Available via license: CC BY 4.0
Content may be subject to copyright.
What AI can do for horse-racing?
Pierre Colle ,Alezan AI
pierre@alezan.ai
Abstract
Since the 1980s, machine learning has been widely
used for horse-racing predictions, gradually ex-
panding to where algorithms are now playing a
huge role as price makers in the betting market.
Machine learning has changed the horse-racing
betting market over the last ten years, but main
changes are still to come. The paradigm shift of
neural networks (deep learning) may not only im-
prove our ability to simply predict the outcome of a
race, but it will also certainly shake our entire way
of thinking about horse-racing - and maybe more
generally about horses.
Since 2012, deep learning provided more and more
state-of-the-art results in computer vision and now
statistical learning or game theory. Following Tuyls
et al. (2021), we describe how the convergence of
the three machine learning fields (computer vision,
statistical learning, and game theory) will be game-
changers in the next decade in our ability to predict
and understand horse-racing.
We consider that horse-racing is a real world lab-
oratory where we can work on the animal-human
interaction and build a non-anthropocentric Artifi-
cial Intelligence. We believe that this will lead us
to understand the horses better and the interactions
between animals and humans in general.
1 Introduction
The current work has been largely inspired by the recent work
regarding AI in football (Tuyls et al. (2021)). This article’s
structure will follow the same structure by first studying the
different fields of AI and then the frontiers between those
fields.
We will study AI from the viewpoint three different axes:
• Computer Vision (CV) – AI analyzes videos and images
to extract information
• Statistical learning (SL) – build models and predict the
race outcome.
• Game theory (GT) – modelling cross-competitor inter-
action and understanding their strategies against each
other.
Within these axes, we will go further into each frontier :
• Frontier 1: (SL and GT) – how to combine statistical
learning with game theory
• Frontier 2: (CV and SL) – how to combine Computer
Vision with statistical learning
• Frontier 3: (CV and GT) – how to combine computer
vision with game theory
Several works have been published regarding soccer, base-
ball (Kelly (2021)), and basketball (Li & Xu (2021)) in the
last few years, but much less has been published regarding
horse-racing. This is mainly because of the betting-centric
nature of the horse-racing ecosystem that makes it very se-
cret. Anyone who has a good idea about horse-racing will
keep it to himself and not share his betting advantage with
other punters, as secrecy is often the best strategy to protect
intellectual assets (see Cohen et al. (2000)). This is even more
important in parimutuel betting, where punters are not betting
against the casino, but against one another.
2 AI Fields
2.1 Statistical Learning
Since the 19th century, horse-racing has been a special field –
maybe the most iconic – for statistical analysis. In the late
20th century with the emergence of computer capabilities,
statistical learning has been used to predict the outcome of
races and then to “beat the odds”.
By 2018, the algorithms used for betting had deeply
changed the horse-racing industry (Malachane & Donner
(2018)). This market is dominated by a few syndicates con-
taining many workers – up to 300 employees (according to
Ziemba (2020)). Those syndicates are not publishing or com-
municating their work.
Current work is limited to public knowledge, not dis-
cussing the actual methods used by betting syndicates.
The first usage of statistical learning is to build a “handi-
capping model” (see Benter (1994)), which basically predicts
a score on each horse about his ability to win the race. To
build such a model, one can leverage a large amount of data
available for each horse. This usually includes:
• The horse’s past results
–the amount of money earned
arXiv:2207.04981v1 [cs.LG] 11 Jul 2022
Figure 1: Diagram of the different fields of AI and the corresponding Frontiers, inspired by Tuyls et al. (2021)
–lifetime records
• Information about the breed, like the dosage index (Ro-
man (2003)).
We are also gradually switching to “big-data” approaches
based on larger datasets like:
• Genomic data (see Danvy & Sabbagh (2018))
• GPS tracking, Nano-bio sensors (see Evans & Mc-
Namee (2020))
• Horserace videos and images, as discussed in Section
2.2
The main challenge for statistical learning in horse-racing
is to leverage a large amount of data, with different structures
inside an integrated end-to-end trainable system.
These heterogeneous big data situations will raise more
challenges for feature engineering (see Wang (2017)). The
perspective of using deep learning to automatically focus on
key data seems to be an obvious yet challenging opportunity.
Handling Historical Data
The most important information about horses on a track is
their past records.
Past records are time-series, and the ability to manage time
series is still a challenging topic (see Ismail Fawaz et al.
(2019)). However, when looking more closely at the past
records of a horse, we would like to consider not only as a
time series but more as a tree-structured graph (see figure 2
from Ko (2018)).
The state-of-the-art of the deep-learning on tree-structured
data is still moving fast, and while some works are using
RNN-based architecture (see Chakraborty et al. (2018); Dong
& Lapata (2016); Alvarez-Melis & Jaakkola (2017)), the
complexity and sparsity of the tree structures make the re-
sulting network rather slow (see Harer et al. (2019))
In the context of horse-racing, to train an end-to-end net-
work understanding historical data, we will face two major
issues:
• Sparsity – most horses participate in 15 to 30 races in
their lives (see Ko (2018)). There are about 10,000
horses trained for thoroughbred racing in France (see
Cl´
es (2021)), so most of the horses never face one an-
other.
• Complexity – given that each race has 10 competitors,
to end-to-end train a network while considering the his-
Figure 2: Image reproduced from Ko (2018) with the authorisation
of the author, Ko Ohashi
torical performance of the 10 last races for each horse,
there will be about 100 000 competitors (input vectors)
to take into account.
Building the Loss
The other main challenge for an efficient end-to-end trained
handicapping model is to build the loss.
The naive approach would be to work on a classifier win-
ner/not winner, and then use a binary cross-entropy loss (see
example in Ko (2018)).
This classifier may give first results but will only use a tiny
portion of the available data, considering that the first horse is
the only interesting thing to take into account while training
the neural network. More complete usage of data will make
it possible to leverage all the available data :
• Final rank information
• Distance on arrival between horses
• Detailed race proceedings
To use the final rank information and the distance on ar-
rival between horses, the learning-to-rank (see Liu (2007))
information-retrieval research seemed to tackle this exact
problem, and libraries are now available publicly to work on
this (see for example Pasumarthi et al. (2019)), the main idea
here is to consider that what matters in not only who is the
winner but what is the exact ordering of the output.
The detailed proceeding of the race is also difficult to
leverage in a loss function, but it could be possible to effi-
ciently leverage the data by using counterfactual analysis as
described in Frontier 1 (see section 3.1).
2.2 Computer Vision
Today’s Computer Vision systems can tackle horse-racing
and numerous challenges with human-based sports like:
• Activity recognition
• Pose estimation – studying the way the horse is mov-
ing and if he suffers from specific injuries. (Chaneyddtt
(2018))
Figure 3: Horses often occlude one another in a horse
race and it is challenging to precisely draw bounding boxes
around each one. Photo made by Peter Miller, from
https://www.flickr.com/photos/pmillera4/40952321920
• Tracking the horses in the race (see for example
Alezan.ai (2021))
These technologies work well in labs, but it is still chal-
lenging to connect them to real-world broadcasting systems
and to pass the usability test for real-world applications.
The main issues with real-world data are:
• The constant changes of cameras in broadcasting sys-
tems, and the need to geometrically combine them (as
described for football in Tuyls et al. (2021))
• The number of real-life scenarios (see a famous example
in Vincent (2020))
• The high number of occlusions present in horse-racing
real-life data, where all the horses are running side by
side (see Figure 3)
A lot more usages for computer vision will arise in sub-
sequent years like emotion recognition or real-time data ex-
traction, but these usages will often be related also to Game
Theory or Statistical Learning, those are discussed in Frontier
2 (see section 3.2) and Frontier 3 (see section 3.3).
2.3 Game Theory
Game theory is about players behavior and decisions. In
horse-racing with strategic decisions are made at different
levels :
• Season level: Participating or not in a race
• Race level: What is the strategy ?
• Action level: When to draft and when to attack ?
Season Level: Participating or not in a race
When an owner makes one of his horses participate in a race,
the choice is made for a reason, and this decision can be
viewed as a game-related decision. From the owner’s point
of view, he can get the following benefits:
• The allocation (money paid to the winner, but also the
2nd, 3rd, 4th and 5th horse)
• The visibility of the horse may raise its value on the
horse market
Figure 4
But each race also has a cost for the owner:
• The horse can be injured and lose its value on the market
• Participating in a race costs money, food, transportation
We could ask for each competitor at the beginning of the
race:
• Why is this horse participating in this race?
• What is his target rank (in the owner’s point of view)?
Apart from the participation of the race, the owner (and the
trainer) needs to decide under what conditions the horse will
participate to this race:
• Which rider should run with the horse? This may in-
fluence not only the short-term outcome of the race but
also the long-term welfare of the horse (Munsters et al.
(2012) have linked, for show-jumping the welfare of the
horse with the choice of the rider).
• What type of blinker (Winkers/blinkers/none) ? De-
pending on the horse’s character, blinkers are a way to
limit his field of view and to limit intimidation by the
crowd/other horses.
• What type of horseshoe (Barefoot/Horsehoes) ? While
Barefoot are faster, this is also more dangerous for the
horse health
• Training intensity and warm up intensity and duration
Race Level: What is the Strategy?
Specific parts of the race may be amenable to game theory,
like the beginning of the race, the positioning inside or out-
side the race field, or deciding how to balance the effort over
the race. The first strategical dilemma is related to the type of
race. Is it a slow, medium, or fast race?
It has been studied for other sports like running. By study-
ing the medallists in the 1500m Championships, it is clear that
the kick of the race (the final sprint) in the last lap depends on
how fast the previous laps were (Sandford et al. (2019)). The
winning strategy regarding when to start the sprint in a race
depends on the speed of the race.
In any case, this “kick” approach of the 1500m running
races is not as relevant for 400m or 800m distances, as the
world-best records are made with athletes who can run the
first half of the race faster than the second half. This sug-
gests that for his best performance, an athlete needs to reach
a high-intensity effort during the whole race that will impact
his performance in the final meters (Sandford et al. (2019)).
The regulation has been studied for horse-racing also, and
it’s seems that it is more similar to the 400m or 800m as the
winning horse is often the one that is able to slow down the
least at the end (see Mercier & Aftalion (2020)).
These considerations regarding “kick” and strategy are
very important for horse-racing because unlike running where
the muscle and the strategic brain belong the same individual,
the horse-rider who decides the tactical aspects of the race
needs to manage the muscles of a horse. This muscle/brain
duality makes horse-racing a very interesting topic to study
for AI (see section 3.1)
Action Level: When to Draft and When to Attack
In the race, each rider needs to handle his tactic between at-
tacking or drafting. This is well explained in the cycling-
related literature (see Mignot (2016)), but it can be easily ex-
tended to horse racing as the average speed of horses was
found to increase with the percentage of the race spent on a
drafting position (Spence et al. (2012)).
The question of whether to use whipping or how to use
it is also a scientific research topic, and had been discussed
regarding horse-welfare. The amount of whip-related pain for
horses whereas recent studies have shown that whipping has
no impact on horses’ speed (see McGreevy et al. (2012) and
Wilson et al. (2018)), and the question of whether whipping
is an aid for the steering (and for the game-theoretic aspect of
the horse riding) is still a matter of debate in the community
(Thompson et al. (2020)).
3 AI Frontiers
3.1 Frontier 1: Interaction Between Horse and
Rider (GT and SL)
The combination of Game Theory and Statistical Learning
is a main topic for horse-racing handicapping models. One
of the key improvement in next years AI (as discussed for
football in Tuyls et al. (2021), and implemented for basketball
in Alcorn & Nguyen (2021)) is to train reinforcement learning
to predict the horse’s next move and use this to compare the
predicted moves with the existing ones.
We may consider that statistical learning will give us the
ability to train, using reinforcement learning, the typical be-
havior of this specific horse in a typical situation. Then,
by comparing the expectations from statistical learning with
the actual behavior of the horse (counterfactual analysis), we
could extract the “unexpected” behaviors, which in the con-
text of machine learning could be interpreted as “tactical”
moves.
Traditionally, for analysis purpose, horse races are divided
by predefined sectors, first sector are the 400 first meters, last
sector are the 400 last meters. This split gives the ability to
consider the race sector per sector and to get information not
only about the arrival, but also about the proceeding of the
race.
Considering the large amount of information given by big
data approach, the first benefit of counterfactual analysis -
(as explained in Tuyls et al. (2021) - is to train an attention
mechanism given the ability, user counter-factual analysis, to
focus on relevant sequences. Then we would replace a non-
trained handmade sector traditional point of view, by a dy-
namic, trainable attention-based mechanism.
The main idea of counter-factual analysis is to compare the
actual behavior of a horse with the expected behavior of the
horse in our model. This difference could be, following the
mathematical definition of information by Shanon (see Shan-
non (1948)), seen simply as an information extraction, or a
compression mechanism, focusing only on the unexpected.
Another benefit, By assuming that the tactical moves are
made by the rider and by doing the statistical learning on the
horse only (with horse historical data), one may even use this
counterfactual analysis to extract the rider’s strategic moves
(see Spence et al. (2012)]) from the (less strategical) horse
behavior.
This analytical split between horse-related and rider-
related components will be very useful to assess the role of
the horse and the rider on a specific result or ranking, and
then it can provide statistical learning with some very accu-
rate analysis to work on.
3.2 Frontier 2: SL and CV
Computer Vision can increase the amount of data available
for statistical learning. A lot of important data are available
right now (GPS trajectories, historical data), but some other
information is still only in the raw videos. High dimensional
videos and the ability to extract that information on a large
scale might feed new statistical learning models.
Some computer-vision accessible data are:
• An event such as horsewhips and falls (In Pinchbeck
et al. (2010), researchers are using video recording as
a primary data source to study the horsewhips)
• The position of the horses’ legs and the frequency of
their movement (As in Mathis et al. (2021))
• The horses’ emotions (nervousness, attention level).
Some studies show that the horses’ facial expressions,
like the orientation of the head and position of the ears
and eyes, yield information about the horses’ attention
(see Wathan & McComb (2014)).
• The horse’s eagerness to enter the starting gates. This in-
formation is already scrutinized by professional punters.
(see Educating the Punter [1994])
The first goal for this kind of study would be, obviously, to
improve handicapping model; but this deep image-to handi-
capping approach can be extremely informative to understand
the horse’s (and more generally animals) behavior. The main
advantage of deep learning compared to other approaches is
the ability to remove the bias in the study and then let the
neural network decide what to look at in a specific image.
As explained in Montavon et al. (2018) we are able to use
visualization techniques to understand meaningful patterns in
the input data and understand, which part of an image plays a
role in final neural network decision.
Having the ability to train an end to end neural network
from videos to race prediction, may give a first answer to
the question of whether the ears’ position before the race has
Figure 5: Behavioral score used to assess the nervosity of a horse
in Munsters et al. (2012). Reproduced with the authorization of the
author
been used by the network to predict the outcome of the race.
This new approach will may anyway gives new tools to study,
on a large scale basis, the correlations between horse emo-
tions, visual markers and race result.
3.3 Frontier 3: GT and CV
In the same way that Computer Vision can be used to im-
prove statistical learning on Frontier 2, we can also use it to
feed game-theoretic models. A lot of Computer-Vision acces-
sible information can provide relevant information for game-
theoretic models such as:
• A warm-up session format: knowing how long and how
fast the warm up is will give us information about the
horse and the owner expectations.
• The global appearance of the horse and some very sim-
ple considerations like the horse’s hair dress will give us
an idea if the horse owner expects to show his horse in
this race.
• The body movements of the horse-rider will provide in-
formation about his relationship with the horse.
A study regarding horse welfare (Munsters et al. (2012))
used a behavioral score built mainly on visual observations
(see Figure 5). It shows that this behavioral score can be used
to measure the matching between a horse and its rider. Only
21 riders and 16 horses were included in the study
Using Computer Vision and working on horse racing, we
will be able to generalize this study for every broadcasted
race; instead of working on 16 horses, we could extract
this kind of behavioral score (see Figure 5) for thousands of
horses and riders competing each year in a league.
Changing the order of the magnitude of this data extraction
will create a lot of opportunities. First of all, it may impart
the ability to characterize horse and rider personality, by cre-
ating a rider-temperament vector for each rider and a horse-
temperament vector for each horse, as it is made in soccer for
players in ?.
Once we have a rider-temperament vector for each rider
and a horse-temperament vector for each horse, it would be
quite direct to give an “impairment score” to every horse-rider
couple even if they had never ridden together before. This
may be used to help the coaches choose the rider to ride a
specific horse.
But this vector approach once trained using the numer-
ous available videos of horse racing could be used outside of
horse-racing for other equine usages: to match a horse with a
rider in contexts like show-jumping when less data are avail-
able but the behavioral score and horse-rider matching could
be transferred.
Eventually, as discussed in section 3.2, working with a
huge horse-racing video dataset gives the opportunity to re-
build a behavioral score analysis and let the neural network
decide exactly what to look for in the image to decide whether
the horse and the rider are matching or not. It sounds feasi-
ble that AI will teach us (humans) how and what to look and
horses, without having all the anthropocentric bias we tend to
project on horses when we look at them.
4 Conclusion: What AI can do for
horse-racing
In the next decades, AI may have the ability to fulfill some
completely new scenarios on horse-racing. These include:
• Improving the odds and making them much more rele-
vant, while taking more and more information into con-
sideration.
• Augmented reality will improve the way people watch
horse-racing, making it more fun to watch horseraces
and bet on a specific horse.
• Splitting the performance between the strategic moves
of the rider and the physical performance of the horse.
• Making an “impairment score” between a horse and a
rider to help horse trainers in their choices.
• Telling the trainer, owners, and the public exactly what
to look at before and during the race to best predict a
horse’s ability.
• Teaching how to look at horses, how to understand their
behavior, when they are nervous, and where they are not.
Our final concern is to know whether this AI revolution
will be made in public labs and published research, aiming
to spread technologies, tools, and knowledge to everyone, or
if it will be made privately and secrecy with only short-term
interest.
In the last decades, the recent AI revolution (with the exam-
ple of Facebook and Google) needed both public and open-
source research communities working together along with
private interests. We feel that AI will encourage horse-racing
and the betting ecosystem to open themselves, open-source
algorithms, and share papers and knowledge.
5 Discussion: What Can horse-racing do for
AI?
Horse racing is an open laboratory with a huge amount of
existing data, existing private funding, and real-life competi-
tion, but it is still missing a public research communities and
widespread technologies for Computer Vision.
Current AI is a very anthropocentric field with, for exam-
ple, human-only pose estimation and human-only emotion
recognition. It would be interesting for AI to focus on a
less anthropocentric approach and use horse racing as a large-
scale real-world laboratory to experiment with the ability of
AI to predict human-animal-related outcomes.
Finally, we always ask ourselves what we can teach AI,
but maybe we should think more about what AI can teach us?
The ability for AI in terms of horse racing to help us under-
stand the horse’s emotion is an opportunity for it to create
knowledge about an animal’s emotions and characters.
References
Alcorn, M. A., & Nguyen, A. (2021). baller2vec:
A multi-entity transformer for multi-agent spatiotemporal
modeling. arXiv preprint arXiv:2102.03291, .
Alezan.ai (2021). Augment betting experience. https://www.
alezan.ai/blog/equidia-course-de-reference.
Alvarez-Melis, D., & Jaakkola, T. (2017). Tree-structured
decoding with doubly-recurrent neural networks. In ICLR.
Benter, B. (1994). Computer Based Horse Race
Handicapping and Wagering Systems: A Report.
https://medium.com/parimutuel-racetrack-analysis/bill-
benter-1994-computer-based-horse-race-handicapping-
and-wagering-systems-a-report-db747c250e77.
Chakraborty, S., Allamanis, M., & Ray, B. (2018). Tree2tree
neural translation model for learning source code changes.
CoRR,abs/1810.00314. URL: http://arxiv.org/abs/1810.
00314. arXiv:1810.00314.
Chaneyddtt (2018). GitHub - chaneyddtt/UDA-Animal-Pose.
https://github.com/chaneyddtt/UDA-Animal-Pose.
Cl´
es, C. (2021). France galop. https://www.france-galop.
com/fr/chiffres-cles.
Cohen, W. M., Nelson, R., & Walsh, J. P. (2000). Protect-
ing their intellectual assets: Appropriability conditions and
why us manufacturing firms patent (or not).
Danvy, S., & Sabbagh, M. (2018). Indices g´
en´
etiques
et de performance chez les ´
equid´
es. https://equipedia.
ifce.fr/elevage-et-entretien/elevage/selection-et-indices/
indices-genetiques-et-de-performance-chez-les-equides.
Dong, L., & Lapata, M. (2016). Language to Logical Form
with Neural Attention. Stroudsburg, PA, USA: Associa-
tion for Computational Linguistics. doi:10.18653/v1/
p16-1004.
Evans, R., & McNamee, M. (2020). Sports Betting,
Horse Racing and Nanobiosensors – An Ethical Evalua-
tion. Sport, Ethics and Philosophy,15, 208–226. doi:10.
1080/17511321.2020.1727946.
Fereira (2018). Multi-agent implementation.
https://github.com/diogodanielsoaresferreira/
RMI-basedDistributedSystem.
Harer, J. A., Reale, C. P., & Chin, P. (2019). Tree-
transformer: A transformer-based method for correction of
tree-structured data, . doi:ArXiv:abs/1908.00449.
Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L., &
Muller, P.-A. (2019). Deep learning for time series classifi-
cation: a review. Data Mining and Knowledge Discovery,
33, 917–963. doi:10.1007/s10618-019-00619-1.
Kelly, T. (2021). What Baseball Can Teach Us About
AI in Business. https://www.netsuite.com/portal/
business-benchmark-brainyard/industries/articles/
cfo-central/artificial-intelligence-in-business.shtml.
Ko (2018). Deep learning and horse race predic-
tion #1 - ko. https://kouohhashi.medium.com/
deep-learning-and-horse-race-prediction-1-ed65727f9d27.
Li, B., & Xu, X. (2021). Application of Artificial In-
telligence in Basketball Sport. Journal of Education,
Health and Sport,11, 54–67. doi:10.12775/jehs.
2021.11.07.005.
Liu, T.-Y. (2007). Learning to Rank for Information Re-
trieval. Foundations and Trends in Information Retrieval,
3, 225–331. doi:10.1561/1500000016.
Malachane, O., & Donner, C. (2018). Hold-up
sur le PMU. https://www.lesechos.fr/2018/04/
hold-up-sur-le-pmu-1020178.
Mathis, A., Biasi, T., Schneider, S., Yuksekgonul, M.,
Rogers, B., Bethge, M., & Mathis, M. W. (2021). Pre-
training boosts out-of-domain robustness for pose estima-
tion. In Proceedings of the IEEE/CVF Winter Conference
on Applications of Computer Vision (pp. 1859–1868).
Mayton, T. (2014). Celtic FC Adds Live Betting to
Stadium App. https://mobilemarketingmagazine.com/
celtic-fc-adds-live-betting-to-stadium-app/.
McGreevy, P. D., Corken, R. A., Salvin, H., & Black,
C. M. (2012). Whip Use by Jockeys in a Sample of
Australian Thoroughbred Races–An Observational Study.
PLoS ONE,7, e33398. doi:10.1371/journal.pone.
0033398.
Mercier, Q., & Aftalion, A. (2020). Optimal speed in
Thoroughbred horse racing. PLOS ONE,15, e0235024.
doi:10.1371/journal.pone.0235024.
Mignot, J.-F. (2016). Strategic Behavior in Road
Cycling Competitions. The Economics of Profes-
sional Road Cycling, (pp. 207–231). doi:10.1007/
978-3-319-22312-4_10.
Montavon, G., Binder, A., Lapuschkin, S., Samek, W., &
M¨
uller, K.-R. (2018). Layer-wise relevance propagation:
an overview. Explainable AI: interpreting, explaining and
visualizing deep learning, (pp. 193–209).
Munsters, C. C., Visser, K. E., van den Broek, J., & Sloet van
Oldruitenborgh-Oosterbaan, M. M. (2012). The influence
of challenging objects and horse-rider matching on heart
rate, heart rate variability and behavioural score in riding
horses. https://www.sciencedirect.com/science/article/pii/
S1090023311001493.
Pasumarthi, R. K., Bruch, S., Wang, X., Li, C., Bendersky,
M., Najork, M., Pfeifer, J., Golbandi, N., Anil, R., & Wolf,
S. (2019). TF-Ranking. New York, NY, USA: ACM.
doi:10.1145/3292500.3330677.
Pinchbeck, G. L., Clegg, P. D., Proudman, C. J., Morgan,
K. L., & French, N. P. (2010). Whip use and race progress
are associated with horse falls in hurdle and steeplechase
racing in the uk. Equine Veterinary Journal,36, 384–389.
doi:10.2746/0425164044868387.
Roman, S. A. (2003). Dosage: Pedigree
& Performance. https://www.amazon.in/
Dosage-Pedigree-Performance-Steven-Roman/dp/
0929346734.
Sandford, G. N., Day, B. T., & Rogers, S. A. (2019). Rac-
ing Fast and Slow: Defining the Tactical Behavior That
Differentiates Medalists in Elite Men’s 1,500 m Champi-
onship Racing. Frontiers in Sports and Active Living,1.
doi:10.3389/fspor.2019.00043.
Shannon, C. E. (1948). A mathematical theory of communi-
cation. The Bell system technical journal,27, 379–423.
Spence, A. J., Thurman, A. S., Maher, M. J., & Wilson, A. M.
(2012). Speed, pacing strategy and aerodynamic drafting in
Thoroughbred horse racing. Biology Letters,8, 678–681.
doi:10.1098/rsbl.2011.1120.
Thompson, K., McManus, P., Stansall, D., Wilson, B. J., &
McGreevy, P. D. (2020). Is Whip Use Important to Thor-
oughbred Racing Integrity? What Stewards’ Reports Re-
veal about Fairness to Punters, Jockeys and Horses. Ani-
mals,10, 1985. doi:10.3390/ani10111985.
Tuyls, K., Omidshafiei, S., Muller, P., Wang, Z., Connor, J.,
Hennes, D., Graham, I., Spearman, W., Waskett, T., Steel,
D. et al. (2021). Game plan: What ai can do for football,
and what football can do for ai. Journal of Artificial Intel-
ligence Research,71, 41–88.
Vincent, J. (2020). Ai camera operator repeatedly con-
fuses bald head for soccer ball during live stream.
https://www.theverge.com/tldr/2020/11/3/21547392/
ai-camera-operator-football-bald-head-soccer-mistakes.
Wang, L. (2017). Heterogeneous data and big data analytics.
Automatic Control and Information Sciences,3, 8–15.
Wathan, J., & McComb, K. (2014). The eyes and ears are
visual indicators of attention in domestic horses. Current
Biology,24, R677–R679. doi:10.1016/j.cub.2014.
06.023.
Wilson, B. J., Thompson, K. R., & McGreevy, P. D. (2018).
The race that segments a nation: Findings from a conve-
nience poll of attitudes toward the Melbourne Cup Thor-
oughbred horse race, gambling and animal cruelty. PLOS
ONE,16, e0248945. doi:10.1371/journal.pone.
0248945.
Ziemba, W. T. (2020). Parimutuel betting markets: Race-
tracks and lotteries revisited. https://www.systemicrisk.ac.
uk/sites/default/files/2020-09/dp-103.pdf.