Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.

pl
Data: 29/07/2019 20:07:44

Annales UMCS Informatica AI XI, 1 (2011) 7–13

DOI: 10.2478/v10065-011-0027-2

CS
A strategy in sports betting with the nearest neighbours search
and genetic algorithms

Damian Borycki1∗
1
Jagiellonian University, Faculty of Physics, Astronomy and Applied Computer Science,
Reymonta 4, 30-059 Kraków, Poland
M
Abstract – The point of sports betting is not merely to correctly predict the outcome of a game, but
to actually win on a bet. We propose a model of sports betting that uses the nearest neighbours search
and genetic algorithms to do the job. It uses data on the teams playing, their respective formations,
individual players, results of previous games, as well as odds offered by bookmakers. The model has
been trained using the data from the seasons 2002/03 until 2008/09 of the English Premier League
U
and tested against the already played games of the seasons 2009/10 and 2010/11.

1 Introduction
Sports betting is becoming more and more popular, as evidenced by an increasing
number of registered players in the online bookmakers. Interest in this area is also
growing in the scientific community. The literature on forecasting results of sporting
events is vast.
Of all sports, football enjoys the greatest interest in Europe. Many papers on pre-
dicting results in this discipline have been published, see for example [1, 2, 3]. [3]
presents a model that uses the offensive and defensive strengths of the teams partici-
pating in the game to predict the result, with the data on previous games providing the
input. The model presented in [2] uses fuzzy logic optimized by a genetic algorithm
and a neural network. The results of the last five games of each team and two direct
matches between them, to a total of 12 games, serve as the input. The model presented
in [1] is based on the rule-based reasoning and Bayesian networks. In addition to the
historical data on previous games, it also uses expert knowledge to construct the priors.

∗
[email protected]
Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.pl
Data: 29/07/2019 20:07:44
8 A strategy in sports betting with the nearest neighbours...

All these models focus on correctly predicting the result of a game, without paying
attention to the financial outcome of a bet. This may be misleading since the payoff
of a string of “obviously winning” bets can be easily offset by a couple of unfortunate
ones.
The model presented in this paper uses much more data than just the historical re-
sults. In contrast to [3] it does not use any expert knowledge or subjective information.
The principal novelty of the present approach is that it tries to optimize the actual
profits from betting, not just correctly predict results of a series of games.

CS
2 Data
Analyzing the chances of Team A winning against Team B, it is important to take
into account as many factors that may influence the result as possible. Analyzing
the results of previous games is not enough. For example, players change teams and
somebody who played for Team A in 2005 may be playing for Team B in 2010. For
this reason, the analysis should include information on individual players. It is also
important to consider which players actually participate in a given game, whether in
M
the starting lineup or on the bench, with a stress on the starting lineup. Random and
unexpected weaknesses of various formations also provide valuable information, as well
as the current disposition of the whole team, as judged from the results of a set of
games immediately preceding the current one, results of direct matches between both
teams etc. Humans consider all sorts of such information before placing a bet, and a
betting algorithm should do the same.
U
In the model presented here, each game is represented by 134 parameters. To test
the model, all games from the seasons 2002/03 until 2009/10 of the English Premier
League, or the total of 3040 games, have been analyzed.

3 The model
The basic strategy employed in our model is the nearest neighbours search [4]. The
probability of winning a bet is calculated on the basis of how many “similar” bets,
found by the KNN search, have been won. The objective is not to predict the result
of a game, but to optimize the profits from the actual bet.

Bets
The bets that our model admits are:
• t1 –home win,
• tx – tie,
• t2 – away win,
• tover – total goals above 2.5,
• tunder – total goals under 2.5.
The model consists of four modules:
Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.pl
Data: 29/07/2019 20:07:44
Damian Borycki 9

Forecasting
This module calculates the probability of winning each of the bets considered. We
use the following notation:
m = [a1 , 22 , . . . , a134 ] is the game, the result of which we try to predict. a1. . . 134 are
the explanatory variables
M = {m : m = [a1 , 22 , . . . , a134 , y], is the set of all matches completed before. y is
the result of the match
Mi is the set of ith attributes of matches from the set M
ni = max(Mi ) − min(Mi )
wi – the weightith attribute (to be discussed in Section 3.4)

CS
dm – the distance of m from m

134
X |am − am |
i i
dm = (1)
i=1
ni

dm serves as a metric in the KNN algorithm, with the size of the neighbourhood, k,
set at 15.
M
Decision-making
After calculating the probability of each of the five types (t1 , tx , t2 , tover , tunder ) this
module decides whether a particular type of bet should be placed or not. This decision
depends on whether the type is worth the risk. Here we use the following notation:
forecastt – probability of winning the bet t resulting from the forecasting
bookieOddt – odds on bet t set by the bookmaker
U
forecastOddt – odds on bet t resulting from the forecasting
1
forecastOddt = forecast t
Under the assumption that we have correctly predicted the result of the game, the
bet t is worth the risk if
bookieOddt ≥forecastOddt
However, as the forecast might not be perfect, a bet is placed only if
bookieOddt ≥ 1.5∗forecastOddt

Assessment
This module is designed to evaluate the performance of the algorithm at a fixed time
interval. To this end each match with the time interval is subjected to the typing. Then
the suggested types are compared with the actual results. On this basis the values of
yield and maxloss are determined. We use the following notation:
T – the set of types with a fixed time interval
T = [t1 , t2 , . . .], time(ti ) ≤time(ti+1 )
#T – the(number of types in the set T
bookieOddt − If the type was successful, t ∈ T
win(t) =
0 otherwise
Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.pl
Data: 29/07/2019 20:07:44
10 A strategy in sports betting with the nearest...

#T
P
win(T ) = win(ti )
i=1
yield(T ) = win(T#T
)−#T

maxloss(T ) – length of the longest substring of zeros in the sequence win(t1 ),

win(t2 ), . . . ,win(t#T ).
In the end, the algorithm is assessed with
 #T
yield(T )
f (T ) = 1 + .
maxloss(T ) + 1
For example, let us

CS
• invest 100 units
• #T = 20
• maxloss(T ) = 9
• yield(T ) = 0.2.
Because maxloss > 0, it follows we cannot invest all available capital in each type
because we will eventually lose. We should not invest 1/9 of the capital, either, because
keeping stakes after each loss we may also go bankrupt. We invest 1/10 of the available
capital in each type. The yield shows that, on average, we gain 20% of the funds
M
invested. If we substitute these to the formula above, we can see that after investing
in 20 types, we gain, under the assumptions specified, 0.0220 ≈ 1.49 of our initial
investment. Our net gain is therefore 49 units.

Optimization
Optimization is crucial for a good performance of the algorithm presented here. The
U
point of the optimization is an appropriate selection of data used to calculate the
distance. The optimization is based on a genetic algorithm [5].
Coding: A chromosome consists of 134 genes. Each gene of the chromosome bears
the information on the explanatory variable it represents, and its weight, w, set to
either 0 or 1. The weight w = 0 means that this particular explanatory variable is not
used in calculating the distance.

Selection was based on the roulette method.

Crossing: Crossing occurs with the probability of 70% and if so it does on a randomly
chosen site.
Mutation: Weights can switch randomly from 0 to 1 and vice versa with the proba-
bility 0.1%.
The fitness function is the same f (T ) that was used in the Assessment module.
The size of each generation was taken to equal 10 and there were 45 generations.
Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.pl
Data: 29/07/2019 20:07:44
Damian Borycki 11

4 Results
Finding the optimal strategy
The model was tested three times against the results of the English Premier League,
the seasons 2007/08 and 2008/09.
Fig. 1 shows the average fitness of each of the 45 generations in each of the three
tests. Note the logarithmic scale of the vertical axis. One can see that starting with
20 generations, the tests bring the returns of the whole generation. One can also see
how quickly the model improves. Fig. 2 shows the fitness of the best chromosome in
each generation. A strategy that gives profits has been found already after the 4th

CS
generation of the genetic algorithm. It can be also seen that the algorithm gets stuck
in a local optimum. The fitness of the best strategy identified during the test equals
843.
Verification
Very promising results obtained during the test have been verified against the games
of the season 2009/10, for 7 best strategies only. These results are presented in Table
1. The best strategies identified during the tests were actually the worst, but some
strategies still brought profits. It appears that in-sample training, or optimization of
M
strategies to the results of the seasons 2007/08 and 2008/09 performed poorly out-of-
sample, as some external conditions might have changed.
The four strategies that brought profits for the season 2009/10 have been further
verified against the games of the 2010/11 season that have been completed until January
1, 2011. The results are presented in Table 2. These four strategies also profit in the
U
2010/11 season.

Fig. 1. The average value of the match in the generation.

Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.pl
Data: 29/07/2019 20:07:44
12 A strategy in sports betting with the nearest...

CS
Fig. 2. The highest value of matches in the generation.
M
Table 1. Verification the 2009/2010 season.
U

Table 2. Verification season 2010/2011.

5 Conclusions
An original, fully automatic algorithm for investing in sports betting has been pre-
sented. Unlike most previous studies on similar subjects, the presented model uses a
Pobrane z czasopisma Annales AI- Informatica http://ai.annales.umcs.pl
Data: 29/07/2019 20:07:44
Damian Borycki 13

lot more publicly available information on football matches. It also focuses not merely
on correctly predicting results of games, but rather on profiting from actual bets. If a
bet is deemed not to be profitable, it is not taken, and the gambler does not sustain
losses, at the price of abstaining from minute winnings. Our results show that, contrary
to the popular opinion, a profitable strategy in investing in sporting bets is possible.
Several possibilities of improving our model appear. For example, introducing the
non-Euclidean metric, separate strategies for different types of bets and changing the
weights from 0-1 to fractional values can help identify even more profitable strategy.
Several programming issues need to be solved in order to run the appropriate software
faster. Eventually we want to be able to use our model in different leagues and in

CS
sports other than football.

References
[1] Min B., Kim J., Choe C., Eom H., McKay R. I., A Compound Framework for Sports Prediction:
The Case Study of Football, Knowledge-Based Systems, 21(7) (2008): 551.
[2] Rotshtein A. P., Posner M., Rakityanskaya A. B., Football Predictions Based on a Fuzzy Model
with Genetic and Neural Tuning, Cybernetics and Systems Analysis 41(4) (2005): 619.
M
[3] Rue H., Salvesen O., Prediction and retrospective analysis of soccer matches in a league, The
Statistician 49(3) (2000): 399.
[4] Larose D. T., Discovering Knowledge in Data. An Introduction to DATA MINING, Chapter 5.
[5] Larose D. T., Data mining methods and models, Chapter 6.
U
)
rg
f.o
pd
.tc
w
w
(w
F
PD
TC
by
ed
er
w
Po