Man v Machine: Greyhound Racing

Predictions

MSc Research Project

Data Analytics

Alva Lyons
x15014274

School of Computing
National College of Ireland

Supervisor: Mr. Oisı́n Creaner

 National College of Ireland
Project Submission Sheet – 2015/2017
School of Computing

Student Name: Alva Lyons

Student ID: x15014274
Programme: Data Analytics
Year: 2016
Module: MSc Research Project
Lecturer: Mr. Oisı́n Creaner
Submission Due 21/12/2016
Date:
Project Title: Man v Machine: Greyhound Racing Predictions
Word Count: 5964

I hereby certify that the information contained in this (my submission) is information
pertaining to research I conducted for this project. All information other than my own
contribution will be fully referenced and listed in the relevant bibliography section at the
rear of the project.
ALL internet material must be referenced in the bibliography section. Students
are encouraged to use the Harvard Referencing Standard supplied by the Library. To
use other author’s written or electronic work is illegal (plagiarism) and may result in
disciplinary action. Students may be required to undergo a viva (oral examination) if
there is suspicion about the validity of their submitted work.

Signature:

Date: 20th September 2018

PLEASE READ THE FOLLOWING INSTRUCTIONS:

1. Please attach a completed copy of this sheet to each project (including multiple copies).
2. You must ensure that you retain a HARD COPY of ALL projects, both for
your own reference and in case a project is lost or mislaid. It is not sufficient to keep
a copy on computer. Please do not bind projects or place in covers unless specifically
requested.
3. Assignments that are submitted to the Programme Coordinator office must be placed
into the assignment box located outside the office.

Office Use Only

Signature:

Date:
Penalty Applied (if
applicable):
 Man v Machine: Greyhound Racing Predictions
Alva Lyons
x15014274
MSc Research Project in Data Analytics
21st January 2017

Research Questions

Can a system be built using machine learning techniques to predict greyhound

racing results? When bench-marked against a human expert how effective is
this system?

Abstract
The purpose of this research is to ascertain if machine learning techniques can
prove advantageous in predicting the outcome of greyhound races. The main focus
of this research is in bridging the gap between existing sports prediction models
which use manual feature selection to creating a model built from machine chosen
subsets by algorithmically sub-setting the feature space. Feature selection is the
process of sub-setting the feature space by analysing the relevance of features both
to each other and to the predicted variable so that only the most relevant fea-
tures are used within the modelling framework. The reason for introducing the
greyhound racing expert is to test whether the system can outperform the average
social gambler who tends to make their betting selection based on tips given to
them by domain experts.

1 Introduction
The greyhound racing industry in Ireland is controlled by the Irish Greyhound Board
(IGB). It is estimated that 720,000 visitors annually attend IGB controlled greyhound
stadia. Shelbourne Park is the premier greyhound stadium in Ireland and hosts one of
the world’s richest greyhound races, The Irish Derby, every September. Racing takes
place in Shelbourne Park every Wednesday, Thursday and Saturday.

This research will attempt to predict the finishing position of a greyhound in a given
race. The data used in this research comprises of 64,908 observations of 10,986 races ran
in Shelbourne Park between January 2009 and August 2016. The prediction rate of this
model is bench-marked against that of the stadium’s resident greyhound expert who is
employed by IGB to predict the winning greyhound, the top 2 and the top 3 finishing
greyhounds for the top of the race card for each race on a given race night.

1
 The use of machine learning techniques in sports prediction is not a new phenomenon
but rather it has gained many more practitioners since the spread of online gambling
markets. While the use of machine learning techniques is prevalent in predicting horse
racing (Butler et al. (1998), Silverman and Suchard (2013), Davoodi and Khanteymoori
(2010), Williams and Li (2008) etc.) results there have only three documented cases of
utilising machine learning in predicting the outcome of greyhound races. While the two
sports are often synonymous there are distinct differences between the two which ensures
that modelling concepts need to be amended. A greyhound race result is the outcome
of 6 greyhounds chasing a mechanical hare in their attempts to catch it; while a horse
race result is the outcome of the interactions between a jockey and its mount as they
traverse the race course. While this might seem trivial, the key difference is apparent
when considering a model’s attempts at predicting the finishing positions of competitors
in a race. A greyhound is bred to chase the hare and will continue its mission even if the
race has already been won. On the other hand, a jockey which surmises it has no chance
of finishing in the first x positions, may choose to pull back so that the horse’s handicap
rating is not affected for its next race. This nuance is one of the factors that led this
researcher to choose greyhound racing as the sport of choice for this research.

The seminal paper in the field of greyhound racing predictions, Chen et al. (1994),
dates back to 1994 and uses the knowledge of a greyhound expert for feature selection in
choosing which performance variables to use when running machine learning techniques
on their dataset. Both of the follow up studies utilise a similar feature selection in their
models (Schumaker and Johnson (2008), Johansson and Sönströd (2003)). This research
paper uses feature selection algorithms to limit the problem space of the domain in order
to avoid the subjectivity of human interactions within the modelling process.

Additionally, this paper combines various data mining techniques from sentiment ana-
lysis through to deep learning ensemble methods in its attempts to test if a machine
learnt model can out-preform an expert in the area of greyhound racing predictions. The
rest of this document is laid out as follows:
• Section 2 discusses the related work in the field of sports predictions and highlights
the role this research plays within this field.
• Section 3 discusses the methodology framework used in completing this research.
• Section 4 reviews and justifies the implementation steps carried out in this re-
search.
• Section 5 evaluates the results of the prediction algorithms.
• Section 6 concludes the research and discusses potential future work to be carried
out.

2 Related Work
2.1 Sports Predictions
The literature and academic work produced in the field of sports predictions is far reach-
ing. Using historical results data to predict the outcome of sporting events has gained
exposure due to the growth of on-line betting markets and the large volumes of historical
data which are easily accessible.

2.2 Horse Racing’s Favourite Long-Shot Bias

The sport of horse racing is often synonymous with greyhound racing and includes a
vast amount of literature and analysis on sports prediction. Many of the earlier models
focus on predicting the efficiency of betting markets in the hopes to apply profit to bets.
Central to these predictions is the perceived notion of the favourite long-shot bias which is
based on the phenomenon that continual betting on an outside selection, or long-shot, will
return less than betting on a favourite despite the higher odds attached to the long shot,
Leighton Vaughan Williams (1997). Kanto et al. (1992) discovered that most gamblers
are willing to risk losses by over valuing long shots due to the enticement of added gain
when winning on a higher priced selection.

Cain et al. (2003) carried out analysis to test whether the favourite long shot bias that
is prevalent in horse racing betting markets is found in other sports. The results of their
analysis shows that betting repeatedly on favourites results in smaller losses than betting
on long shots in most sports; with the exception of soccer and greyhound racing, Lyons
(2016).

2.3 Problem Space and Feature Set

An important step in the data mining process is choosing which features to include in
your model. Feature selection can either be done manually through the use of domain
knowledge or algorithmically with the use of machine learning methods. The race card
available on tracks includes 50 variables which could potentially affect the outcome of a
race. Adding all of these variables into a model would increase it’s complexity and be
algorithmically inefficient, Lyons (2016).

Many of the works performed on predicting results of horse and greyhound races focus
on the model used for prediction and it’s tuning parameters rather than the selection of
the feature subset (Pudaruth et al. (2013), Davoodi and Khanteymoori (2010), Williams
and Li (2008)). Their feature subset are listed but the motivation behind choosing which
features to include in their model is not elaborated on. This lack of formal explanation
for the feature subset doesn’t allay the assumption of human subjectivity in the choosing
of what performance variables affect the outcome of these sporting events.

McCabe and Trevathan (2008)’s paper focuses more on the feature set than the model
used in sports prediction. This paper provides an interesting discussion on why variables
were included in the model however they are very vague on the potential ”subjective”
variables not added. Similar to the papers listed above feature selection in the research
by McCabe and Trevathan is a manual process and does not use machine learning to
choose the optimal subset of features to include in the modelling.
2.3.1 Historical Feature Selection Techniques in Greyhound Racing Predic-
tions
Chen et al. (1994) in their prediction of greyhound racing results chose their feature
set following discussions with domain experts who informed them which 10 performance
variables they believed were most important in predicting winners. They admit that
while this is not optimal it is a consequence of their chosen algorithms being unable to
handle noisy data. Remarkably neither Johansson and Sönströd (2003) nor Schumaker
and Johnson (2008), in their follow up studies, chose to research further attempts at
feature selection. Rather they used a similar feature subset to those used in the study by
Chen et al.. (Lyons (2016))

2.3.2 Bridging The Research Gap

This research attempts to apply various feature selection algorithms to the transformed
dataset in order to ascertain which features have a greater impact on influencing a grey-
hounds finishing position within a race. The features which are extracted as relevant to
this domain problem are then chosen as the final dataset to be used in the model. The
choice of using a neural network in the modelling phase of this research is to test if the
use of machine based feature selection can outperform those as used by Chen et al. (1994)
and Johansson and Sönströd (2003). As this research is focusing on classification rather
than regression modelling the model choice of Schumaker and Johnson (2008) (Support
Vector Regression) is discounted from the outset.

3 Methodology
The methodology used in this research is Knowledge Discovery in Databases (KDD). The
KDD methodology allows for an iterative approach to the processes involved in extracting
knowledge from raw data. Initial plans were to utilise the SEMMA notation, as developed
by SAS, but the sequential nature of this methodology couldn’t rival the flexibility and
interactivity of KDD, Azevedo and Santos (2008). KDD focuses on the entire process
from data selection through pre-processing, extraction, data mining to interpretation,
Fayyad et al. (1996). An illustration of the KDD methodology as it pertains to this
research is shown in Figure 1

Figure 1: KDD Methodology

3.1 Selection
The raw data used in this research is extracted from the Irish Greyhound Board’s (IGB)
website 1 using a Python script. This data consists of 383,746 observations of 28,271 races
ran throughout Ireland between 1st April 2007 and 3rd September 2016. This data is
collected from multiple embedded pages and loaded into 6 tables in a MySQL database.
The flow of this script is illustrated in the diagram in Appendix A.

3.2 Pre-Processing
The benefits of a thorough pre-processing phase ensure a strong knowledge of the dataset
is gained before transformations commence.The data from IGB’s website contains numer-
ous inaccuracies and missing data points ensuring the pre-processing phase of the KDD
methodology plays an integral role in this research. Errors in the data are discovered
when the data is examined using visualization and descriptive statistics.

3.2.1 Dealing with Missing Values

(a) TrapData Table (b) DogRaceHistory Table

(c) Races Table (d) Dogs Table

Figure 2: Missing Values in Raw Data Tables

In dealing with missing values it is necessary to ascertain why a value is missing or

incorrect and to decide if action is to be taken, Witten et al. (2011). The proportion
of missing values to actual data can be seen in Figure 2. In these diagrams the red
squares represent missing values while the blue squares represent the presence of data. It
is evident that the DogRaceHistory table, in particular, has a large proportion of missing
1
www.igb.ie/results
values. As this table is integral to determining the performance history of a greyhound
it is important to ascertain the best method of handling this missing data.

Domain knowledge plays an important part in deciding what steps to take in handling
missing values. For instance, a NULL value in the Seed column does not represent a
true missing value. A greyhound’s seed indicates their preferred running style; Inside (I)
seeded greyhounds tend to run toward the bend; Middle(M) seeded greyhounds tend to
run in the centre of the track; Wide(W) seeded greyhounds run toward the rails. The
lack of one of these characters in the seed column is likely to indicate that a greyhound
has no preferred running style. For this reason the missing data points in this column
are replaced with ”A” (Any).

Further exploratory analysis of the DogRaceHistory table shows up explanations for

some of the high volume of missing data points within this dataset. Table 1 depicts the
percentage of missing data in this table that can be attributed to the inclusion of time
trials.

Table 1: % of Missing Values in DogRaceHistory Table that are attributed to Time Trials

DogRaceHistory Table
Field Responsible for % of Missing Data
Weight 100
NumberOfDogs 100
WinTime 100
Going 100
PlacedDistance 99.9
RunnerGrade 86.90
RaceGrade 73.1
SP 70.5
Remarks 61.2
SectionalPosition 46.9
SectionalTime 44.5
EstimatedTime 8.4

Time trials take place on race nights before racing commences whereby 1 or more
greyhounds run around the track in a non competitive setting to see how fast they can
chase the mechanical hare. Due to the non competitive nature of these events it is deemed
appropriate to remove these from the dataset. Removal of time trials significantly reduce
the number of missing data points in the DogRaceHistory table.

3.2.2 Dealing With Incorrect Values

Outlier detection is performed on the dataset to allow for purification of the data before
modelling commences, Hodge and Austin (2004). For instance, in the Weight column of
the DogRaceHistory table the minimum weight of a greyhound is listed as 0lbs while the
maximum is listed as 677lb. The average weight of a greyhound is 65-75lbs. The outlier
values are imputed by taking the mean value of a greyhounds weight in its preceding
and succeeding 2 races and inputting this as the new value in the weight column. In the
instances where a greyhound does not have any other races against it the mean value
of the weights of its competitors is imputed as its new value. Similar imputations are
performed on other outliers within the dataset.

3.2.3 Pruning The Dataset

The track ratings of greyhound stadia in Ireland differ depending on the ground condi-
tions; as such the RunnerGrade variable has varying significance depending on where a
race took place. As this research attempts to predict results in Shelbourne Park it is
deemed appropriate to limit the race history of greyhounds to their performance at this
race track. Additionally, only A grade (middle distance) races are used in this research.
The reason for omitting sprint races lies in the short distance between the first bend and
the finish line. In sprint races if a runner gets knocked at the first bend their chances of
recovery are limited Lyons (2016). The remaining dataset consists of 64,908 observations
of 10,986 races ran between January 2009 and August 2016.

3.2.4 Limitations Of The Dataset Discovered During Pre-Processing

During the pre-processing phase limitations of the dataset are exposed. These limitations
lie in the scraped data being from a view of the IGB’s database at the time of scraping.
While this doesn’t affect historical race content it does ensure that owner and trainer data
is inadmissible for modelling as there is no way of ascertaining whether the greyhound
was attached to its current owner or trainer at the time of each historical race. Including
the data in these two tables in the modelling phase of this research could potentially lead
to incorrect predictions based on corrupt data.

Similiarily, litter distribution data in the Dogs table, which depicts the number of starts
and race placings of a greyhound’s siblings, is also inadmissible as it is that of a view of
the to-date total of the litter at the time of scraping rather than the time of racing.

While omitting this data does have its benefits in that it cuts down the processing time
of feature selection in the data mining phase; it restricts the model, in that it does not
have access to the same data available in real time to the greyhound expert the model
will be bench-marked against.

3.3 Transformation
3.3.1 Text Analysis
The remarks column in the DogRaceHistory table provides a shorthand of comments on
how a greyhound ran in a given race eg. FAw (Fast Away), BBkd (Badly Baulked),
TRec (Track Record) etc. The full list of 320 possible remarks can be found on the IGB’s
website 2 .
2
https://www.igb.ie/upload/pdf/Abbreviations.pdf
 In order to analyse these remarks across the dataset the basic premise of sentiment
analysis is performed such that the text is classified as expressing a positive or negative
tenet, Liu (2010). While sentiment analysis deals with ”the computational treatment of
opinion, sentiment, and subjectivity in text”, Pang and Lee (2008), and is considered to
be more suitable for text mining of unstructured datasets; the simplicity and power of
this method is deemed appropriate to analyse this variable.

The first step in utilising the premise of sentiment analysis is to create domain specific
lexicons of the shorthands used in the remarks column. Once the lexicons are created
confirmation is received from 2 experts working within the greyhound racing industry
to ensure subjectivity is minimized during this phase. These dictionaries are created
by assigning each shorthand comment into one of 5 categories; Very Positive, Positive,
Neutral, Negative, Very Negative.

The motivation for utilising this variable and performing text analysis lies with the
possibility that the greyhound’s ability is not properly reflected by it’s finishing position.
For instance, a greyhound may only finish 5th in a race despite being quick out of the
traps due to being impeded by another greyhound. By the same respect the 1st placed
finisher in this race might have missed the early fighting and received a clear run despite
being slow away.

In order to run sentiment analysis on this column the 5 dictionaries are loaded into
MySQL tables and the remarks column is scored using an SQL statement which scans
each of these tables and matches the word in remarks to those in the dictionaries. A
scoring is given to the words depending on which category they fall into:

• Very Positive = +2 points

• Positive = +1 point

• Neutral = 0 points

• Negative = -1 point

• Very Negative = -2points

The scoring for each remark is totalled and set as the RemarkScore for each greyhound
in a given race.

3.3.2 Feature Engineering

Feature engineering is the task of extracting and transforming raw data into features that
better represent the problem domain in order to improve a model’s prediction when new
data is introduced, Brownlee (2014). The importance of feature engineering lies in finding
the best representation of variables in a dataset in order to bridge the gap between the
features in the initial problem domain to the structure of features needed for the solution
architecture, Dash and Liu (2003). There are many elements to feature engineering from
framing the problem domain to data cleansing and formatting. The element discussed in
this research pertains to the manual construction of new features from the raw data.
 The raw data is transformed to create meaningful information from each dog’s race his-
tory. While Chen et al. (1994) and Schumaker and Johnson (2008) average their variables
over 7 races, this research looks to emulate the on track race card; which provides histor-
ical data on each greyhound’s last 5 races; by averaging the greyhound history statistics
over 5 races. This ensures that the model has access to the same data as the ordinary
social gambler. Where a greyhound has run less than 5 races the data is averaged over
the number of runs of that greyhound up to a maximum of 5 races. The below formula
shows how the rolling averages are calculated; this example calculates a greyhounds av-
erage position at the first bend in each of it’s last n races.

F irstBendr 1 + F irstBendr 2 + · · · + F irstBendr n

BreakAvg5 =
n

Similar formulae are used to transform other variables in the raw data. Figure 3
depicts a table of the transformations that took place in this phase of the research meth-
odology.

Figure 3: Variables Created From Raw Data

3.4 Data Mining
Data mining is the process of analysing datasets to find unobserved and often unsuspected
relationships within the data by combining statistics, artificial intelligence and machine
learning features, (Hand et al. (2001)). Fayyad et al. 1996 address the importance of
understanding the data mining activity before including it in the KDD process. Similar
to the KDD methodology, the choosing of an algorithm to use in tackling a prediction
problem can involve many interactions and iterations before knowledge is gleaned. An
important first step is to decide which data mining process of predictive analysis is re-
quired in ascertaining the value of the predictor variable. Classification analysis deals
with predicting which category or class an object falls into. The required output is a dis-
crete variable. Regression analysis is used to predict missing or unavailable numerical
data values; the output variable is a continuous variable. Han (2005)

The application of predicting the outcome of a competitive event does not strictly
fall into either a classification or a regression problem and as a result both regression
and classification techniques are possible within the realms of this research domain. As a
classification problem the output variable can be a matter of predicting the binary output
of win or lose. As a regression problem it is possible to look at the finishing order of a
race with the view to regressing on the FinishingPosition variable.

This research approaches the problem of predicting greyhound racing results as a clas-
sification problem. However, rather than choosing binary classification of ”win” or ”lose”
it attempts are made to classify a greyhound’s Finishing Position. The reasoning for not
choosing binary classification is partly due to class imbalance; for each race 6 greyhounds
are entered and 5 greyhounds cannot win as such the number of observations in the lose
class in the training set is larger than that of winners and random predictions could result
in a higher rate of prediction due to chance alone. Additionally, by choosing to classify
the problem using the Finishing Position as the predictor variable this allows for testing
how wrong a predicted class is. For instance, incorrectly predicting a 1st place finisher
will finish in 2nd place is ”less wrong” than predicting the same greyhound will finish
in 6th place. The choice of algorithms and justification for their uses is discussed in the
implementation section of this paper.

4 Implementation
4.1 Tools Used
The tools used in implementing this research are:

• Python (Version 2.7.12)

• MySQL

• R (Version 3.3.1)

• R Studio 64bit

• Amazon EC2
Python is used in the selection phase of this research to scrape the raw data due to
the power of its BeautifulSoup library; which provides an easy to use framework for
parsing HTML into a tree representation. The data mining phase utilises R; a statistical
programming language which is widely used for data analysis, Lantz (2013).

4.2 Examining The Dataset

Once the feature engineering phase is complete the next step is to combine and explore
the dataset. A flattened correlation matrix of the processed dataset is produced in R
using the corrplot library (See Figure 4).

Figure 4: Flattened Correlation Matrix

As is evident in this visual representation of the correlation between features there

are a number of strong correlations amongst variables. While some correlations are
expected; such as the positive correlation between the number of starts of a greyhound
and the number of wins; others are unexpected, such as the negative correlation between
a greyhound’s win percentage (the percentage of wins over all races) and its finishing
position in its last 5 races. A negative correlation depicts an inverse proportionality
between variables in that as one increases the other will decrease. This would suggest
that over time a greyhound’s recent form has an obvious affect on its overall performance
in that rather than remaining consistent it will either improve or deteriorate over time.

4.3 Feature Selection

Feature selection is the process of binning variables into subsets of relevant and irrelevant
features such that only the most relevant features are used within the modelling frame-
work, Dash and Liu (1997). A feature is deemed to be relevant if it affects the target
problem in any way. The benefits of feature selection lie in reducing the complexity and
run-time of the machine learning algorithm. The reducing of complexity allows for better
understanding of the patterns that arise in the data mining process. Additionally, feature
selection when performed correctly, can improve model performance.

4.3.1 Methods of Feature Selection

There are three categories of feature selection methods; Wrapper, Filter and Embedded.

• Filter Methods - are concerned with exploring only the inherent features of a
dataset. They are based on statistical tests and are independent of the variable to
be predicted.

• Wrapper Methods - Unlike filter methods wrapper methods are used to find
features subsets which interact with the variable to be predicted. In this way the
choosing of a wrapper method is closely linked to the choosing of a modelling
algorithm as the feature subset space is wrapped around the classifying model,
Saeys et al. (2007).

• Embedded - Embedded methods are an extension of the Wrapper Method frame-

work and attempt to combine the best properties of the preceding two methods.
The feature selection is ’embedded’ in the modelling algorithm which runs feature
selection and prediction concurrently.

This research focuses on wrapper and embedded methods as they interact with the
variable to be predicted, are less likely to get stuck in a local optima and model feature
dependencies. The limitations of these methods, however, lie in the increased risk of
over-fitting, Saeys et al. (2007).

4.3.2 Wrapper with Boruta Package

Several packages are available in order to ascertain the importance of independent vari-
ables in predicting the dependent variable. The Boruta package in R comprises of a
wrapper algorithm which utilises random forests in order to extract relevant features
from a data set. This is achieved by comparing a variable’s importance against import-
ance that is achievable at random, Kursa and Rudnicki (2016). The application of this
package on the dataset did not dramatically reduce the feature space; removing only one
variable (Colour ) from the original 30 inputted into the algorithm.
4.3.3 Wrapper With caret & randomForest Packages
The caret, an acronym for Classification and Regression Training, package contains
functions to organise a model’s training approach, Kuhn (2016), and utilises a number of
other packages in R. In this example, the caret package wraps around the randomForest
package in order to rank the variable importance of the features in the dataset. Variable
importance ratings are assigned to each feature and they are then ranked according to
how important they are to the dependent variable, FinishingPosition. The ranked feature
results are shown in Appendix B. Interestingly Colour, which is the only variable to be
dismissed when running the Boruta package with random forests, ranks above 9 other
features when the the caret package is wrapped around a random forest.

4.3.4 Embedded - Recursive Feature Elimination with caret

Recursive feature elimination (RFE) is an embedded method of feature selection. It
attempts to find the optimal subset of features by iterating through all features, assigning
weights to each feature depending on its value to the dependent variable. Features are
eliminated based on their ranked weighting, in that those with a smaller weighting are
eliminated first. Once a feature is pruned the remaining features are reassigned weights
and the process is iterated until a stopping criterion is reached whereby the optimal
number of features is selected, Guyon et al. (2002).

The caret packages provides a set of predefined functions to embed RFE with al-
gorithmic functions; such as Naı̈ve Bayes; Random Forests; and Bagged Trees. These 3
functions are modelled on the dataset in order to attempt to find the optimal subset of
features to use in prediction.

1. Naı̈ve Bayes is a classification algorithm which is based on Bayes Theorem and as-
sumes independence amongst the feature space. The running of a recursive feature
elimination using Naı̈ve Bayes limits the feature subspace to 3 features; Finishing-
PositionAvg5, OverallAvgTime, Avg2ndBend. Examining the flattened correlation
matrix in figure 4 tells us that the basic assumptions of Naı̈ve Bayes are violated
in the our dataset, in that the features are not independent. Therefore, the results
of this analysis are dismissed for the remainder of this research.

2. Random Forests are the result of combining decision trees such that each tree
depends on the values of a randomly sampled independent vector whereby the
entire forest is distributed homogeneously Breiman (2001). The output of the ran-
dom forest RFE is shown in Appendix C. This depicts the top 10 features se-
lected to be NumberOfDogs, TimeAvg5, OverallAvgTime, PrizeMoneyWonAvg5,
DogsAge, FinishingPositionAvg5, SecTimeAvg5, PlacedPercent, RankedGradeAvg5
and Avg2ndBend5.

3. Tree Bagging is an ensemble method which uses decision trees to generate multiple
versions of a predictor and aggregates the result, Breiman (1996). The output of the
Tree Bagging RFE method is shown in Appendix D. The top 10 features returned
embedding RFE with Tree Bagging are DogsAge, Weight, TimeAvg5, SecTimeAvg5,
OverallAvgTime, RaceNumber, RankedGradeAvg5, PrizeMoneyToDate, BreakAvg5
and TrapNumber.
4.3.5 Combined Feature Selection Results
The results of the caret & randomForest wrapper, embedded treebagging and embedded
randomForest are combined to create an optimal subset of the data for use in the final
stage of the modelling process. The top 10 features of each of these methods are combined
in order to ascertain if there are any features which are prevalent across all feature
selection methods. The ranked table is shown in Table 2.

Table 2: Top Ten Features Selected

Features Selected
caret & randomForest RFE with RandomForest RFE with TreeBagging
DogsAge NumberOfDogs DogsAge
OverallAvgTime TimeAvg5 Weight
SecTimeAvg5 OverallAvgTime TimeAvg5
TimeAvg5 PrizeMoneyWonAvg5 SecTimeAvg5
Weight DogsAge OverallAvgTime
RankedGradeAvg5 FinishingPositionAvg5 RaceNumber
PrizeMoneyWonAvg5 SecTimeAvg5 PrizeMoneyToDate
BreakAvg5 PlacedPercent BreakAvg5
RaceNumber RankedGradeAvg5 RankedGradeAvg5
DaysSinceLastRace Avg2ndBend5 TrapNumber

The features highlighted in red indicate those that are selected as a top 10 ranking
feature across all 3 feature selection methods. These 5 variables are selected for the
final feature subset. Five more features for this subset are selected by choosing the
highest ranked features amongst the 3 methods deployed. The transformed variable,
AvgRemarks5, resulting from the Text Analysis phase of this research (section 3.3.1),
fails to make the cut despite finishing amongst the top 13 features across all feature
selection methods.

Resulting Optimal Feature Subset: The final feature subset selected for the model-
ling phase consists of DogsAge, OverallAvgTime, SecTimeAvg5, TimeAvg5, RankedGradeAvg5,
Weight, RaceNumber, PrizeMoneyWonAvg5, BreakAvg5 and Avg2ndBend5 .

5 Evaluation
5.1 Model Performance
For the purpose of modelling the dataset is split in a 60/20/20 ratio for training, valid-
ation and testing and all variables are scaled to reduce variance across the dataset. An
important design criterion for model performance is choosing the correct parameters. It
is important to ensure that while these parameters are tuned the test set is not utilised
so as to avoid the model learning from iterations over the test set.
5.1.1 Neural Network
In order to emulate the research in the field of greyhound racing by Chen et al. (1994)
and Johansson and Sönströd (2003); who used shallow neural networks in their predic-
tions; the optimal feature subset chosen following feature selection is inputted into a
Deep Learning Neural Network using the H2O package in R. Deep Learning reduces the
complexity of an algorithm but is better suited to larger datasets.

This research initially utilises deep neural networks in its attempts at classifying the
finishing position of a greyhound. A deep neural network is ran several times across
different subsets of the data and the average prediction performance is found to be 18.92%.
The use of shallow neural networks improves the performance of the feature subset;
using just one hidden layer on subsets of the dataset improves the average prediction
performance to 19.89%. The performance of the shallow neural network approaches the
prediction rate of Chen et al. (1994), whose model correctly predicted 20% of winners.
Similar comparisons cannot be made on Johansson and Sönströd’s paper as their model
is assessed against monetary gain on bets rather than win percentages.

5.1.2 Random Forest

A random forest algorithm is ran on the feature subset to see if the use of an ensemble
method can boost the prediction accuracy of the feature set. The random forest fails to
improve upon the performance of the neural network; only predicting 19% of finishing
positions correctly when tested against the validation set and 18.28% when run on the
test set.

5.2 Model Evaluation

The breakdown of the model’s success rates as discussed in Section 5.1 are shown in table
3.
Table 3: Model Success Rates

Model Success Rate

Random Forest 18.28%
Deep Neural Network 18.92%
Shallow Neural Network 19.89%

5.2.1 Greyhound Expert Prediction

The greyhound expert’s predictions are scraped from embedded pop-ups on the IGB’s
website and the percentage of first place finishers correctly predicted is derived. A limita-
tion of this research lies therein. This research, similar to the greyhound expert, attempts
to predict the finishing order of greyhounds in a race; however the expert makes 3 pre-
dictions per race; what greyhound will win the race; what two greyhounds will finish in
the top 2 in any order; and what 3 greyhounds will finish in the top 3 in any order. It
is necessary to only choose the percentage of first place finishes correctly predicted as
the other 2 predictions turn the problem from a 6 class multi-class problem to a binary
classification problem. The greyhound expert correctly predicted 23.7% of first placed
finishers in the time frame used in this research (Jan. 2009 - Sept. 2016).

5.2.2 Random Chance

A typical greyhound race consists of 6 greyhounds chasing a mechanical hare. However,
not all races will have 6 runners. The mean number of greyhounds per race in the final
dataset is 5.864. The relative probability of correctly predicting winning greyhounds in
1
a race should approach 5.864 or 17.05% with increased data.

5.2.3 Comparative Performance in Greyhound Racing Prediction

Table 4: Comparative Performances

Model Greyhound Expert Random Chance Chen et al.

(Deep) Neural Network -4.78% +1.87% -1.08%
(Shallow) Neural Network -3.81% +2.84% -0.11%

Neural Network: A neural network was chosen for the modelling phase of this project
to analyse whether algorithmically limiting the feature space is more, less or equally as
effective as using domain expert knowledge in choosing an optimal subset of features. As
discussed in section 2.3.1, Johansson and Sönströd and Chen et al.; who limited their
feature space by seeking advice on what variables to add into their model by the same
experts they hoped their model would outperform; both used neural networks for their
modelling phase. This research shows that algorithmic and manual feature selection in
the domain of greyhound racing when combined with the neural network model have
comparable results, with the manual feature selection performing 0.11% better; as shown
in table 4.

6 Conclusion and Future Work

6.1 Conclusion
The aim of this research was threefold. The first was to bridge the gap of previous research
in the domain of greyhound racing prediction modelling by algorithmically subsetting the
feature space. The outcome of this research shows that the use of a machine learnt feature
subset can be effectively used in place of a domain expert knowledge feature set without
significant loss to prediction accuracy.

The second aim was to build a system to predict greyhound racing results. As discussed
in section 5.2.2, with an increased number of races the random probability of correctly
predicting the winner in a greyhound race should reach 17.05%. The models are tested
across 20% of the dataset, 2197 races, which should bring the probability close to 17%.
Table 3 shows that all 3 machine learning models preform better than random chance;
thus highlighting that the use of machine learning techniques is advantageous in predicting
greyhound racing results.
 Finally, the system built was to be bench marked against a human expert in greyhound
predictions. The system which is created in this research needs further tuning to rival the
hit ratio of the human expert. While the model’s performance is improved during tuning
the results in table 4 depict that the use of shallow neural networks correctly predict
3.81% less winners than the human expert.

It can be concluded that machine learnt feature selection must at all times be accom-
panied by domain knowledge; it is in combining the two that an optimal feature set can
be obtained. Although feature selection plays an important role in this research it is only
1 step within the iterative framework. It alone, cannot adequately account for a model’s
success or failure; rather the amalgamation of domain knowledge, feature engineering,
feature selection and model selection when combined optimally ensure success.

6.2 Future Works

Future work for this research should concentrate on the modelling phase. Suggestions on
possible work to bolster the success of this research are listed below:

1. Modelling in this paper is performed on each individual greyhound separately in

order to ascertain their individual probability of winning a race given their historical
performance data. The features subset developed in this research could be combined
with conditional statistics to allow that the sum of probabilities for all greyhounds
in a given race equal 1 so that within race competition can be accounted for.

2. Deep learning algorithms improve with added data. A future focus to improve
model performance could be to generalise the feature selection across all 28,271
races scraped from tracks throughout Ireland.

3. Utilize the optimal feature subspace in regression modelling.

Acknowledgements
This research would not have been possible without the help of my supervisor Mr. Oisı́n
Creaner, alongside Mr. Michael Bradford and Dr. Simon Caton, whose help and
guidance throughout the completion of this Masters programme proved invaluable. I
would like to take this opportunity to thank them sincerely for their generosity of time
and spirit.

I would also like to thank the Irish Greyhound Board who generously allowed me
to use their data in the completion of this Masters programme.

Without the continued support of my family and friends I would not have had the
courage nor strength to be able to complete my Masters. Particularly, I would like
to dedicate this thesis to my mam, Eileen Lyons, who sadly passed away before its
completion. Ar dheis Dé go raibh a h-anam.
References
Azevedo, A. and Santos, M. F. (2008). Kdd, semma and crisp-dm: a parallel overview.,
in A. Abraham (ed.), IADIS European Conf. Data Mining, IADIS, pp. 182–185.
URL: http://dblp.uni-trier.de/db/conf/iadis/dm2008.htmlAzevedoS08
Breiman, L. (1996). Bagging predictors, Machine Learning 24(2): 123–140.
URL: http://dx.doi.org/10.1007/BF00058655
Breiman, L. (2001). Random forests, Machine Learning 45(1): 5–32.
URL: http://dx.doi.org/10.1023/A:1010933404324
Brownlee, J. (2014). Machine learning mastery.
URL: http://machinelearningmastery.com/discover-feature-engineering-how-to-
engineer-features-and-how-to-get-good-at-it/
Butler, J., Tsang, E. P. K. and Sq, C. C. (1998). EDDIE beats the bookies.
Cain, M., Law, D. and Peel, D. (2003). The favourite-longshot bias, bookmaker margins
and insider trading in a variety of betting markets, Bulletin of Economic Research
55(3): 263–273.
URL: http://dx.doi.org/10.1111/1467-8586.00174
Chen, H., Rinde, P. B., She, L., Sutjahjo, S., Sommer, C. and Neely, D. (1994). Expert
prediction, symbolic learning, and neural networks: An experiment on greyhound ra-
cing., IEEE Expert 9(6): 21–27.
URL: http://dblp.uni-trier.de/db/journals/expert/expert9.htmlChenRSSSN94
Dash, M. and Liu, H. (1997). Feature selection for classification, Intelligent Data Analysis
1: 131–156.
Dash, M. and Liu, H. (2003). Consistency-based search in feature selection, Artificial
Intelligence 151(1–2): 155 – 176.
URL: http://www.sciencedirect.com/science/article/pii/S0004370203000791
Davoodi, E. and Khanteymoori, A. R. (2010). Horse racing prediction using artificial
neural networks, Proceedings of the 11th WSEAS International Conference on Nural
Networks and 11th WSEAS International Conference on Evolutionary Computing and
11th WSEAS International Conference on Fuzzy Systems, NN’10/EC’10/FS’10, World
Scientific and Engineering Academy and Society (WSEAS), Stevens Point, Wisconsin,
USA, pp. 155–160.
URL: http://dl.acm.org/citation.cfm?id=1863431.1863457
Fayyad, U., Piatetsky-Shapiro, G. and Smyth, P. (1996). The kdd process for extracting
useful knowledge from volumes of data, Commun. ACM 39(11): 27–34.
URL: http://doi.acm.org/10.1145/240455.240464
Guyon, I., Weston, J., Barnhill, S. and Vapnik, V. (2002). Gene selection for cancer
classification using support vector machines, Machine Learning 46(1): 389–422.
URL: http://dx.doi.org/10.1023/A:1012487302797
Han, J. (2005). Data Mining: Concepts and Techniques, Morgan Kaufmann Publishers
Inc., San Francisco, CA, USA.
Hand, D., Mannila, H. and Smyth, P. (2001). Principles of Data Mining, A Bradford
book, MIT Press.
URL: https://books.google.ie/books?id=SdZ-bhVhZGYC

Hodge, V. J. and Austin, J. (2004). A survey of outlier detection methodologies, Artificial

Intelligence Review 22(2): 85–126.
URL: http://dx.doi.org/10.1007/s10462-004-4304-y

Johansson, U. and Sönströd, C. (2003). Neural networks mine for gold at the grey-
hound racetrack, Proceedings of the International Joint Conference on Neural Networks
pp. 1798 – 1801 vol.3.

Kanto, A. J., Rosenqvist, G. and Suvas, A. (1992). On utility function estimation of

racetrack bettors, Journal of Economic Psychology 13(3): 491 – 498.
URL: http://www.sciencedirect.com/science/article/pii/016748709290006S

Kuhn, M. (2016). A short introduction to the caret package.

URL: https://cran.r-project.org/web/packages/caret/vignettes/caret.pdf

Kursa, M. B. and Rudnicki, W. R. (2016). Boruta package in r - documentation.

URL: https://cran.r-project.org/web/packages/Boruta/Boruta.pdf

Lantz, B. (2013). Machine Learning with R, Packt Publishing.

Leighton Vaughan Williams, D. P. (1997). Why is there a favourite-longshot bias in

british racetrack betting markets, The Economic Journal 107(440): 150–158.
URL: http://www.jstor.org/stable/2235276

Liu, B. (2010). Sentiment analysis and subjectivity, Handbook of Natural Language Pro-
cessing, Second Edition. Taylor and Francis Group, Boca.

Lyons, A. (2016). RIC research proposal for greyhound racing predictions modelling.

McCabe, A. and Trevathan, J. (2008). Artificial intelligence in sports prediction, Fifth

International Conference on Information Technology: New Generations .

Pang, B. and Lee, L. (2008). Opinion mining and sentiment analysis, Found. Trends Inf.
Retr. 2(1-2): 1–135.
URL: http://dx.doi.org/10.1561/1500000011

Pudaruth, S., Medard, N. and Dookhun, Z. B. (2013). Article: Horse racing prediction
at the champ de mars using a weighted probabilistic approach, International Journal
of Computer Applications 72(5): 37–42. Full text available.

Saeys, Y., Inza, I. n. and Larrañaga, P. (2007). A review of feature selection techniques
in bioinformatics, Bioinformatics 23(19): 2507–2517.
URL: http://dx.doi.org/10.1093/bioinformatics/btm344

Schumaker, R. P. and Johnson, J. W. (2008). An investigation of svm regression to

predict longshot greyhound races, Communications of the IIMA: Vol. 8: Iss. 2, Article
7 pp. 67–82.
URL: ”http://scholarworks.lib.csusb.edu/ciima/vol8/iss2/7”
Silverman, N. and Suchard, M. (2013). Predicting horse race winners through a regular-
ized conditional logistic regression with frailty, Journal of Prediction Markets 7(1): 43–
52.
URL: http://EconPapers.repec.org/RePEc:buc:jpredm:v:7:y:2013:i:1:p:43-52

Williams, J. and Li, Y. (2008). A case study using neural networks algorithms: Horse
racing predictions in jamaica., in H. R. Arabnia and Y. Mun (eds), IC-AI, CSREA
Press, pp. 16–22.
URL: http://dblp.uni-trier.de/db/conf/icai/icai2008.htmlWilliamsL08

Witten, I. H., Frank, E. and Hall, M. A. (2011). Data Mining: Practical Machine Learning
Tools and Techniques, 3rd edn, Morgan Kaufmann Publishers Inc., San Francisco, CA,
USA.
A Python Script Flow Chart

Figure 5: Python Script Flow Chart

B Caret and RandomForest Variable Importance Rank-
ings

Table 5: Variable Importance Results with caret & randomForest packages

Variable Importance in order

Rank Variable
1 DogsAge
2 OverallAvgTime
3 SecTimeAvg5
4 TimeAvg5
5 Weight
6 RankedGradeAvg5
7 PrizeMoneyWonAvg5
8 BreakAvg5
9 RaceNumber
10 PrizeMoneyToDate
11 DaysSinceLastRace
12 Avg2ndBend5
13 AvgRemarks5
14 FinishingPositionAvg5
15 Starts
16 TrapNumber
17 ShowPercent
18 RaceGrade
19 PlacedPercent
20 WinPercent
21 Colour
22 RaceDay
23 ShowPercent5
24 PlacedPercent5
25 WinsToDate
26 WinPercent5
27 Going
28 RaceDistance
29 Seed
30 Gender
C RFE with Random Forest Results

Table 6: Optimal Features as Selected by RFE & Random Forest

Feature Importance Rankings in order

Rank Variable
1 NumberOfDogs
2 TimeAvg5
3 OverallAvgTime
4 PrizeMoneyWonAvg5
5 DogsAge
6 FinishingPositionAvg5
7 SecTimeAvg5
8 PlacedPercent
9 RankedGradeAvg5
10 Avg2ndBend5
11 ShowPercent
12 RaceGrade
13 AvgRemarks5
14 BreakAvg5
15 WinPercent
16 ShowPercent5
17 PlacedPercent5
18 TrapNumber
19 Weight
20 RaceDay
21 RaceDistance
22 Seed
23 RaceNumber
24 Going
D RFE with Tree Bagging Results

Table 7: Optimal Features as Selected by RFE & Tree Bagging

Feature Importance Rankings in order

Rank Variable
1 DogsAge
2 Weight
3 TimeAvg5
4 SecTimeAvg5
5 OverallAvgTime
6 RaceNumber
7 PrizeMoneyToDate
8 BreakAvg5
9 RankedGradeAvg5
10 TrapNumber
11 PrizeMoneyWonAvg5
12 AvgRemarks5
13 RaceGrade
14 Avg2ndBend5
15 DaysSinceLastRace
16 FinishingPositionAvg5
17 WinPercent
18 PlacedPercent
19 ShowPercent
20 Starts
21 Colour
22 RaceDay
23 ShowPercent5
24 PlacedPercent5
25 RaceDistance
26 Going
27 Seed
28 WinPercent5
29 WinsToDate
30 Gender
31 NumberOfDogs