International Journal of Mathematical and Computational Methods

K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

Football Match Statistics Prediction using Artificial Neural Networks

1
K. Sujatha, 1T. Godhavari and 2 Nallamilli P G Bhavani
1
Dept of Electrical and Electronics Engineering/CSE Dept, Center for Electronics Automation and Industrial
Research (CEAIR), Educational and Research Institute University, Chennai, India,
[email protected]
2
Electrical and Electronic Engineering, Educational and Research Institute University, Chennai, India,

Abstract-The predictions of the outcomes of football (American soccer) matches are widely done using the if-else
case based Football Result Expert System (FRES). The proposed prediction technique uses a neural network approach
to predict the results of football matches. The neural network detects patterns from a number of factors affecting the
outcome of a match making use of historical cases of training. This paper describes the inputs, outputs and compares
the results of this kind of a system.
Keywords: - Artificial Neural Networks, Back propagation, sport result prediction, pattern prediction, FRES,
Activation function

1 Introduction 2 Input factors considered

The input factors are considered for a Bundes Liga
To predict the outcome of a match between two teams, a (German League) match between FC Bayern Munich
person will generally take into account certain factors and FC Borussia Dortmund. Training data available is
like the recent performances of the teams, whether the for all meetings between the two sides in the Bundes
match is going to be played at home or away, recent
Liga during the period 2005 to 2011, making a total of
player transfers, recent coach and staffing changes, etc.
14 matches. The various inputs considered and their
This is what human 'intuition' is made up of, a decision
based on a combination of solid tangible facts available effects on the outcome of the match are as follows:
to the person making the decision. The problem when a 2.1 Transfer money spent
human predicts the outcome of a match is that the
Generally, twice during a season (during the transfer
persons decision will be influenced by factors like the
window) the teams are allowed to buy or sell players to
humans' preference in teams, the humans' perception of
change the makeup of a team. The indication is that if a
certain players in a team, some studies show decisions
team spends more money on transfers then the team is
may even be influenced by the color of the kit of the
looking to improve its bench strength or team makeup.
team. [1]
Historically the performance of a team generally tends to
One of the methods widely used for predicting match improve after spending during the mid-season transfer
outcome is the Football Result Expert System used window. Money spent during the transfer season is
widely by media and book makers. The problem with directly normalized as shown and used as an input to the
FRES is that it uses a rule based approach or an if-else neural network.
statement based approach that considers and rejects
Generally, the first step for normalization involves
various factors one by one.
applying a median filter to normalize data as in equation
The approach described in this paper takes advantage of 1. [2] This step was not performed for the simple reason
the fact that neural networks are good at recognizing that there were not enough training sets available to allow
patterns and mapping these patterns to outputs. The for this process of elimination of outlying data. This lead
approach proposed in this paper uses a number of factors, to certain mid-range values getting squashed, but was
some which may even not be available to a standard acceptable since the two teams in consideration, Borussia
human decision maker as inputs to the neural network. Dortmund and FC Bayern Munich generally tended to be
The neural network is given a number of previous among the top spenders.
performances of the teams in question as a training set
The formula used for normalization is
after which it is used to predict a match in the future.
v' = (v – average) / SD (1)

ISSN: 2367-895X 1 Volume 3, 2018

 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

where v' – normalized value, v – previous value, average cup. Care should be taken to ensure that only records
– average of the training set, SD – standard deviation of pertaining to that league are considered as the number of
the set. matches played may otherwise tend to be different for
the two teams.
2.2 UEFA co-effecient
2.5 Team cost
UEFA is the European football governing body that
organizes the annual Champions League in which all the Each player on the team is given a figure that is
top European teams appear. Even though the Champions indicative of how much the club would have earned
league is not related to the German Bundes Liga, this from the sale of that player to another club. [4]
rank has been taken into consideration since both Bayern The sum of all the individual player costs gives the team
Munich and Borussia Dortmund have been regular cost. These values were found to vary between 2005 and
fixtures in the league since 2005 [3]. UEFA rank is given 2012, reflecting changes in the value of the Euro.
by the "Champions league point system". A certain
number of points are awarded to a team at each stage of 2.6 Year of match
the Champions league as given in Table 1. The inclusion of year of match is found to improve
the performance and the prediction capabilities of
Table 1: Champions league point system
the neural network. This is analogous to a human
S.no Stage Points earned being coming to the conclusion that the recent
1 1st qualifying round elimination 1 performance of the team is more important than past
2 2nd qualifying round elimination 2 performances. In simple terms Borussia Dortmund
3 Group stage participation 4 is on a rising trend with more and more victories
4 Group stage win 2
5 Group stage draw 1 than Bayern Munich while in the past Bayern
6 Round of 16 participation 4 Munich was the superior team. [5]

The neural network was amazingly able to pick up

Since 2009 onwards, an additional point is awarded to a on this trend and predict more accurately the recent
team taking part in the round of 16. Eliminated teams are matches played out between the two teams.
not awarded points since they move to the second tier
Europa league and earn points there.
Normalization of the year of match is easy since the
upper and lower bounds are fixed with a difference
The above system is used for awarding points to a team between the maximum and minimum value less
taking part in the Champions league. The total team co- than ten. For future predictions, the value used for
efficient is given by the sum of points earned by the team normalization will have to be changed appropriately
in the past five years plus 33% of the association co-
efficient during the same period. This results in the 2.7 Wins and losses
UEFA coefficient of a team increasing with time. Home and away wins, home and away losses and home
and away draws are very vital inputs to the neural
2.3 League rank
network. These statistics are to be considered for the
The position of the team in the league standings is an league upto the point of the match. Home and away
indication of their performance so far, and as such is performance are considered separately since there was
used as one of the inputs to the neural network. The found to be a difference in the home and away
number of teams in the Bundes Liga has remained performance of a team, with a team generally performing
constant at 18. Normalization is done by simply dividing better in home matches than in away matches [6].
position of team in the league by 18.
2.8 League points
2.4 Goals scored and conceded
League points awarded in the Bundes Liga are as shown
Goals scored are used as a number of inputs to the neural in Table 2.
network. The categories these come under include:
 Away goals scored – Goals scored by a team Table 2: League point award system
playing away fixtures
 Away goals conceded – Goals conceded by a S.no Number of points Result
team playing in away fixtures. 1 3 points Win
 Home goals scored – Goals scored by a team in 2 1 point each Draw
home fixtures 3 0 points Loss
 Home goals conceded – Goals conceded by a
team playing home matches. The points are used to rank the team within the league.
Goals scored either home or away have to be considered The final standings in the league are given in decreasing
for all matches prior to that meeting in the league. It was order of points earned by a team, meaning the team that
found that most sites kept a record of goals scored in all
wins the maximum points takes the top position in the
formats including the Champions league and the German
ISSN: 2367-895X 2 Volume 3, 2018
 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

league while the team with the minimum number of

points takes the bottom position in the table. Points The resulting predictions therefore are not clear cut true
earned till the match day in consideration is used as one and false cases (or 1s and 0s), but instead give a value
of the neural network inputs. between 0 and 1 which when multiplied by 100 gives the
2.9 Home advantage prediction of a team A or a team B win. For all purposes,
the team showing the greater win percentage can be
Home advantage is given by the last two nodes. These thought of as the ultimate winner.
nodes are used to indicate which team has the home
ground advantage. For example, if the match is being 3.2 Goals scored
played at home for team A, the input for 'home ground
A' node will be 1 and correspondingly will be 0 for Goals scored are another two outputs of the neural
'home ground B' input node. network. The neural network is expected to predict the
goals scored based on past patterns resulting in certain
2.10 Factors rejetced number of goals scored. The performance of the neural
A number of factors that, if were taken into network for prediction of goals scored was found to be
consideration would improve the performance of the as good as a human guess. That is, close at best and way
predictor, had to be rejected as being too impractical of mark at worst.
included the following:
 Possession – ball possession by a team is generally In cases where the win percentage of the two teams was
an indication of the performance of the team as a close to one another (indicative of a drawn match), the
whole. goals scored would often be the same for both teams
 Passing – passing gives an indication of how well indicating that the neural network is able to make
the team is performing as a unit associations between goals scored and other factors such
 Clean sheets – Clean sheets are credited a team that as previous goals scored and conceded and previous
hasn't conceded any goals in a match. The number wins and losses.
of clean sheets in a league prior to a match gives a 3.3 Bookings
fair indication of that teams' defense performance. The ability of the neural network to predict the number of
 Bookings – Bookings are of two types, yellow or red and yellow cards shown to each team was found to be
red. A yellow card is shown for a minor infraction, close at best. The reason for the neural network
while two yellow cards in a match amount to a performing so badly at predicting the number of cards is
player being red carded whereupon he will no because the number of cards depends on a lot of
longer be allowed to participate in the match and his intangible factors that cannot be predicted like, player
team will be reduced to ten players. A team which temperament, referee temperament and match day ground
has a large number of bookings may indicate an conditions.
inferior team. The neural network used was the standard back
propagation neural network as explained below. A few
The inclusion of the above factors would have benefited features were used to enhance the performance of the
the prediction capability of the neural network but had to network such as the addition of logarithmic and
be rejected since records of these values could not be exponential input and output neurons. In practice, any
found even upon considerable searching. kind of neural network algorithm can be used to model a
predictor, but back propagation was found to provide the
3 Output of the network best controllability in terms of rate of learning and
More than predicting the winning team, the network also generalization capabilities of the neural network.
attempts to predict a number of other factors given
below. This was met with a certain degree of success as
illustrated in the results section re-enforcing the idea that
a neural network is as if not more capable than a human
at providing unbiased predictions and at detecting
complex patterns.

3.4 Backpropogation Algorithm

A back propagation neural network was written in
3.1 Match Winner MATLAB for a single hidden layer. The inputs are
Match winner is provided by the neural network as a assumed to be stored in a column wise matrix, and the
percentage. One output provides the win percentage of outputs of the training set likewise in a column wise
team A and another separate output the win percentage of matrix. The number of nodes in the hidden layer is the
team B. For training, if team A has won the match, the same as the number of input layer nodes. The BPN
input team A will be 1 while the input team B will be 0 performs updating operations taking the entire input
and vice-versa.
ISSN: 2367-895X 3 Volume 3, 2018
 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

matrix into consideration. The algorithm for back by Y '  Oh  d ' . Where d is given
propagation is as follows:
by d  (TO  OO )  OO  (1  OO ) .
Step 1: Inputs and outputs are normalized as Step 14: 'delW' is updated using the formula given as
explained above. Total number of inputs to delW  ( momentum  delw)  (learningRate  Y )
the network is given by l = iO×3 and number Step 15: The complete training set 'nSet' error is
of outputs is given by n = jO×3, where iO and calculated as e  w  d .
jO are the number of inputs and outputs. The Step 16: 'X' is calculated as X  Oi  dStart ' .
number of hidden layer neurons m equals the
number of input neurons i.e. m = l. Where dStar  e  Oh  (1  Oh ) . (5)
Step 2: Let 'nTest' be the number of training sets. Step 17: Change in input layer weights is given by
This means the size of the input matrix will delv  ( momentum  delv )  (learningRate  X )
be lxm and the size of the output matrix will and weights are adjusted as v  v  delv
be nxm. and w  w  delw .
Step 3: The input and output pattern are stored in Step 18: Repeat steps 9 to 18 until the error rate is
rows of the input and output matrices leaving lesser than tolerance value. Save the weights
two places for the augmented neurons. Every and exit.
second row of the input and output matrix is
given by computing the ln (loge) of the The given algorithm explains the learning algorithm
preceding row to give the logarithmic neuron. used and how the weights have been adjusted.
Every third row of both matrices is computed
by the exponential of the original pattern. 3.5 Number of hidden nodes
Step 4: Assign learning rate and momentum factor to The number of nodes in the hidden layer is generally
some initial value. kept to the minimum required to classify all input pattern
Step 5: Initialize the input layer – hidden layer weight areas accurately. This is done to keep the memory
matrix v (lxm) and the hidden layer – output requirement of the nodes to a minimum. The number of
layer weight matrix w (mxn) to some random separable regions in space is one less than the number of
values. hidden nodes. It is also argued that the number of hidden
Step 6: Let the thresholds given by delv and delw, nodes should be one less than the number of training
both be zero matrices initially. sets.
Step 7: Variable 'iterate' is used to store the number
of iterations that training is going to take Based on this, the number of training sets is found to
place for. be13 for matches played between 2005 and 2011,
Step 8: Since the input neurons use a linear activation meaning that the number of hidden nodes should be 12
function, output of input layer 'Oi' is made (given as 13-1). But, since the number of inputs being
equal to input to input layer 'Ii' for each passed to the network is 29, or 87 taking into account
pattern (stored as a column). logarithmic and exponential neurons, the number of
Step 9: Input to the hidden layer is calculated by hidden nodes is kept at 87 to ensure separability of each
multiplying the output of the input layer with region within the training set.
corresponding weight values. That
is, I h  v ' Oi , where 'Ih' represents the input General rule of thumb dictates that the number of hidden
to the hidden layer and is a column matrix of layer nodes for a single hidden layer feed forward
length m. network equals the mean (rounded up) of the number of
Step 10: Hidden layer outputs 'Oh' are calculated using input layer and output layer nodes.
the sigmoidal function as shown- 3.6 Logarithmic and exponential neurons
1
OO  (2) Logarithmic and exponential neurons are generally used
1  e Io to augment the training capabilities of the network. For
Step 11: Target output 'To' (nx1) is calculated from the every one input neuron, its logarithmic and exponential
output matrix by taking the appropriate neurons are computed as shown and passed as additional
column. inputs to the network. On the output side also, the
Step 12: Error is calculated in two steps. First, the part logarithmic and exponential neurons are calculated and
error ePart is calculated as- used for every output neuron.
ePart   (TO  OO )
2 (3)
These neurons are collectively termed as augmented
Final error is given as Root Mean Square neurons. The addition of these neurons provides accurate
ePart (4) boundary mapping and increases the speed of training.
(RMS) value of ePart or ERMS  . These increase the accuracy of the neural network at the
n
cost of the generalization capabilities.
Step 13: Calculated output 'Y' (mxn) is given
ISSN: 2367-895X 4 Volume 3, 2018
 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

Though these neurons play no role in the actual inputs borders will be clearly defined and fresh data that falls
and outputs of the network, they are used to modify the just outside a border will be disregarded even though it
interconnecting weights of the neural network and may belong to that data set.
therefore improve training.

4 Performance Parameters
4.1 Input nodes
The list of inputs given to the neural network (given
separately for team A and team B):
 Transfer money spent A, B
 UEFA ranking A, B
 League position A, B
 Away goals scored A, B
 Home goals scored A, B
 Away goals conceded A, B
 Home goals conceded A, B
 Player cost A, B
 Year of match
 Away wins A, B
 Home wins A, B
 Away losses A, B
 Home losses A, B Figure 1: Comparison of learning rates
 Total draws A, B
 League points A, B
 Home ground A, B
The above given 29 inputs are used with the network.
Addition of augmented neurons gives a total of 87 input
neurons.
4.2 Output nodes
The following are the outputs of the neural network:
 Match winner A, B
 Goals scored A, B
 Yellow cards A, B
 Red cards A, B
The total number of bookings or cards shown for each
team is given by the sum of the corresponding teams red
and yellow cards.
4.3 Selection of learning rate Figure 2: Comparison of various momentum factors
Experimentally, the optimal learning rate is found to be
0.7. As shown below, a learning rate of 0.7 provides the The ideal solution involves classifying all training data
best convergence of error rate in 500 iterations. with minimum error, while at the same time maintaining
4.4 Selection of momentum factor a minimum error for fresh data. In order to decide when
to stop training, 3 out of 13 training patterns are used as
When momentum factor is zero, the neural network the control group. Training is done on the remaining 10
takes 854 iterations to achieve a convergence of 0.02. patterns while error is calculated at the end of each set of
The most suitable momentum factor which does not weight iterations for the 3 control patterns. It is found
make the neural network overshoot its minima is found the error value for the control set initially decreases
to be 0.9. The figure given below compares the error rate along with decrease in overall error of the training set,
for different values of momentum factor. but starts to increase at a certain point. This point is the
minimum error for the control pattern after which it has
4.5 Generalisation started to lose its generalization capabilities.
Often the problem with training for a number of fixed
iterations or training with the goal of reducing error is The 3 control patterns are selected randomly from the
that the neural network will lose its generalization given input test patterns, choice of different control
capability. That is, since the error will now be small, the
ISSN: 2367-895X 5 Volume 3, 2018
 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

patterns is found not to noticeably affect the neural case, the neural network has to be trained separately for
network. such matches using suitable historical test cases, so that
the neural network makes the association for a reduced
goal margin.

This method of prediction is also appropriate for

matches in other ball sports. For example, such a model
is readily applicable to predict the outcome of a hockey
match with more or less identical inputs and outputs.
5.2 Matches predicted with actual result

The predictions of goals scored by each team by the

neural network have been compared against actual
results in the table given below.

Figure 3: Error rate of training set for 65 iterations Table 3: Predicted results Vs actual results

Date Predicted result Actual result

5 Prediction Capabilities Munich Dortmund Munich Dortmund
5.1 Teams considered 13/08/2012 2.3 0.6 2 1
13/05/2012 1.5 3.3 2 5
The two teams used for testing the prediction capabilities 12/04/2012 0.7 1.3 0 1
of the neural network are FC Bayern Munich and FC 20/11/2011 0.2 0.8 0 1
Borussia Dortmund. Both teams perform in the top flight
German league the Bundes Liga.
5.3 Matches predicted against human guesses
To test how the neural network predictions compare to
human predictions, the results of the neural network
were compared against book maker odds. Book maker
odds are figures released by bookmakers or betting
managers that represent the chance of a team winning
the match as a fraction. The fraction represents how
much a better can expect to earn for backing a particular
team. Conversion of the fraction into each team's odds
involves multiplying by the denominator for both teams.
For example, if Bayern Munich is given odds of 1/7, it
means that for every seven rupees put on Bayern
Munich, there is a chance of earning back one rupee.
Figure 4: Error rate of control set in 65 iterations This means that the booker thinks Dortmund have a one
in seven odd of winning or a 14.3% chance of winning.
Both teams have also participated at every Champions The below table is a comparison of booker odds with
league since 2003 and feature in their ranking system. predicted win percentages to test how the neural network
One reason that two top teams have been selected is that compares to human intuition.
in matches involving a weaker team and a stronger team,
the result is generally unpredictable, both Bayern
Munich and Borussia Dortmund are top teams and
matches between them are keenly contested. The other
reason is that it is much more practical to select two well
known and popular teams since past statistics of these
teams will be easy to locate.

In practice, this method of sports prediction can be

applied to any two teams, including for instances where Table 4: Predicted results Vs Booker odds
a good team is playing against a weak team. During such
matches, the better team tends to win by a smaller Date Predicted result Booker odds
Munich Dortmund Munich Dortmun
margin [7] than expected due to factors such as the
d
'underdog factor' where the weaker team performs much
13/08/2012 70% 40% 20% 80%
better than expected against the stronger team. In such a 13/05/2012 10% 95% 20% 80%
ISSN: 2367-895X 6 Volume 3, 2018
 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

12/04/2012 44% 72% 50% 50% case of predicting the winner, the system is found have
20/11/2011 43% 80% 66.7% 33.3% an RMS error value slightly more than the FRES system
of prediction. Another feature of the system of
Note that the sum of percentages of prediction of either prediction is that the later the match predicted, better the
team winning may not be equal to 100%. This is because accuracy. This clearly highlights the fact that this
the neural network has not created the connection that method of prediction is a learning based method of
sum of returned percentages should be equal to 100%, as prediction and that the system predictions improve with
the training pattern makes use of a 100 - 0 relationship to more test cases.
indicate a winning team.
CITATIONS
5.4 Comparison with FRES
1. "The jersey colors losers wear",
FRES makes use of a rule-based network akin to an if- www.news.menshealth.com/the-jersey-colors-
else system to predict the outcome of the match. FRES is winners-wear/2012/05/02/
found to be very successful in predicting the final 2. "Normalizing NN data",
outcome of the match closer to match completion. This www.elitereader.com/vb/showthread.php?threadid=
is because FRES divides the match into a number of 118179
time segments and inputs are used for each given time 3. "UEFA club coefficients",
segment. www.uefa.com/memberassociations/uefarankings/cl
ub/index.html
The FRES approach involves factors such as emotional 4. "Bundesliga stats",
state (depending on current score line), the teams' www.bundesliga.com/en/about/questions/marketing
offensive and defensive capabilities etc. That is, it is /
more of a here and now kind of system used to predict 5. "Bayern Munich Wikipedia page",
matches based on numerous current factors. What it en.wikipedia.org/wiki/FC_Bayern_Munich
lacks is the capability to make deductions based on past 6. "Home advantage",
factors or performances of the team. en.wikipedia.org/wiki/Home_advantage
When FRES was applied to predict the result of a match 7. "How teams rise to the occasion",
it was found that given the first half data input, FRES http://www.soccerhelp.com/Soccer_Tactics_Weak_
would predict the full time winner and score line very Teams.shtml
accurately. The performance of FRES is better in this
regard when compared to the system of prediction using ACKNOWLEDGEMENTS
neural networks. It is also found that the use of neural
networks to predict matches before the start of the match The author would like to thank Mr. Arun Prakash,
produces results more accurate than those of FRES. manager and coach of Andrew football club, Hyderabad
Another advantage is that this approach can predict for his valuable guidance on factors affecting a teams'
matches weeks even months in advance, whereas FRES performance in a match. The author also owes a debt of
only starts predicting matches once they have begun. gratitude to a number of Bundes Liga fans on the Bundes
Liga forums who answered questions about the league
Table 5: Predicted RMS error Vs FRES RMS error and the two teams and provided certain hard to find
statistics.
S. FRES RMS error Prediction RMS error
No. Goals Winner Goals Winner
1 6.593 6.108 1.500 6.403
2 5.920 5.306 1.772 9.014
3 6.809 6.358 0.761 2.280
4 5.702 6.212 0.283 5.237

6 Conclusion
From the table it is clear that this method of prediction is
much better at predicting goals than FRES. This is REFERENCES
mainly because of the vast number of factors taken into [1] Byungho Min, Jinhyuck Kim, Chongyoun Choe and
consideration that give an inkling of the defensive and Robert Ian, "A compound framework for sports
offensive capabilities of the two teams in question. In the prediction: The case study of football", Knowledge-
ISSN: 2367-895X 7
Based Systems, Vol. 21, No. 7, pp. 551-562, 2008.
Volume 3, 2018
 International Journal of Mathematical and Computational Methods
K. Sujatha et al. http://www.iaras.org/iaras/journals/ijmcm

[2] J. Sindik and N. Vidal, "Uncertainty coeffecient as a Football match outcomes", Working Papers, Queen
method for optimization of the competition systems Mary University, 2012.
in various sports", Sport Science, Vol 2, No. 1, pp. [5] Andreas Heuer and Oliver Rubner, "Towards he
95-100, 2009. perfect prediction of soccer matches", Westfalische
[3] Y. Y. Petrunin, "Analysis of the football Wilhelms University, Germany, In press, 2012.
performance: from classical methods to neural [6] Andrew James Moore, "Predicting football results",
network", Conference Paper, unpublished, 2011. Unpublished, Module code: COM 3021, May 6,
[4] A. C. Constantinaou, N. E. Fenton and M. Neil, "A 2004.
Bayesian network model for forecasting Association

ISSN: 2367-895X 8 Volume 3, 2018