Content uploaded by Malhar Bangdiwala
Author content
All content in this area was uploaded by Malhar Bangdiwala on Oct 13, 2022
Content may be subject to copyright.
Using ML Models to Predict Points in Fantasy
Premier League
Malhar Bangdiwala
Computer Engineering
Sardar Patel Institute of Technology
Mumbai, India
malhar.bangdiwala@spit.ac.in
Rutvik Choudhari
Computer Engineering
Sardar Patel Institute of Technology
Mumbai, India
rutvik.choudhari@spit.ac.in
Adwait Hegde
Computer Engineering
Sardar Patel Institute of Technology
Mumbai, India
adwait.hegde@spit.ac.in
Abhijeet Salunke
Computer Engineering
Sardar Patel Institute of Technology
Mumbai, India
abhijeet salunke@spit.ac.in
Abstract—Fantasy Premier League is an ever-growing game,
with millions of people playing the game. To outperform the rest,
it is imperative for the players to accurately predict the expected
points the footballer would earn over the course of the match.
However, doing so is not easy as there are several aspects to
consider as well as the human bias towards the players’ favourite
footballers and teams. This paper attempts to build and compare
three machine learning models to accurately predict the number
of points that each footballer would earn over the course of the
season. For doing so, the Linear Regression, Decision Tree, and
Random Forest algorithms have been leveraged. Features such as
fixture difficulty, form of the two teams, creativity, and threat of
the footballer have been considered. This would help the players
of this game to make more informed decisions while making their
respective teams.
Index Terms—Linear Regression, Decision Tree, Random For-
est, Fantasy Premier League
I. INTRODUCTION
A fantasy sport is a form of game in which players assemble
fictitious or virtual teams made up of proxies for real players
in a professional sport. It is commonly played on the Internet.
These teams compete based on their players’ statistical per-
formance in actual games. This performance is transformed
into points, which are collated and totaled based on a roster
chosen by the manager of each fantasy club. In fantasy sports,
the owners, like in actual sports, select, trade, and remove
players from their teams.
Fantasy Premier League, or FPL is a popular game that casts
the player in the role of a Fantasy manager. The manager is
given the task of picking a squad of fifteen players who would
give the manager points based on their real life performances.
This game is based on the English Premier League. The
English Premier League (EPL) is the highest level in the
English football league system. It is contested by 20 clubs
and follows the English Football League’s promotion and
relegation system. Each team plays 38 matches during the
season, which runs from August to May. This means that each
team plays each other twice: once at home and once at the
oppositions home ground. FPL is an ever growing community:
from less than two million managers in the 2007/08 season, it
has grown to around eight million for the 2020/21 season[1].
The managers must make a squad of fifteen players and
would receive points for each player based on their real
life performance for their clubs in the EPL matches. Each
player has a certain price and the manager can not exceed
the stipulated budget of 100 million while building the squad.
Players earn points for goals, assists, saves, and clean sheets.
As a reward for a solid performance in a match, players can
receive additional bonus points. The starting XI will score the
manager’s team’s points for the match round or ”Gameweek”.
However, if a starting player does not appear for his club in
that round of fixtures, the points earned by the first player
off the bench will be tallied instead. If two or three starting
players fail to show up, the same thing happens. Extra points
can be earned by selecting a captain from the starting lineup.
In that Gameweek, the captain’s points are doubled. The exact
breakdown of points for each action is mentioned in Table 1
[2].
Fantasy Premier League releases its own set of statistics for
each player that includes, but is not limited to: form, influence,
creativity, threat and opposition difficulty. With increasing
number of statistics, it has become hard for the managers to
keep track of which set of numbers play an important role in
the squad making procedure. Further, guessing the number of
a points a play would accumulate in the future is based on
estimation and the bias of the managers.
This paper aims to eliminate the above-mentioned by
building highly accurate models using the algorithms: Linear
Regression, Decision Tree and Random Forest. The model has
been validated by looking at each player’s performance for
2022 2nd Asian Conference on Innovation in Technology (ASIANCON)
Pune, India. Aug 26-28, 2022
978-1-6654-6851-0/22/$31.00 ©2022 IEEE
1
TABLE I
POI NTS F OR E ACH AC TIO N
Sr. No. Action Points
1Playing up to 60 minutes 1
2Playing 60 minutes or more (excluding stop-
page time)
2
3Each goal scored by a goalkeeper or de-
fender
6
4Each goal scored by a midfielder 5
5Each goal scored by a forward 4
6Each assist for a goal 3
7Clean sheet by a goalkeeper or defender 4
8Clean sheet by a midfielder 1
9Every 3 shots saved by a goalkeeper 1
10 Each penalty save 5
11 Each penalty miss -2
12 Bonus points for the best players in a match 1-3
13 Every 2 goals conceded by a goalkeeper or
defender
-1
14 Each yellow card -1
15 Each red card -3
16 Each own goal -2
each gameweek of the 2020/21 season. For each gameweek,
the model is fitted by using all historical data prior to that
week, and then calculate the mean absolute error for the
following 6 gameweeks. This historical data consists of all
the data dating back to the 2016/17 season.
II. LITERATURE SURVEY
There are not many papers when it comes to predicting
points for Fantasy Premier League. Thus, the scope needs to
be widened to prediction of points for players in any sport.
M. Pardee[3] presents and explores methods for predicting
Quarterback Fantasy Football scores with minimal data. It also
contains various recommendations for improving the data in
order to attain better results. The author includes the game
data from the previous six NFL in the dataset. Support Vector
Regression (SVR) and Neural Networks were utilized by
the author to predict the Fantasy Football scores of NFL
players. The author states that the results of these models were
promising given that the data used for training was limited
In 2013, Bonomo et al. [4] applied two mathematical
programming models that operate as virtual coaches, selecting
a virtual squad roster for each round of the real Argentinian
soccer league. When the season is over and all the results
are known, the a posteriori model determines what would
have been the best lineup for the game. The apriori model
was submitted to the fantasy game and received results that
placed it among the top scorers. The authors state that further
development of such models would provide important tools to
assist real-world coaches and managers in making decisions.
The authors created a posterior model which is a descriptive
one because it addresses the question of what would have been
the best team in each round of the tournament if the outcomes
had been known in advance. As a result, it can discover an
ideal set of teams that would have to be created in order to
get the highest possible total points while staying within the
game’s limits. The second model which is the prescriptive
one was created a priori and offers adjustments to the team
lineup round by round in order to improve team robustness
as the tournament advances without knowing future results. It
accomplishes this by combining data on player performance
in previous tournaments and past matches in the present one
with data on the round’s essential qualities.
Katara et al. [5] used deep neural networks to predict a
team of eleven players for Dream11, the Indian Cricket version
of Fantasy sports. They also compared the accuracy of the
neural network against that of well-known machine learning
approaches like k-nearest neighbours, logistic regression, naive
bayes, random forest, support vector machines.
One of the earliest attempt on FPL prediction is that of
Stolyarov et al. [6] in 2017, who predicted the players’
performance with the help of gradient boosted trees. The
author offers a model of optimal sequential decision-making in
the Fantasy Premier League (FPL). The author uses XGBoost,
one of the most powerful recent data science approaches, to
forecast predicted points. The author then creates an automated
manager that employs this strategy and produces decent team
formation results.
Saifi et al. [7] combined random forests and gradient
boosted machines to analyse the accuracy of the predictions.
Delano[8] has made predictions using convolutional neural
networks and has compared its accuracy with other models.
More recently Gupta[9] combined Autoregressive Integrated
Moving Average (ARIMA) and Long Short Term Memory
for the predictions. Gupta went one step further by predicting
the squad as well by using liner programming techniques. Al-
though, Gupta expanded the horizon in this research direction,
the model was overfitted. In 2020, Sonderberg et al.[10] used
several machine learning models and compared their accuracy.
However, in that paper, the subject of creating the squad was
not touched upon. Similar to Sonderberg, Gunjan[11] created
different models for each position: goalkeeper, defender, mid-
fielder and attacker. This seemed to increase the accuracy of
the predictions.
Nicholas Bonello et al.[12] states that the performance
predictors for the Fantasy Premier League (FPL) are frequently
based only on past statistical data. External factors such as
injuries, managerial decisions, and other tournament match
data can never be integrated into the final forecasts with this
approach. The author suggests a new strategy for forecasting
future player performances by automatically incorporating
human feedback into the model. The author has described a
novel strategy for selecting FPL player picks for upcoming
gameweeks in this work that automatically incorporates human
feedback into the algorithm via publicly available online data
such as blogs published by subject experts. His recommender
system can take into account expert and fan feedback by
looking at news articles, fantasy football prediction blogs, and
2
even tweets utilizing fantasy premier league specific hashtags.
These are employed as additional parameters in the predictive
model, allowing it to automatically incorporate external factors
such as midweek game lineups and performances, managerial
decisions, rotations, and injuries, as well as unexpectedly good
or dismal prior week results.
All of this work done in earlier times is a testament that the
prediction of points is possible by leveraging several statistical
and machine learning models.
III. METHODOLOGY
A. Dataset
The data for training the models are taken from Vaastav’s
Github repository[13]. Here, there is complete information
about every player and team season-wise, starting from the
2016/17 session. The repository is being updated every week
based on the real-life matches that take place in the Premier
League. For the purpose of this paper, player and team data
from the 2016/17 season and onwards has been considered,
upto the 2020/21 season. The 2021/22 season has not been
used since that is the ongoing season and complete data for
the same is not available.
The available data can be broadly classified into two cat-
egories: player-specific data and team-specific data. Team-
specific data is the data which is common to all the players
of that specific team. This includes the following data:
1) Difficulty rating of the opposition team relative to their
team.
2) Whether the match is being played at home or away.
3) Strength of the various departments (attack, midfield and
defense) at home and away.
Some of the player-specific data would include:
1) Goals scored
2) Assists
3) Minutes Played
4) Clean sheets kept
B. Data Cleaning
From all the available data, only the important features are
used for training the models. Such important features include
the goals, shots, blocks, tackles etc. Certain predictive statistics
which are provided officially by the FPL platform such as
form, ict(innovation, creativity, threat) index, expected goals,
expected assits, probability of playing the next match etc.
have also been used. For each gameweek, a new dataset is
curated that contains all the data from the beginning of the
2016/17 season until the last completed gameweek. Certain
columns which contain data of a specific window have also
been added. This includes columns such as minutes played in
the last three matches, points scored in the last five matches
etc. The exact features to be selected was based on Fig.1. This
is done because these parameters will be used as features to
predict the points of future game weeks.
Fig. 1. Feature vs importance graph
The columns where data is not available are replaced with a
zero. The reason for this absence of data can be either due to a
new player being added to a squad during the transfer window
or if a player has been sold to another team during the same.
Major injuries can also lead to a player missing out on many
of the matches in that season. Further, encodings have been
carried out to convert the season and playing position of the
player into numerical data. This cleaned data is now fed to
the models for training, testing and validation.
C. Models
For the purpose of this paper, three regression models were
used: Linear Regression model, Decision Tree Regression
model, and Random forest Regression model.
1) Linear Regression Model: A linear model is one in
which the input variables (x) and the single output variable
(y) are assumed to have a linear relationship. That y can be
determined using a linear combination of the input variables
is more detailed (x). The procedure is known as simple linear
regression when there is only one input variable (x). Multiple
linear regression is the term used in statistics literature when
there are multiple input variables. In our case, we will use the
multiple regression model as there are multiple features that
may affect the points of the player. Some of the independent
variables are the number of assists, goals, clean sheets, goals
conceded, goals scored, minutes played, red cards, yellow
cards, etc. The dependent variable in this scenario is the points
of the player.
2) Decision Tree Regression Model: In the form of a tree
structure, a decision tree creates regression or classification
models. It incrementally breaks down a dataset into smaller
and smaller subsets while also developing a decision tree.
A tree with decision and leaf nodes is the end result. The
Root Node is the first node in the tree, which represents the
3
complete sample and can be further divided into nodes. The
branches indicate decision rules, whereas the inside nodes
reflect data set features. Finally, the outcome is represented
by the Leaf Nodes. This algorithm is extremely beneficial for
handling problems involving decisions. Decision trees offer
the advantages of being simple to grasp, requiring minimal
data cleansing, non-linearity having no effect on the model’s
performance, and having a small number of hyper-parameters
to modify. It may, however, have an issue with over-fitting.
3) Random Forest Regression: Random Forest Regression
is a supervised learning approach for regression that uses the
ensemble learning method. The ensemble learning method
combines predictions from several machine learning algo-
rithms to produce a more accurate forecast than a single model.
During training, a Random Forest constructs many decision
trees and outputs the mean of the classes as the prediction
of all the trees. Random Forest Regression is a powerful and
precise model. It usually works well on a wide range of issues,
including those with non-linear relationships. However, there
are some drawbacks: there is no interpretability, overfitting is a
possibility, and we must choose the number of trees to include
in the model. In our case we will have 7 decision tress in the
Random Forest Regressor.
D. Comparison Parameters
To compare the above mentioned models, Root Mean
Square Error (RMSE) and Mean Absolute Error(MAE) have
been used.
1) Root Mean Square Error: The root mean square error
(RMSE) is the residuals’ standard deviation. The RMSE is a
measure of how spread out the residuals are in comparison
to the regression line data points. In other words, it indicates
how tightly the data is clustered around the line of best fit. It
is calculated according to the formula:
RM SE =r1
nΣn
i=1(xi−ˆxi)2(1)
Where,
i = Summation variable
n = Total number of observations
xi=Actual value of observation
ˆ
xi=Predicted value of observation
2) Mean Absolute Error: The Mean Absolute Error (MAE)
is a regression model evaluation statistic. The mean absolute
error of a model with regard to a test set is the average of all
individual prediction errors on all instances in the test set. The
difference between the true value and the anticipated value for
each instance is the prediction error. It is calculated using the
formula:
MAE =1
nΣn
i=1|yi−xi|(2)
Where,
i = Summation variable
n = Total number of observations
xi=Actual value of observation
yi=Predicted value of observation
E. Results
After the models are trained, RMSE and MAE are calcu-
lated. RMSE for training dataset for Linear model, Decision
Tree, and Random Forest Regressor are 2.172224, 0.260842,
0.260842. It is observed that the RMSE of the Linear model,
Decision Tree, and Random Forest Regressor are 2.159776,
2.987197, and 2.145552 respectively for the test dataset. This
has been visualized in Fig.2 and 3 respectively.
Fig. 2. RMSE for test dataset
Fig. 3. RMSE for training dataset
The MAE for the training dataset is 1.289776, 0.024551,
and 0.024551 in the same order, and for the test dataset
1.264294, 1.527351, 1.214767. This can be seen above in Fig.
4 and 5 respectively.
4
Fig. 4. MAE for test dataset
Fig. 5. MAE for training dataset
Model-wise accuracy has been tabulated in the tables 2, 3
and 4.
TABLE II
LINEAR REGRESSION MOD EL ACC URA CY SUMMARIZATION
Sr. No. Metric Train Test
1RMSE 2.172224 2.159776
2MAE 1.289776 1.264294
TABLE III
DECISION TRE E REGRESSION MODE L ACCURACY SUMMARIZATION
Sr. No. Metric Train Test
1RMSE 0.260842 2.987197
2MAE 0.024551 1.527351
TABLE IV
RANDOM FOREST REGRESSION MOD EL ACCURACY SUMMARIZATION
Sr. No. Metric Train Test
1RMSE 0.260842 2.145552
2MAE 0.02455 1.214767
Thus we can observe that the decision tree and random
forest regressor are overfitted as the error metrics for the
training dataset are low but are high for the test dataset. The
linear regression model and random forest model perform
similarly for the training dataset but we can notice that the
linear regression model is not overfitted to the training dataset.
IV. CONCLUSION
In conclusion, this paper has successfully predicted the
points a player would accumulate in the upcoming match
on the FPL platform. Three models have been created and
their accuracy has been compared. It can be inferred that the
Linear Regression Model performed the best among the three.
Predicting the points a player would earn is the key decision
any manager on FPL has to make. The entire game revolves
around this sole aspect. This decision is rather daunting given
the amount of data and statistics that is available. With the
help of our models, this task has been made easier for all the
managers on FPL. This paper also contributes to the work that
has taken place in this domain by validating three models that
can be used for making predictions.
V. FUTURE WORK
This line of work can be extended by creating the entire
squad of fifteen players based on the points predictions.
Simple mathematical and statistical models such as Linear
Programming can be used for the same. Further selecting the
starting eleven from the fifteen can also be done. The team can
then be used on the platform to check how well the models
perform compared to the real world managers.
REFERENCES
[1] J. Gatsman, ”Fantasy Premier League grows to +7 million users this sea-
son” fanarena.com. https://fanarena.com/fantasy-premier-league-growth-
users-per-year/ (accessed: May 8,2022)
[2] The Scout, ”FPL basics: Scoring points” premierleague.com.
https://www.premierleague.com/news/2174909 (accessed: May 8,2022)
[3] Pardee, M. An Artificial Neural Network Approach to College Football
Prediction and Ranking. 1999
[4] Bonomo, Flavia, Guillermo Duran, and Javier Marenco. “Mathematical
programming as a tool for virtual soccer coaches: A case study of
a fantasy sport game”. In: International Transactions in Operational
Research 21. DOI: 10.1111/itor.12068. 2013
[5] Katara, Karthik Krishnan, Gokul Shetty, Shashank Bankapur, Sanjay
Kolkar, Ranjit T S, Ashwin Vanahalli, Manjunath. (2020). Analysis and
Prediction of Fantasy Cricket Contest Winners Using Machine Learning
Techniques. 10.1007/978-981-15-5788-0 43.
[6] Stolyarov, Arseniy and Gleb Vasiliev. “Predict To Succeed: Optimal
Sequential Fantasy Football Squad Formation Using Machine Learning
Tools”. 2017
[7] Saifi, Murtaza. “Implementation of Machine Learning Techniques to
Predict Player Perfor-mance using Underlying Statistics”. PhD thesis.
Dublin, National College of Ireland. 2018
5
[8] Ramdas, Delano. Using Convolution Neural Networks to Predict
the Performance of Footballers in the Fantasy Premier League.
10.13140/RG.2.2.10010.72645/2. 2022
[9] Gupta, Akhil. Time Series Modeling for Dream Team in Fantasy Premier
League. 2019
[10] Lindberg, Adrian and David Sonderberg. “Comparison of Machine
Learning ApproachesApplied to Predicting Football Players Perfor-
mance”. English. MA thesis. Gothenburg: ChalmersUniversity of Tech-
nology and University of Gothenburg. 2020
[11] Gunjan, Kumar. “Machine learning for soccer analytics”. In: University
of Leuven. 2013
[12] Bonello, Nicholas, Joeran Beel, Seamus Lawless, and Jeremy Debattista.
“Multi-streamData Analytics for Enhanced Performance Prediction in
Fantasy Football”. In: arXiv preprintarXiv:1912.07441 2019.
[13] Anand, Vaastav. ”Fantasy-Premier-League” github.com.
https://github.com/vaastav/Fantasy-Premier-League (accessed: May
8,2022)
6
