such that, for each football match in any of 10 differ-
ent European leagues, full and half time results, the number of shots, shots
on target and number of corners taken by each team are given along with
betting odds of the final result (home win, away win or draw) and of the
over/under 2.5 goal market from multiple bookmakers. The maximum and
average odds for the match outcome and over/under 2.5 goal markets over
many bookmakers, collected from the BetBrain odds comparison website, are
also given. From the 2017/2018 season onwards, the number of leagues in
which all of this information is available has been increased to 22 (though
data from the extra 12 leagues are not utilised in this paper).
In this paper, a rating system is introduced, called the GAP rating system,
and those ratings are used as inputs to produce probabilistic forecasts. The
available data provide an opportunity for robust evaluation of forecasts and
of potential betting strategies. In this paper, the performance of the forecasts
is compared with different inputs to the GAP ratings representing different
measures of attacking performance, including shots and corners. Only leagues
for which this information is available along with the maximum and average
BetBrain odds in the over/under 2.5 goal market are thus considered. The
leagues considered in this paper are summarised in table 1. The total number
of football matches considered is thus 54,437. The data set does not include
cup games, playoffs or any other extra matches during the regular season and
thus these are not considered in this paper.

4 Generalised Attacking Performance (GAP)

Ratings
Generalised Attacking Performance (GAP) ratings are now formally defined.
Consider a football league consisting of N different teams who all play each
other a number of times over a season. For a given football match, let Sh
and Sa be some measure of the attacking performance of the home and away

7
 League First available season Number of matches
English Premier League 2005/2006 6460
English Championship 2005/2006 6624
English League One 2005/2006 6624
English League Two 2005/2006 6624
English National League 2005/2006 6488
Scottish Premier League 2005/2006 2940
Spanish Primera Liga 2005/2006 4910
Italian Serie A 2005/2006 4908
French Ligue One 2005/2006 4908
German Bundesliga 2005/2006 3951

Table 1: Football league data used in this paper.

teams respectively, where the definition of ‘attacking performance’ (hereafter

frequently referred to as the ‘GAP rating input’) is given by the user and
is usually derived from match statistics. Most obviously, the input could be
defined as the number of goals scored by each team but could also be the
number of shots taken, the number of shots on target, the number of corners,
a combination of these, or any other indication of attacking performance.
Each team is given a separate attacking and defensive GAP rating for home
and away games defined as follows:

• Hia - The attacking GAP rating of the ith team for home games

• Hid - The defensive GAP rating of the ith team for home games

• Aai - The attacking GAP rating of the ith team for away games

• Adi - The defensive GAP rating of the ith team for away games

The attacking GAP ratings relate roughly to the expected attacking per-
formance that a team should achieve against an average team in the league
whilst the defensive ratings relate to the expected attacking performance of
an average opposing team. The best teams in the league will thus have a
high attacking rating and a low defensive rating. The following rule is used
to update the ith team’s GAP ratings after a home match against the jth

8
team:
Hia + Adj
Hia = max(Hia + λφ1 (Sh − ), 0),
2
a a
Hia + Adj
Ai = max(Ai + λ(1 − φ1 )(Sh − ), 0),
2 (1)
d d
Aaj + Hid
Hi = max(Hi + λφ1 (Sa − ), 0),
2
Aaj + Hid
Adi = max(Adi + λ(1 − φ1 )(Sa − ), 0)
2
The GAP ratings for the jth team (away to the ith team) are updated as
follows:
Aaj + Hid
Aaj = max(Aaj + λφ2 (Sa − ), 0),
2
Aaj + Hid
Hja = max(Hja + λ(1 − φ2 )(Sa − ), 0),
2 (2)
Hia + Adj
Adj = max(Adj + λφ2 (Sh − ), 0),
2
Hia + Adj
Hjd = max(Hjd + λ(1 − φ2 )(Sh − ), 0),
2
where, λ > 0, 0 < φ1 < 1 and 0 < φ2 < 1 are parameters to be selected.
For a given match, a home team can be said to have outperformed expec-
tations in an attacking sense if its attacking performance is greater than the
mean of its attacking rating and the opposition’s defensive rating and thus,
when this is the case, both its home and away attacking ratings are increased
(or decreased, if its attacking performance is lower than expected). The same
is true if an away team outperforms expectations. A team’s defensive ratings
are impacted in a similar way such that, if an opposing team’s attacking
performance is lower than expected, its defensive rating is decreased (note
that lower defensive ratings indicate better defensive performance). The pa-
rameter λ determines the influence of the last match on the ratings of each
team whilst φ1 and φ2 determine the influence of a home match on a team’s
away ratings and the influence of an away match on a team’s home ratings
respectively. If φ1 = 0, the home team’s away ratings will not be affected
whilst the same is true of the away team’s home ratings if φ2 = 0. The

9
maximum operator is included to ensure that the GAP ratings for each team
cannot become negative. Parameter selection is discussed in section 6.
The GAP rating system described above has some similarities to, and, to
some extent, is inspired by, the pi-rating system defined in Constantinou &
Fenton (2013). Some of the similarities between the two ratings systems are
described below:

1. The parameter λ performs the same function as for both GAP rat-
ings and pi-ratings. The Elo rating system also has a similar learning
parameter.

2. Both GAP ratings and pi-ratings consist of separate home and away
ratings for each team (in order to account for the impact of home
advantage).

3. In GAP ratings, the parameters φ1 and φ2 govern the impact of a home

match on a team’s away ratings and the impact of an away match on a
team’s home rating respectively. In the pi-rating system, a parameter
γ is defined with a similar function which serves both purposes (it is
necessary for the one parameter to serve both purposes to ensure that
the sum of the ratings over the entire league remains zero).

4. A pi-rating can roughly be interpreted as a team’s expected winning

(or losing) margin over an average team in a league. GAP ratings can
similarly be interpreted as the expected number of attacking plays (as
defined) of a team against an average opponent.

Despite the above similarities, there are a number of key differences:

1. GAP ratings are defined more generally than pi-ratings such that any
measure of attacking performance can be used as an input. It would,
however, be straightforward to extend pi-ratings in this way also.

2. In the GAP rating system, each team has four ratings corresponding to
its attacking and defensive performance and the location of the match
(home or away). Pi-ratings do not distinguish attacking and defensive
ability and thus each team is assigned two ratings corresponding to the
expected difference in the number of goals scored by each team. This
makes GAP ratings more suitable for forecasting the total number of
goals in a match.

10
 3. GAP ratings define the expected number of attacking plays by a home
(away) team to be the average of their home (away) GAP rating and
the away (home) defensive GAP rating of the opposition. Pi-ratings
define the expected winning (losing) margin of the home team to be
the difference between the pi-ratings of the two teams.
4. pi-ratings are constrained such that the sum of the ratings over the
league is zero. For GAP ratings, the only constraint is that all ratings
must be greater than or equal to zero.
5. As described in the list of similarities above, GAP ratings have separate
parameters for the effect of a home team’s performance on its away
rating and an away team’s performance on its home rating. This extra
flexibility is possible in the former case because of the difference in the
constraint described above.

5 From GAP Ratings to Probabilistic Fore-

casts
Here, an approach to using GAP ratings to predict the probability that there
will be three or more goals in the match is described. Although a variety
of approaches could be taken to do this, in this paper, a simple logistic
regression approach is taken, allowing for multiple variables to influence the
p̂
estimated probability. Let y = log( 1−p̂ ) where p̂ is the estimated probability
of the total number of goals exceeding 2.5. A regression model is defined as
ŷ = α̂ + βˆ1 (H a + H d + Aa + Ad ) + βˆ2 ri
i i i i (3)
1
where ri = is the odds implied probability (which may be derived from a
oi
single bookmaker or averaged over several). The parameters are estimated
using least squares estimation. From ŷ, the estimated probability p̂ can easily
be derived.

6 Parameter Selection and Forecast Evalua-

tion
GAP ratings require the selection of three parameters: λ, a parameter that
governs the impact of the most recent match on a team’s ratings and φ1 and

11
φ2 , parameters that govern the impact of a home match on a team’s away
ratings and of an away match on a team’s home ratings respectively. In
addition, the selection of a number of extra parameters is required depending
on the approach taken to forming probabilistic forecasts. Under the approach
described in section 5, and which is used to calculate the results shown in this
paper, a further three parameters are used in the logistic regression making
the total number of parameters six.
Given the probabilistic nature of the forecasts, parameter selection is per-
formed with the aim of maximising forecast performance given a scoring rule,
a function of a probabilistic forecast and its outcome that, when averaged
over many forecasts and outcomes, measures the skill of the forecasting sys-
tem (other approaches such as attempting to maximise profit could also be
used but are not considered here). An important property of scoring rules
is propriety. A score is proper if it is optimised in expectation by the true
underlying probability or probability distribution from which the outcome
was drawn. It is often argued that only scoring rules that are proper should
be used in practice since, otherwise, a forecaster may be encouraged to issue
forecasts that don’t reflect their true beliefs (Bröcker & Smith (2008)).
The scoring rule of choice in this paper is the ignorance score (Good
(1952), Roulston & Smith (2002)), given by S(p(Y )) = − log2 (p(Y )) where
p(Y ) is the probability or (in the continuous case) probability density placed
on the outcome Y . The ignorance score is proper and is also conveniently
linked to the more well known concept of maximum likelihood in that the
aim is to maximise the mean of the logarithm of the probability placed on
the outcomes. A benefit of the ignorance score is in its interpretation. The
difference between two ignorance scores calculated using two different forecast
systems of the same outcome can be interpreted as the number of bits of
information gained from using one over the other. For example, if forecast
system A has a mean ignorance of 0.75 and forecast system B has a mean
ignorance of 0.5, forecast system B provides 0.25 extra bits of information,
on average, than forecast system A since ignorance is a negatively oriented
score. This can be interpreted as placing 20.25 ≈ 1.19 more probability on
the outcome, on average.
Parameter selection is performed by minimising the mean ignorance score
over all previous matches. This is done using the Nelder-Mead simplex algo-
rithm as implemented in the fminsearch function in Matlab. For computa-
tional reasons, this is done with respect to the three parameters that impact
the ratings and the logistic regression parameters are optimised using least

12
squares given the ratings that result. The parameters are therefore not all
optimised simultaneously. Since it is computationally expensive to optimise
the ratings parameters, this is done only between seasons and over all leagues
simultaneously. The logistic regression parameters, however, are optimised
after every round of games since it is computationally cheap to do so.
A useful concept commonly used in weather forecasting is that of the ‘cli-
matology’. A climatology is simply defined as a distribution of past states,
over some given time period. An example would be the distribution of max-
imum temperatures at a specific location on a given day over the last 100
years or, in the binary case, the proportion of days on which it has rained
over that time. The climatology gives a useful summary of past states but
can also be treated as a forecast itself. For example, weather forecasts can-
not generally be expected to have skill beyond around two or three weeks
and thus, if one wanted to gain insight into how, say, the temperature might
pan out a year from now at a given location, the climatology would likely
provide the best information available. The concept of climatology can easily
be transferred into sports forecasting where, in this case, observing three or
more goals in a football match is conceptually similar to observing a rainy
day. The climatology is therefore used as an alternative baseline forecast,
the skill of which is compared with that of the forecast system defined in
this paper. If the forecast system cannot outperform the climatology, it is of
little use in practice. The probability of exceeding 2.5 goals is, in fact, found
to be around 50 percent. All mean ignorance scores in this paper are quoted
relative to that of the climatology (i.e. with the climatological ignorance
subtracted) such that, if the mean ignorance is less than zero, the forecast
system outperforms the climatology, on average.
An alternative approach to the parameter selection approach performed
in this paper would be to maximise the profit made through the betting
strategy. However, this would cause the function to be optimised to be less
smooth and would therefore make optimisation more challenging as the risk
of falling into a local maximum (minimum) would be increased. This is
therefore left as a suggestion for possible future work.

7 Dealing With Promotions and Relegations

It is almost universally the case that football leagues feature promotion
and/or relegation, i.e. if a team finishes top or close to the top of a league,

13
they are offered the chance to play in the league above, whilst teams that
finish close to the bottom are ‘relegated’ to the league below. This links into
the broader question of how a team’s ratings should transfer from one season
to another. Whilst a team that remains in the same league from one season
to the next already has a set of ratings with which to start the following
season, teams that enter the league will not and thus some approach is re-
quired to initialise the ratings of such teams. A number of approaches could
be taken to this, varying somewhat in complexity. For example, data on
summer signings, preseason bookmakers’ odds and predictions from pundits
could all be utilised (though the latter may be hard to quantify). There is
also a question of how realistic it is for those teams that stay in the same
league to maintain their ratings from one season to the next given managerial
changes, player transfers etc. and thus extra data could be used to adjust
the ratings before the start of each season. In this paper, however, for sim-
plicity, if a team remains in the same division, they retain their ratings from
the previous season whilst teams that are promoted to the league (i.e. come
from the league below) are given the average ratings of the teams relegated
from that league in the previous season and teams relegated to that league
are given the average rating of the promoted teams. Note, however, that
it is common for promoted teams to outperform the relegated teams they
replace. For example, it has been found that, in the English Premier League,
the promoted teams tend to achieve around 8 more points on average than
the relegated teams in the previous season (Constantinou & Fenton (2017)).

8 Betting Strategies
This paper uses two different betting strategies based on the concept of ‘value
betting’. In value betting, whether a bet is placed or not is determined by
whether the probabilistic forecast (if taken as a true probability) implies that
the odds have ‘value’. The odds on a particular outcome are said to have
value if they are longer (i.e. generate a larger return) than they should be,
given the true probability1 of that outcome. Specifically, the odds are said
to have value if pi > ri where, as defined in section 2, pi represents the
‘true’ probability and ri represents the odds implied probability. When this
1
Here, we take the frequentist definition of probability, i.e. that the probability is the
expected relative frequency of outcomes were the match able to be repeated independently
an infinite number of times

14
is the case, the expected profit for the gambler is positive and, thus, by only
betting when there is value, they would profit in the long term. In practice,
however, due to inevitable limitations inherent in any real world predictive
model, the underlying probability can only be estimated and thus will differ
from the true probability. Nonetheless a betting strategy can be defined
by replacing the true probability, which is not known, with an estimated
probability. This strategy is referred to throughout this paper as the Level
Stakes strategy. Under this strategy, a unit bet is made on the ith outcome
if and only if the forecast probability exceeds the odds implied probability,
i.e. if p̂i > ri , where p̂i is the forecast probability.
An alternative strategy makes use of the Kelly Criterion (Kelly Jr (1956)).
Under the Kelly criterion, the proportion of one’s wealth placed on a partic-
ular outcome is given by
rp − 1
f = max( , 0) (4)
r−1
where p is the gambler’s estimated probability of the outcome and r is the
decimal odds offered by the bookmaker. If the forecast probability is less
than the odds-implied probability, the bet is not taken since the stake is
zero. The idea behind the Kelly criterion is that the bankroll of the gambler
‘grows’ over time by increasing the stake proportionally to their bank. In this
paper, however, the approach of Boshnakov et al. (2017) is taken, in which
equation 4 is always taken as a proportion of 1 and therefore does not depend
on the bankroll. Each stake is multiplied by a constant such that the average
stake when a bet is taken is exactly 1. This is done so that the results are
directly comparable with the level stakes case above. This is referred to in
the results section of this paper as the Kelly strategy.
In value betting, considering forecast probabilities rather than actual
probabilities removes the guarantee of making a profit in expectation and
thus the performance of the forecasts and betting strategy will depend heav-
ily on the skill of the forecasts or, more specifically, the ability of the forecasts
to identify profitable betting opportunities. With this in mind, some addi-
tional conditions are placed on whether to bet or not on a specific match.
Bets are not placed until both teams have played at least six league games
during that season in order for the ratings to have sufficiently ‘learned’ about
the attacking and defensive ratings of the teams. Although teams remaining
in the same league from one season to the next retain their rating, no account
is taken of transfer activity and other close-season changes and so the ratings

15
are prone to being uninformative at this stage of the season. The first six
matches are chosen to be ineligible for betting in this case because there is
some evidence that this is the point at which the league starts to settle down
and reflect the overall strength of the teams (Cronin (2019)). In addition,
bets are never placed on the last six games for each team as games at this
time of the season can be particularly unpredictable due to differences in mo-
tivation between teams. For example, a team fighting relegation at the end of
the season will have more to play for than a team safely in mid-table at that
stage and so there is more incentive for the former to pick their best team or
put more effort into winning the game. Indeed, bookmakers will often make
the relegation threatened team favourites to win in those situations. These
kinds of factors are hard to factor into a mathematical model and whether
to bet or not may require some human judgement beyond the scope of this
paper.

9 Experimental Design
The performance of the probabilistic forecasts formed using GAP ratings
along with the two betting strategies is demonstrated with the following
experimental design. Attacking and defensive GAP ratings are calculated for
each team and updated after each match. Promotions and relegations are
dealt with using the approach described in section 7 in which newly promoted
teams are initialised with the average ratings of the relegated teams from the
previous season and relegated teams with the average ratings of the promoted
teams. For each match, the current attacking and defensive GAP ratings of
each team and the odds-implied probabilities from the BetBrain maximum
odds are then used to calculate a forecast probability of the number of goals
exceeding 2.5 using the logistic regression model defined by equation 3. The
parameter values used to calculate the ratings are optimised over all available
data in previous seasons and used for the entirety of that season whilst the
logistic regression parameters are optimised after every round of games (see
section 6 for details). Data from the 2000/2001 season are used solely for
parameter selection purposes since no previous data are available. Both the
Level Stakes and the Kelly strategies defined in section 8 are then used and
the profit/loss for each match based on these strategies are calculated. This
is repeated under the assumptions of both maximum and average Betbrain
odds. As explained in section 8, the first and last six matches of the season

16
for each team are not treated as eligible for betting.
The above experimental design is applied six times using different mea-
sures of attacking performance as inputs to the GAP ratings. The six differ-
ent inputs tested are listed in table 2. The results of the experiment using
each GAP rating input are described and compared in the results section.

Measure
1 Goals
2 Shots
3 Shots on Target
4 Corners
5 Shots and Corners
6 Shots on Target and Corners

Table 2: GAP rating inputs tested

10 Results
The stated aim of this paper is to build strategies with which to make a
long term gambling profit on European football matches. The obvious first
consideration is therefore whether the forecasts and betting strategy are able
to achieve this. Here, the profitability of both betting strategies is assessed
in two different settings. The majority of the analysis is performed in the
case in which the maximum BetBrain odds are assumed to be available. The
main results are then repeated but with the profit/loss calculated using the
average odds. The effect of betting with odds that are a weighted average of
the maximum and average odds is then assessed. Finally, the forecast skill
of the forecasts under each GAP rating is compared.

10.1 Maximum Odds

The cumulative profit a gambler would have made by applying the Level
Stakes strategy over all of the considered seasons using the maximum Bet-
Brain odds is shown as a function of time in figure 1 for each input to the
GAP ratings. The profit in each season for each input is shown in table 3.
Under four of the considered inputs, a profit would have been made over

17
this period. In the cases in which only shots and both shots and corners are
considered, the total profit is substantial, reaching around 500 units. This
corresponds to an average profit of around 0.8 percent per bet taken. It
is unclear whether the small financial gain in considering corners alongside
shots is due to chance or a genuine improvement in the ability to find value
bets. There is a substantial difference between the profit made from using
only shots on target and from including corners as well, however, suggesting
that corners may be a useful addition in terms of the performance of the
strategy. It is also interesting to note that using only corners as inputs out-
performs using only shots on target. Finally, strikingly, when goals scored
is used as the GAP rating input, the gambling return is far lower than for
each of the other measures, resulting in a large loss. It is worth noting that,
were the gambler to bet on the number of goals exceeding 2.5 in all possible
matches (that is including those deemed not to have value), they would make
an average loss of 1.81 percent whilst, if they were to bet that the number of
goals would be less than 2.5, they would make an average loss of 1.46 percent.
There is thus a small bias in which better odds are typically offered on there
being fewer than 2.5 goals. This, however, is not enough to be exploitable in
terms of making a profit.
Season G S ST C S+C ST + C
2005/06 +35.14 +14.16 +0.95 −6.50 +41.54 −7.38
2006/07 −18.50 −37.09 −88.54 +0.12 +11.81 −3.33
2007/08 −51.35 +80.93 +47.69 +45.70 +65.74 +45.20
2008/09 −115.48 −26.17 +46.45 −12.60 −39, 82 −1.57
2009/10 −70.59 +76.49 −14.41 −39.40 +66.06 −10.01
2010/11 −106.79 +58.15 −9.07 +55.23 +45.22 +39.80
2011/12 −28.46 +173.16 +131.73 −13.41 +127.94 +88.79
2012/13 −31.59 +38.15 +16.50 +9.61 +91.54 +47.25
2013/14 −19.60 +13.44 −16.41 +29.17 +17.38 −4.38
2014/15 −99.30 −49.63 −86.33 +1.88 +16.64 −19.26
2015/16 −82.91 −13.05 −127.51 −94.08 +5.29 −110.63
2016/17 −2.91 −18.06 −102.92 −56.59 −13.65 −102.91
2017/18 +31.50 +89.51 +53.51 +105.94 +99.36 +88.45
Total −631.12 +414.72 −148.36 +25.07 +535.01 +50.02

Table 3: Total profit in each season from using the Level Stakes strategy with
maximum BetBrain odds when using goals (G), shots (S), shots on target
(ST), corners (C), shots and corners (S+C) and shots on target and corners
(ST+C) as the input to the GAP ratings. The input that results in the
highest profit is highlighted in green.

Whilst the above results consider the performance of the forecasts over
all leagues considered, it is also of interest to assess the performance over
individual leagues. The overall profit that would have been made in each

18
Figure 1: Cumulative profit over time from using the Level Stakes betting
strategy with the maximum BetBrain odds for each input to the GAP ratings.

league under the Level Stakes strategy is shown in table 4 for each GAP rating
input. The strategy appears to be fairly robust. For example, using shots and
corners as the GAP rating input would have made a profit in six out of the
ten leagues considered with the English Premier League and Championship
returning the largest profits. In fact, for each league considered, at least one
of the inputs would have resulted in a profit. Notably, however, using goals
as the GAP rating input would have resulted in an overall loss in all ten
leagues considered.
The results above demonstrate that using the forecasts with the Level
Stakes betting strategy would have produced a cumulative profit over the
seasons considered. However, from these results alone, it is not possible to
tell whether this could have occurred entirely by chance, i.e. the forecasts

19
 League G S ST C S+C ST + C
English Premier League −56.14 +129.70 −46.19 +41.18 +139.63 −10.72
English Championship −82.19 +187.79 +34.00 +78.91 +204.05 +62.29
English League One −27.24 −40.74 −26.60 −68.85 −14.14 −31.40
English League Two −44.11 +3.62 +9.93 −22.93 −2.16 −49.36
English National League −47.30 +21.39 +90.31 +8.38 +99.58 +97.05
Scottish Premier League −43.49 +35.53 −33.33 +133.27 +80.82 +22.32
Spanish La Liga −105.74 +41.74 +6.58 −33.79 +44.38 +23.18
French Ligue One −98.56 +26.82 −70.57 −53.97 −33.45 −31.15
Italian Serie A −65.56 −30.92 −69.15 −62.19 −26.65 −8.75
German Bundesliga −60.79 +39.79 −43.34 +5.06 +42.95 −23.44
Combined −631.12 +414.72 −148.36 +25.07 +535.01 +50.02

Table 4: Total profit for each league, along with the combined profit, from
using the Level Stakes strategy with maximum BetBrain odds when using
goals (G), shots (S), shots on target (ST), corners (C), shots and corners
(S+C) and shots on target and corners (ST+C) as the input to the GAP
ratings. The input that results in the highest profit is highlighted in green.

got lucky in identifying bets that resulted in a profit over time. In order
to assess this, actual betting performance is compared with the results that
would typically be expected were the gambler to bet randomly with the same
probabilities. A simple approach is therefore taken whereby the number
of bets actually taken in a season is preserved but those bets are chosen
randomly, rather than using the outlined betting strategy. This process is
repeated multiple times with different realisations to give a range of possible
scenarios of how the profit and loss might turn out when this is done. If the
actual results are typical of what happens when bets are made randomly,
then little confidence should be had in the performance of the forecasts for
identifying value bets. In figure 2, the results of doing this using shots and
corners as the input to the GAP ratings are shown in the upper panel. Here,
the green line shows the actual profit and loss over time whereas each black
line shows the profit and loss achieved by randomly choosing bets under
different realisations. In the lower panel, the red line shows the profit/loss
obtained from only betting when the forecasts suggest negative value and the
black lines show the profit/loss from betting randomly at the same frequency.
Since the actual profit made is far higher than what would be typical if the
bets were chosen randomly, a great deal of confidence is gained that the
forecasts are informative for the purpose of this strategy. Moreover, it is clear
that the forecasts are also capable of identifying bets that offer negative value.
The latter could potentially be useful when gambling on a betting exchange
in which gamblers are able to act like a bookmaker and ‘lay’ bets for other
gamblers. Given the lack of available betting exchange data, this is not tested

20
in this paper, however.

Figure 2: Upper panel: Cumulative profit as a function of time using the

Level Stakes betting strategy (green line) and when bets are randomly taken
at the same frequency (black lines) using maximum BetBrain odds. Lower
panel: Cumulative profit as a function of time betting only when the forecast
suggests there is negative value (red line) and when bets are taken randomly
at the same frequency (black lines).

In order to illustrate the results of the above process for each GAP rating
input, the proportion of random bet scenarios with a higher profit/loss is
shown in figure 3 (note the logarithmic scale) as a function of time (calculated
at the end of each season). Here, in all cases other than when goals are used
as inputs to the GAP ratings, the betting strategy is eventually able to
outperform all random bet scenarios tested. When goals are used as the
input, on the other hand, there is little evidence that betting performance is

21
better than that which could be achieved by chance.

Figure 3: The proportion of random bet scenarios with higher cumulative

profit as a function of time for each GAP rating input considered when using
maximum BetBrain odds.

Whilst the results given above suggest that the Level Stakes betting strat-
egy is able to identify enough value bets to make a robust long term profit
for certain GAP rating inputs, it is important to rule out the possibility that
the forecasts don’t simply identify arbitrage opportunities. An arbitrage
opportunity is a case in which opportunities exist for strategically placing
bets with bookmakers that have differing odds and guaranteeing a profit.
Such an opportunity is available when the overround is less than zero. The
mean overround of bets selected by the forecasts and betting strategy in each
league and for each GAP rating input is shown in table 5. Here, in all cases,
the mean overround is greater than zero, showing that a profit is made de-

22
spite the built in bookmakers’ profit margin of the chosen bets and hence
the favourable results do not simply occur as a result of identifying arbitrage
opportunities.
League G S ST C S+C ST + C
English Premier League 0.011 0.011 0.011 0.011 0.012 0.011
English Championship 0.013 0.014 0.014 0.013 0.014 0.014
English League One 0.016 0.016 0.016 0.016 0.016 0.016
English League Two 0.016 0.016 0.016 0.016 0.016 0.016
English National League 0.017 0.017 0.018 0.018 0.018 0.018
Scottish Premier League 0.015 0.016 0.016 0.016 0.016 0.016
Spanish La Liga 0.012 0.012 0.012 0.012 0.012 0.012
French Ligue One 0.015 0.015 0.015 0.015 0.013 0.014
Italian Serie A 0.012 0.016 0.012 0.013 0.012 0.013
German Bundesliga 0.012 0.014 0.013 0.012 0.012 0.012
Combined 0.014 0.014 0.014 0.014 0.014 0.014

Table 5: Mean overround of bets placed in each league, along with that of
all the leagues combined, when using goals (G), shots (S), shots on target
(ST), corners (C), shots and corners (S+C) and shots on target and corners
(ST+C) as inputs to the GAP ratings and using maximum BetBrain odds.

The forecasts considered in this paper are formed using logistic regression
with the sum of both the attacking and defensive ratings for the home and
away team as a single term (equation 3). However, an alternative is to use
the GAP ratings as separate inputs to allow each of the ratings to impact
the forecasts to a potentially different extent. Although details of the results
are omitted here, this was tested over all eligible matches (i.e. more than
six matches into the season and more than six matches before the end of
the season) for each GAP rating input to assess whether this improved the
forecasts. The resulting forecasts were found to have less support than the
forecasts considered in this paper, when tested using Akaike’s Information
Criterion (AIC). These results imply that there is little evidence to distin-
guish the importance of each rating in terms of building the forecasts and so
this is not considered further.
Another issue worth visiting is to check that the profit/loss achieved us-
ing the Level Stakes betting strategy does not simply occur as a result of
testing too many inputs to the GAP ratings and finding one that works well
in terms of making a profit by chance. A simple way to account for multiple
testing is to use the Bonferroni correction (Bonferroni (1936)). The Bonfer-
roni correction accounts for multiple tests by adjusting the significance level
α to account for the number of hypotheses tested. Under the Bonferroni
correction, if a significance level of α is required for a single test, and m tests

23
are performed, the required significance level of each of the tests is adjusted
α
to m .
In figure 3, the profit/loss from each GAP rating input was compared
with that achieved from randomly betting at the same frequency. This was
done by testing 1024 different random sets of bets. The proportion of ran-
domly selected sets of bets achieving a greater profit than that of the Level
Stakes betting strategy can then be interpreted as an estimated p-value of
the profit/loss. Here, this is done again but with one million random sets
of bets in order to estimate the p-value more accurately. This can then
be compared to adjusted significance levels calculated using the Bonferroni
correction. The estimated p-value for each GAP rating input is shown in
table 6, along with the proportion of random sets of bets that would have
returned a profit. Each are compared with adjusted significance levels at the
5%, 1% and 0.1% levels (denoted with one, two and three stars respectively).
Since the profit/loss for each GAP rating input other than goals is highly
significant even when multiple testing is accounted for, the results in this
paper do not appear to result from testing too many inputs. In addition, the
fact that, in each case, only a handful of randomly chosen sets would have
resulted in a profit demonstrates the difficulty of profiting over this number
of bets without truly having some ability to identify value.

GAP rating input Estimated p-value Prop. trials with profit

Goals 0.541765 0.000017
Shots 0.000000∗∗∗ 0.000009
Shots on Target 0.000124∗∗ 0.000000
Corners 0.000003∗∗∗ 0.000004
Shots and Corners 0.000000∗∗∗ 0.000008
Shots on Target and Corners 0.000000∗∗∗ 0.000000

Table 6: Estimated p-values for each GAP rating input along with the pro-
portion of random sets of bets returning a profit when using maximum Bet-
Brain odds. Estimated p-values with one, two and three stars next to them
indicate significance at the 5%, 1% and 0.1% levels respectively, adjusted
using the Bonferroni correction.

All of the above results make use of the Level Stakes strategy in which
a unit bet is taken if a forecast implies the odds offer ‘value’. However, in

24
section 8, another strategy, called the Kelly strategy, was defined in which the
stake is allowed to vary based on the difference in implied probability between
the forecast and the odds on offer. Odds perceived to have greater ‘value’ are
assigned higher stakes. Here, the profit/loss of applying this strategy with the
forecasts defined in this paper is demonstrated. In each case, the stakes are
adjusted so that the average stake is one. This makes the profit/loss directly
comparable with the level stakes strategy which has formed the basis of the
results so far.
The profit/loss of using each GAP rating input along with the Kelly
strategy with maximum BetBrain odds is shown as a function of time in
figure 4. Here, the solid lines represent the profit/loss of the Kelly strat-
egy whilst, for comparison, the dotted lines represent the profit/loss of the
Level Stakes strategy. The results here are striking. In all cases, there is
an improvement in terms of profit, whilst forecasts formed using each of the
measures of attacking performance other than goals result in a profit. This
represents a substantial improvement and highlights the potential for the use
of alternative betting strategies alongside the forecasts produced.

10.2 Average odds

Selected results are now repeated for the case in which average rather than
maximum BetBrain odds are assumed to be available. The cumulative profit
under each GAP rating input both for the Level Stakes and Kelly betting
strategies are shown in figure 5 as a function of time. In contrast to the
maximum odds case, a substantial loss is made, demonstrating that the fore-
casts do not appear to be informative enough to overcome the substantial
overround in this case (around 6.8 percent).
Despite the substantial loss, it is worthwhile to consider the performance
of each of the betting strategies compared to a case in which bets are ran-
domly taken at the same frequency. Similarly to figure 3, the proportion
of random bet scenarios (i.e scenarios in which the same number of bets are
selected as under the Level Stakes betting strategy but randomly rather than
using the forecast probability) in which the cumulative profit is higher than
that achieved by the Level Stakes strategy is shown at the end of each season
in figure 6. Here, despite the fact that a loss is made, the strategy performs
better than would be the case were bets taken randomly. This means that,
whilst the strategy is unable to find enough value bets to make a long term
profit, it is able to identify bets that have a low ‘negative value’ and is there-

25
Figure 4: Cumulative profit/loss of using GAP ratings with each input along-
side the Kelly strategy (solid lines) and the level stakes strategy (dotted lines)
when using maximum odds.

fore successful in reducing losses over time (though in practice it would be

better simply not to bet at all).

10.3 Contrasting Betting Performance Under Maxi-

mum and Average Odds
There is a stark difference between the results achieved when assuming maxi-
mum and average Betbrain odds. In the former case, a substantial and robust
profit is made whilst, in the latter, there is a clear loss. It is worth considering
the implication of this both in terms of the performance of the forecasts and
the efficiency of the market. To make a profit under a value betting strategy

26
Figure 5: Cumulative profit/loss of using GAP ratings with each input along-
side the Kelly strategy (solid lines) and the level stakes strategy (dotted lines)
when using average BetBrain odds.

such as the Level Stakes and Kelly strategies, two essential ingredients are
required. Firstly, there must be odds available in which the probability of the
outcome is higher than the odds-implied probability, that is odds that offer
value. If such opportunities exist, it can be said that the market is inefficient
since a market that takes into account all information would be reflective of
the true probability (plus a profit margin). The second requirement is that
the forecasts are able to identify such opportunities. Someone in possession
of the true probabilities would have the tools to do this and would therefore
easily be able to make a profit over time by only betting on odds that offer
value. In practice, given the inherent limitations in all mathematical models,
the forecasts are not expected to reflect the probabilities of each outcome

27
Figure 6: The proportion of random bet scenarios with higher cumulative
profit as a function of time for each GAP rating input considered when using
average BetBrain odds.

perfectly. Given this, in order to make a long term expected profit, the ex-
pected profit from value bets that are successfully identified must not be
cancelled out by bets that are incorrectly selected and in fact have negative
value.
The fact that a robust profit is made in the maximum odds case implies
that both of the requirements to make a profit are satisfied, that is the market
defined by the maximum BetBrain odds is inefficient and that the forecasts
are informative enough to identify value bets. For the average odds case, the
fact that a loss is made has a profound implication in that it confirms that
the forecasts cannot represent perfect probabilities. This is because, if the
average odds market were efficient, no bets would be taken and the profit/loss

28
would be zero. If the market were inefficient, value bets would be identified
and an expected profit made. The substantial loss can therefore only result
from limitations in the performance of the forecast probabilities. In addition,
since a profit cannot be demonstrated, it is impossible to conclude whether
the average odds market is inefficient or not.
The results above clearly show that the overround in the odds impacts
the profitability of both betting strategies. Whilst using the maximum odds
clearly results in a sustained profit, using the average odds results in a loss. In
fact, there is a substantial difference in the overround of the maximum odds
and the average odds. In the former case, the mean overround is 1.57 per-
cent whilst, in the latter, it is a substantially higher 6.8 percent. In practice,
finding and betting with the maximum odds may be a prohibitively time con-
suming task. However, by carefully choosing a selection of bookmakers and
selecting the best odds among those, a substantially lower effective overround
may be possible than if the average odds were taken. Another potential way
to reduce the overround is to use a betting exchange. Betting exchanges often
have a far lower effective overround than bookmakers, even once commission
is taken into account. For example, if even odds (i.e. decimal odds of 2) are
offered on both outcomes in the over/under 2.5 goal market, a commission
rate of 5 percent such as that charged by Betfair would result in effective
odds of 1.95 on both outcomes, corresponding to an effective overround of
1 1
1.95
+ 1.95 − 1 = 0.026 or 2.6 percent.
The performance of the Level Stakes and Kelly betting strategies in a
setting in which the odds on offer are higher than the average but lower than
the maximum BetBrain odds is now considered. To find odds that are a
weighted average of the two, the following relation is used:
1
r̃ = 1 1 (5)
ω rmax + (1 − ω) rave
i i

where rimax and riave are the maximum and average odds respectively, ω deter-
mines the distance between them, and the resulting odds are r̃. The reason
the odds are inverted first is so that they are weighted in terms of their
odds-implied probabilities. The overall profit from both strategies is shown
as a function of ω in figure 7 under both the Level Stakes strategy (top) and
the Kelly strategy (bottom). Here, as previously shown, for both strategies,
using the average odds produces a loss for all GAP rating inputs. However,
a profit can still be made when the odds are substantially lower than the

29
maximum and this suggests that the forecasts may be effective in achieving
a profit even when the best odds are not available or it is not practical to
compare the odds of a large number of bookmakers.

Figure 7: Total profit as a function of ω for the level stakes (top) and Kelly
(bottom) betting strategies for each GAP rating input.

10.4 Assessing Forecast Accuracy

Whilst the profit made is a useful indication of the performance of the fore-
casts under each GAP rating input, it is also of interest to evaluate the
forecast skill (accuracy). The advantage of this is twofold. Firstly, someone
may wish to use the forecasts for some other purpose, perhaps with some
different betting strategy or even for some different purpose altogether such
as choosing whether to attend a match or not! Secondly, it is interesting to

30
see whether there is general agreement between the relative forecast skill and
betting performance when using each of the GAP rating inputs. The mean
ignorance score of the forecasts under each GAP rating input is shown for
each league in table 7. Here, there appears to be reasonable agreement be-
tween the skill of the forecasts and the betting performance. This is a useful
observation since, if increased forecast skill generally results in larger profits,
some confidence can be had that optimising the parameters with respect to
the ignorance score is a reasonable thing to do. This, in fact, need not be
the case, however. In the two betting strategies outlined in section 8, a profit
is expected to be made when the ‘true’ probability is higher than the odds
implied probability and thus a higher profit will be expected from forecasts
that are able to identify value bets more often. The ignorance score, on the
other hand, favours forecasts that place the most probability on the outcome,
on average. It is therefore the case that, for the Level Stakes betting strategy,
the forecasts need not be the most accurate, they just need to be capable of
successfully identifying enough value bets to make a profit.
There is, in fact, a connection between the ignorance score and gambling
returns. It can be shown that, if in a series of bets, a gambler sets the
proportion of their wealth placed on each outcome to be their estimated
probability distribution, their expected return is 2 raised to the power of the
mean ignorance of the house (note that the ignorance score of the house can
be calculated even if the sum of the odds-implied probabilities do not add up
to one; this, however, gives the house a distinct advantage) relative to that of
the gambler (Roulston & Smith (2002)). For both of the betting strategies in
this paper, however, bets are only taken if the odds offer ‘value’. The efficacy
of both strategies can thus only be tested empirically.

11 Discussion
In this paper, the GAP rating system has been defined and demonstrated as a
basis with which to build probabilistic forecasts. Alongside two value betting
strategies, the resulting forecasts have been shown to be capable of producing
a robust profit in the over/under 2.5 goal betting market using match data as
inputs. At first glance, the results may seem somewhat counterintuitive since
one might have expected the number of goals actually scored in past matches
to be the main factor impacting the number of goals scored in future matches.
In football, however, perhaps more than in many other sports, scoring is a

31
 League G S ST C S+C ST + C
English Premier League −1.36 −1.48 −1.46 −1.52 −1.62 −1.48
English Championship −0.39 −0.39 −0.39 −0.38 −0.58 −0.39
English League One +0.05 +0.12 +0.12 +0.08 +0.04 +0.12
English League Two +0.08 +0.12 +0.11 +0.12 −0.00 +0.12
English National League −0.33 −0.52 −0.52 −0.54 −0.47 −0.52
Scottish Premier League −0.38 −0.62 −0.41 −1.02 −0.63 −0.62
Spanish La Liga −3.08 −3.36 −3.44 −3.23 −3.12 −3.36
French Ligue One −1.37 −1.79 −1.59 −1.84 −1.99 −1.79
Italian Serie A −1.05 −1.22 −1.15 −0.98 −0.86 −1.22
German Bundesliga −2.39 −2.20 −2.17 −2.03 −2.10 −2.22
Combined −0.87 −0.97 −0.94 −0.96 −0.99 −0.97

Table 7: Mean ignorance score for each league, along with the combined
score, when using goals (G), shots (S), shots on target (ST), corners (C),
shots and corners (S+C) and shots on target and corners (ST+C) as the
input to the GAP ratings. The input that results in the best mean ignorance
score is highlighted in green. For clarity, each entry in the table has been
multiplied by 100.

relatively difficult task with a substantial element of chance involved. This

means that, whilst a team may dominate a match, they may simply be
unlucky or have an ‘off day’ in terms of converting chances. Conversely, a
team with relatively few chances may simply get lucky and still manage to
score. If one team has a large number of chances and fails to score while
another has relatively few but still manages to score at least one goal, this
information may not be very informative when predicting the number of goals
that might occur in future matches. A team that has more shots and corners
is usually far more likely to score than those that achieve fewer. It therefore
seems reasonable that shots and corners should provide a better indication
of how the game went than the number of goals scored.
The GAP rating system has a number of similarities with another well
know rating system in the football forecasting literature called the pi-rating
system. These similarities and a number of key differences have been de-
scribed. One key difference is that, unlike pi-ratings, which estimate the dif-
ference in the number of goals scored by two teams, GAP ratings differentiate
between the attacking and defensive capabilities of each team. Consequently,
whilst there is no obvious way of using pi-ratings to create informative fore-
casts for the over/under market (since the expected difference in the number
of goals scored does not give a strong indication of the total number of goals
scored), GAP ratings are ideally suited to this. GAP ratings are also directly
applicable to the match outcome market in which the potential profitability
of pi-ratings has been demonstrated (Constantinou & Fenton (2013)). Given

32
the demonstrated success of using match statistics rather than goals as inputs
for the ratings, an interesting avenue for further research is to test both the
pi-rating and GAP rating systems with different match statistics as inputs
in this market. In fact, this will form the basis of another paper (in prepa-
ration). Preliminary results suggest that both sets of ratings are capable of
yielding a robust profit and that, consistently with the results in this paper,
match statistics other than goals tend to produce far more accurate forecasts
and a larger profit in each case.
It is worth noting that, whilst the forecasts and betting strategies pre-
sented in this paper have been found to perform extremely favourably in that
they produce a clear long term profit in the over/under market, there appears
to be some evidence that the results from the strategy would have worsened
in recent years, since the cumulative profit appears to have ‘levelled off’ (i.e.
the cumulative profit does not appear to have been growing as quickly). It
is not clear why this should be the case and this forms the basis of potential
future work. One possible explanation for this is that, in recent years, the
betting odds have started to take better account of additional performance
data such as shots and corners. This seems like a plausible explanation given
the increase in the availability of in-match data and greater interest in sta-
tistical modelling for sporting events. One explanation that can be ruled out
is an upward shift in the typical profit margin of the bookmakers, which has
in fact fallen in recent years.
The efficacy of GAP ratings for producing forecasts for the over/under
2.5 goal has been demonstrated and the potential to extend their use to the
match outcome market discussed. Testing the performance of the ratings
in these two markets is straightforward because historical odds are easily
available. However, in the modern game, a vast number of other betting
markets are often available. For example, many bookmakers give odds on
whether the total number of goals will exceed 1.5, 3.5, 4.5 etc and it would
be easy to extend the use of GAP ratings to these markets. Other examples
of markets in which the ratings could be utilised include the total number of
corners, the half time score and the number of goals scored by a single team.
There is also significant potential for the GAP rating system to be applied
to other sports, particularly those in which there are well defined statistics
other than goals or points that signify attacking performance. In basketball
or ice hockey, for example, the number of shots in a game could be used
as an input for the ratings. In American football, a measure of attacking
performance could be the number of yards gained by each team, the number

33
of times each team got close to scoring or some other similar measure.
The results in this paper demonstrate that measures of attacking perfor-
mance other than goals can be more informative in predicting future goalscor-
ing performance than goals themselves, so much so that the former can result
in profitable betting strategies whilst the latter does not. Given the wide ar-
ray of match statistics available, there is great potential to come up with
innovative measures of the attacking performance of football teams. The
GAP rating system described provides a solid basis on which to incorporate
these data to build informative and profitable probabilistic forecasts.

34
A Parameter Values
The following tables show the parameter values selected at the beginning of
the stated season for a given GAP rating input.
Shots Goals
Season λ φ1 φ2 Season λ φ1 φ2
2005/06 0.459 0.481 0.566 2005/06 0.395 0.109 0.909
2006/07 0.449 0.449 0.700 2006/07 0.103 0.569 0.490
2007/08 0.446 0.478 0.654 2007/08 0.474 0.144 0.924
2008/09 0.476 0.480 0.689 2008/09 0.478 0.487 0.391
2009/10 0.405 0.477 0.852 2009/10 0.210 0.566 0.439
2010/11 0.426 0.476 0.813 2010/11 0.226 0.495 0.483
2011/12 0.464 0.491 0.660 2011/12 0.158 0.534 0.500
2012/13 0.468 0.491 0.637 2012/13 0.121 0.559 0.504
2013/14 0.451 0.491 0.560 2013/14 0.121 0.519 0.520
2014/15 0.407 0.492 0.568 2014/15 0.424 0.492 0.542
2015/16 0.437 0.496 0.566 2015/16 0.086 0.595 0.528
2016/17 0.445 0.497 0.551 2016/17 0.075 0.534 0.665
2017/18 0.437 0.496 0.566 2017/18 0.078 0.515 0.847
Shots and Corners Shots on Target
Season λ φ1 φ2 Season λ φ1 φ2
2005/06 0.183 0.491 0.739 2005/06 0.176 0.551 0.724
2006/07 0.455 0.454 0.715 2006/07 0.449 0.464 0.735
2007/08 0.450 0.459 0.710 2007/08 0.466 0.471 0.694
2008/09 0.435 0.477 0.671 2008/09 0.450 0.478 0.762
2009/10 0.475 0.483 0.625 2009/10 0.413 0.482 0.868
2010/11 0.466 0.484 0.650 2010/11 0.434 0.475 0.800
2011/12 0.469 0.496 0.603 2011/12 0.500 0.485 0.738
2012/13 0.481 0.492 0.553 2012/13 0.480 0.491 0.734
2013/14 0.422 0.492 0.541 2013/14 0.467 0.490 0.786
2014/15 0.436 0.497 0.543 2014/15 0.478 0.490 0.793
2015/16 0.452 0.499 0.535 2015/16 0.499 0.504 0.804
2016/17 0.425 0.497 0.543 2016/17 0.494 0.513 0.816
2017/18 0.436 0.493 0.546 2017/18 0.500 0.504 0.796

35
 Shots on Target and Corners Corners
Season λ φ1 φ2 Season λ φ1 φ2
2005/06 0.163 0.510 0.795 2005/06 0.173 0.468 0.818
2006/07 0.445 0.463 0.723 2006/07 0.464 0.495 0.763
2007/08 0.407 0.460 0.779 2007/08 0.405 0.474 0.771
2008/09 0.434 0.475 0.662 2008/09 0.388 0.482 0.640
2009/10 0.423 0.484 0.622 2009/10 0.321 0.485 0.572
2010/11 0.445 0.484 0.653 2010/11 0.366 0.503 0.581
2011/12 0.432 0.495 0.614 2011/12 0.342 0.498 0.571
2012/13 0.444 0.503 0.558 2012/13 0.311 0.491 0.548
2013/14 0.480 0.491 0.567 2013/14 0.348 0.493 0.552
2014/15 0.422 0.496 0.560 2014/15 0.426 0.488 0.547
2015/16 0.432 0.509 0.541 2015/16 0.297 0.487 0.525
2016/17 0.441 0.517 0.557 2016/17 0.297 0.490 0.538
2017/18 0.490 0.519 0.552 2017/18 0.426 0.488 0.547

GAP rating input α β1 β2

Goals −2.0629 −0.0026 +4.0918
Shots −2.3009 +0.0070 +3.8969
Shots on Target −2.1576 +0.0086 +3.9135
Corners −2.4001 +0.0176 +3.9910
Shots and Corners −2.4287 +0.0070 +3.8912
Shots on Target and Corners −2.3010 +0.0081 +3.8815

Table 8: Logistic regression parameter values for each GAP rating input,
calculated using all available years of data.

Acknowledgements
I am indebted to two anonymous reviewers and a handling editor who pro-
vided invaluable feedback on earlier versions of this paper.

References

36
References
Baker, R. D. & McHale, I. G. (2015), ‘Time varying ratings in association
football: the all-time greatest team is..’, Journal of the Royal Statistical
Society: Series A (Statistics in Society) 178(2), 481–492.

Berrar, D., Lopes, P., Davis, J. & Dubitzky, W. (2019), ‘Guest editorial:
special issue on machine learning for soccer’, Machine Learning 108(1), 1–
7.

Bonferroni, C. (1936), ‘Teoria statistica delle classi e calcolo delle prob-

abilita’, Pubblicazioni del R Istituto Superiore di Scienze Economiche e
Commericiali di Firenze 8, 3–62.

Boshnakov, G., Kharrat, T. & McHale, I. G. (2017), ‘A bivariate weibull

count model for forecasting association football scores’, International Jour-
nal of Forecasting 33(2), 458–466.

Bröcker, J. & Smith, L. A. (2008), ‘From ensemble forecasts to predictive

distribution functions’, Tellus A 60(4), 663–678.
URL: http://dx.doi.org/10.1111/j.1600-0870.2008.00333.x

Carbone, J., Corke, T. & Moisiadis, F. (2016), ‘The rugby league prediction
model: Using an elo-based approach to predict the outcome of national
rugby league (nrl) matches’, International Educational Scientific Research
Journal 2(5), 26–30.

Cintia, P., Giannotti, F., Pappalardo, L., Pedreschi, D. & Malvaldi, M.

(2015), The harsh rule of the goals: Data-driven performance indicators for
football teams, in ‘Data Science and Advanced Analytics (DSAA), 2015.
36678 2015. IEEE International Conference on’, IEEE, pp. 1–10.

Constantinou, A. C. (2019), ‘Dolores: a model that predicts football match

outcomes from all over the world’, Machine Learning 108(1), 49–75.

Constantinou, A. C. & Fenton, N. E. (2013), ‘Determining the level of ability

of football teams by dynamic ratings based on the relative discrepancies
in scores between adversaries’, Journal of Quantitative Analysis in Sports
9(1), 37–50.

37
Constantinou, A. C., Fenton, N. E. & Neil, M. (2012), ‘pi-football: A
bayesian network model for forecasting association football match out-
comes’, Knowledge-Based Systems 36, 322 – 339.

Constantinou, A. & Fenton, N. (2017), ‘Towards smart-data: Improving

predictive accuracy in long-term football team performance’, Knowledge-
Based Systems 124, 93–104.

Cronin, B. (2019), ‘How important are the first six games of the
season?’, https://www.pinnacle.com/en/betting-articles/Soccer/
importance-of-first-six-games-in-soccer/J3R2FJ6DY6MBZVW2. Ac-
cessed: 18/04/2019.

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), 265–280.

Dixon, M. J. & Pope, P. F. (2004), ‘The value of statistical forecasts in the

uk association football betting market’, International journal of forecasting
20(4), 697–711.

Eggels, H. (2016), Expected goals in soccer: Explaining match results using

predictive analytics, in ‘The Machine Learning and Data Mining for Sports
Analytics workshop’, p. 16.

Elo, A. E. (1978), The rating of chessplayers, past and present, Arco Pub.

ESPN (2018), ‘Espn’s nba basketball power index playoff odds’,

http://www.espn.co.uk/nba/story/_/page/BPI-Playoff-Odds/
espn-nba-basketball-power-index-playoff-odds. Accessed:
07/07/2018.

Fifa (2018), ‘Revision of the fifa / coca-cola world

ranking’, https://resources.fifa.com/image/upload/
fifa-world-ranking-technical-explanation-revision.pdf?
cloudid=edbm045h0udbwkqew35a. Accessed: 27/04/2019.

FiveThirtyEight (2019a), ‘The complete history of the nfl’, https:

//projects.fivethirtyeight.com/complete-history-of-the-nfl/.
Accessed: 29/03/2019.

38
FiveThirtyEight (2019b), ‘Nba elo ratings’, https://fivethirtyeight.
com/tag/nba-elo-ratings/. Accessed: 29/03/2019.
Football Data (2018), ‘Football data’, http://www.football-data.co.uk/.
Accessed: 09/10/2018.
Goddard, J. (2005), ‘Regression models for forecasting goals and match
results in association football’, International Journal of forecasting
21(2), 331–340.
Good, I. J. (1952), ‘Rational decisions’, Journal of the Royal Statistical So-
ciety: Series B 14, 107–114.
Huang, K.-Y. & Chang, W.-L. (2010), A neural network method for predic-
tion of 2006 world cup football game, in ‘Neural Networks (IJCNN), The
2010 International Joint Conference on’, IEEE, pp. 1–8.
Hubáček, O., Šourek, G. & Železnỳ, F. (2019), ‘Learning to predict soccer
results from relational data with gradient boosted trees’, Machine Learning
108(1), 29–47.
Hucaljuk, J. & Rakipović, A. (2011), Predicting football scores using ma-
chine learning techniques, in ‘MIPRO, 2011 Proceedings of the 34th Inter-
national Convention’, IEEE, pp. 1623–1627.
Hvattum, L. M. & Arntzen, H. (2010), ‘Using elo ratings for match re-
sult prediction in association football’, International Journal of forecasting
26(3), 460–470.
Joseph, A., Fenton, N. E. & Neil, M. (2006), ‘Predicting football results
using bayesian nets and other machine learning techniques’, Knowledge-
Based Systems 19(7), 544–553.
Karlis, D. & Ntzoufras, I. (2003), ‘Analysis of sports data by using bivariate
poisson models’, Journal of the Royal Statistical Society: Series D (The
Statistician) 52(3), 381–393.
Kelly Jr, J. (1956), ‘A new interpretation of the information rate’, Bell Sys-
tem Technical Journal 35, 917–926.
Lee, A. J. (1997), ‘Modeling scores in the premier league: is manchester
united really the best?’, Chance 10(1), 15–19.

39
Maher, M. J. (1982), ‘Modelling association football scores’, Statistica Neer-
landica 36(3), 109–118.

Opta (2018), ‘Assessing the performance of premier league goalscorers’,

https://www.optasportspro.com/about/optapro-blog/posts/2012/
blog-assessing-the-performance-of-premier-league-goalscorers/.
Accessed: 09/10/2018.

Opta Football (2018), ‘Opta Football’, https://www.optasports.com/

sports/football/. Accessed: 09/10/2018.

O’Donoghue, P., Dubitzky, W., Lopes, P., Berrar, D., Lagan, K., Hassan,
D., Bairner, A. & Darby, P. (2004), ‘An evaluation of quantitative and
qualitative methods of predicting the 2002 fifa world cup’, Journal of Sports
Sciences 22(6), 513–514.

Rathke, A. (2017), ‘An examination of expected goals and shot efficiency in

soccer’.

Roulston, M. S. & Smith, L. A. (2002), ‘Evaluating probabilistic forecasts

using information theory’, Monthly Weather Review 130, 1653–1660.

Rue, H. & Salvesen, O. (2000), ‘Prediction and retrospective analysis of

soccer matches in a league’, Journal of the Royal Statistical Society: Series
D (The Statistician) 49(3), 399–418.

Suznjevic, M., Matijasevic, M. & Konfic, J. (2015), Application context based

algorithm for player skill evaluation in moba games, in ‘2015 International
Workshop on Network and Systems Support for Games (NetGames)’,
IEEE, pp. 1–6.

The Economist (2017), ‘How data changed gambling - the economist ex-
plains’, https://www.economist.com/the-economist-explains/2017/
07/19/how-data-changed-gambling. Accessed: 07/10/2018.

Van Haaren, J. & Van den Broeck, G. (2015), Relational learning for football-
related predictions, in ‘Latest Advances In Inductive Logic Programming’,
World Scientific, pp. 237–244.

40