See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/251880935

Analyzing the Placement Odds of Favorite Horses in the Thoroughbred Racing

Industry of the British Isles

Article in CHANCE · December 2011

DOI: 10.1007/s00144-011-0038-1

CITATIONS READS

0 3,014

2 authors, including:

Fernando Mata
Instituto Politécnico de Viana do Castelo
119 PUBLICATIONS 346 CITATIONS

SEE PROFILE

All content following this page was uploaded by Fernando Mata on 08 August 2014.

The user has requested enhancement of the downloaded file.

 Newcastle University ePrints

Mata F, Watts S. Analyzing the placement odds of favourite horses in the

thoroughbred racing industry of the British Isles. Chance 2011, 24(4), 35-40.

Copyright:

This is an Author's Accepted Manuscript of an article published in Chance, 2011, copyright Taylor &
Francis, available online at: http://www.tandfonline.com/doi/abs/10.1080/09332480.2011.10739885

Always use the definitive version when citing.

Further information on publisher website: http://www.tandfonline.com

Date deposited: 4th December 2013

Version of article: Author final

This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported License

ePrints – Newcastle University ePrints

http://eprint.ncl.ac.uk
Analyzing the placement odds of

favorite horses in the thoroughbred

racing industry of the British Isles.

Fernando Mata

Sarah Watts

This article describes a logistic regression model to estimate the probability of at

least one of the two “favorite to win” horses, being placed in the end of the race (ending up in

one of the three first positions), based on the number of horses entered and on the time of the

day the race takes place (day, evening), and handicapping (handicapped race or not). The

association between placement of favorites and the following variables, was also explored

but not found significant: type of race, type of surface, surface conditions, and whether or not

it is a stakes race.

The Thoroughbred horse racing industry is a multimillion pound global

industry. Annually in the United Kingdom (UK) the sport attracts almost 6 million

spectators, not only to view but to place bets on, the 9000 plus races that are run. The
turnover of the horserace betting market in the UK alone in 2006 has been reported to be

in excess of £10 billion. However, the money does not only lie within the betting market

itself as the racing industry supports the wider economic community not only in the UK

but worldwide. The American Horse Council put the impact of the racing industry on the

economy in the United States of America (USA) at $26.1 billion, which puts the industry

on a par with one of the 75 largest companies in the USA. Similar importance has been

stated on the impact of the industry on the Australian economy, and certainly the same

kind of information could easily be found for several other different countries.

A lot of research has been done regarding the prediction of final results in

thoroughbred horse racing; from predictive modeling (e.g. a E.M. White and coauthors

article published in the “International Journal of Forecasting” in 1992) to pre-race

behavior (e.g. a G.D. Hutson and M.J. Haskell article published in “Applied Animal

Behavior Science” in 1997) and heart size (L.E. Young and coauthors, published in the

“Journal of Applied Physiology” in 2005). Competition data has also widely been used,

across all equine sports, in genetic evaluation and breeding value studies. Economy and

betting markets (M.A. Smith, and L.V. Williams, article published in the “International

Journal of Forecasting” in 2010), favorite-long shot bias and bettor preferences (L.D.

Brown and coauthors article published in “CHANCE” in 1994) are examples of other

topics picked up by research, to predict final results. However, anecdotal information is

still, and probably will always be, widely used by bettors to justify the odds.

A simple introduction to the structure of thoroughbred racing

Horse racing can be split in two types: flat racing and national hunt racing.

Both types are widespread across the more than 60 racecourses in the British Isles.
Racecourse circuits can be left- or right-handed, flat or undulating, triangular, square,

oval, figure of eight shaped or even just straight. The majority of races are run on turf but

there are five courses offering the alternative to run flat races on an all weather track

(AWT).

A flat race is run over a predetermined distance in which the horses are not

required to jump any obstacles. Total distances vary from 5-16 furlongs (1-3.2 km) and

are generally split into sprints, mile and classic distances or stayers races.

National hunt (NH) racing is split into steeplechases and hurdle races. They

are run over distances of 2-4.5 miles (3.2-7.2 km) and include solid obstacles at a

minimum height of 4’6’’ (~1.37 m). Hurdle races are run over distances of 1.75-3 miles

(1.9-4.8 km) and include less dense obstacles at a minimum height of 3’6’’ (~1.07 m).

National hunt racing also includes flat races known as national hunt flat races or bumpers

races designed for young horses. Hunter-chases are another classification of NH racing,

designed for horses that have regularly hunted throughout the season.

In stake races part of the prize money is put up by the owners of the horse

running. Competitors often race against horses of the same age, gender or class.

Conditions races may also involve all horses carrying the same weight or with weights

being adjusted to reflect age, races of a certain value or number of races won. In a

handicap race all horses are allocated weights to be carried based on their previous

performance. The principle of handicapping is to allow every horse an equal chance of

winning. In other words the horse deemed to have the best chance in a race will carry the

most weight and the least fancied runner will carry the least weight.
Previous research

Rowe in his book from 2004, “how to win in horseracing” discusses 2 studies
with 2,317 and 10,466 races, where the favorite won 33% and 32.6% of the times
respectively, and got placed around 50% of the times. Rowe’s data comes from the USA
where flat races on dirt tracks are common. In contrast, UK races are more likely to be on
turf, and there is mixture of flat races and races that include jumps.
Duncan’s 2005 book “Winning horse racing formulae”, provides a British
perspective, with the analysis broken down and presented according to race type,
handicap and age range. When considering the figures given by Duncan for just flat
racing, the average percentage for one of the top three winning is 74%. Rowe calculated a
top three win percentage of 69%.
Both Rowe and Duncan discuss studies carried out into finishing positions in
Thoroughbred horseracing but neither considers in depth the effects of variables such as
type of race, age and surface.
G.D. Hutson, and M.J. Haskell, conducted a study into pre-race behavior and its

importance in allowing the prediction of race winners in their 1997 article in “Applied

Animal Behavior Science”.

They found that no single behavior or appearance could improve upon the
predictions made by the betting market. By considering multiple behavior variables,
howevers, horses deemed unable to win could be eliminated, and therefore their results
were considered of potential high economic value.
A number of other studies have used predictive modeling and forecast
combination in order to predict future racing outcomes. S. Lessmann and coauthors in
their “European Journal of Operational Research” article (2009) use a complex two
stage model: First a machine learning technique (support vector machines for
classification), estimates the likelihood of a given runner being a winner. Then a
conditional logit (CL) model, estimates the wining probability of one horse in
conjunction with the other competitors. This is a 2 stage approach based on W. Benter
(Efficiency of Racetrack Betting Markets, 1994), where support vector regression (SVR)
was used to model the relationship between variables and horses’ finish positions, being
then combined with horses final market prices (odds) using a CL model in a second step.
These 2 steps were found to be complementary to each other, once CL accounts for
within race competition, whereas SVR uses a large number of variables and
automatically models non-linear relationships between these.
E.M. White and coauthors test the accuracy of a forecast combination of
judgmental and statistical methods, to predict wining outcomes in horse races. These
authors caution that “horse racing possesses a strong element of random chance which
makes it impossible to predict every winner with perfect accuracy”.
The bulk of current research into surface and racetrack design is aimed at
minimizing the risk of injury to the locomotory system of the horse. Such information is
of high economic value if it can extend the racing careers of the horses. Studies into
performance and environmental factors are contradictory, such as that by M.D.S. Mota
and coauthors in 1998 (“Journal of Animal Breeding Genetics”), reporting varying
effects of surface type. It has been suggested that turf tracks are much faster than sand
tracks, with light turf producing the best results. The same study also found that drenched
sand was faster than smooth, dry sand, but this contradicts results found by R.L. Hintz,
and L.D. Van Vleck (published in 1978 in the “Journal of Animal Science”) who
suggested dry sand to produce faster times due to less resistance.
Most of the race type studies have also been directed at injury risk and

reduction. An example of an extensive study into this area was carried out by J.R.

Newton and coauthors (published in 2005 in the “Equine Veterinary Journal”) who

found that the risk of epistaxis increased with an increasing racing effort. This meant that

the incidence of blood visible at the nostrils increased between flat and hurdle racers and

again between hurdlers and chasers.

Method and Results

To estimate the probability of a favored horse placing, we considered data

from The Racing Post for all thoroughbred races in the British Isles between June 1 and

October 31, 2008. The date, racecourse location, surface, period of the day the race took

place, ground conditions, number of runners, type of race, handicapped or not and the

finishing positions of the favorite and second favorite were noted.

Previous research did not use the technique used in this study. We used a

logistic regression where the outcome dichotomy was considered to be: at least one of the

favorites placed or none of the favorites placed. The predictors (independent variables)

used were: the ground condition or “going” (heavy, soft to heavy, soft, good to soft,

good, good to firm, firm and hard); the period of the day the race took place (day,

evening); handicapped race or not; type of race (flat, national hunt flat, chase, hurdles,

stakes); track surface (turf, AWT); and number of runners. All the variables were used as

factors, with the exception of the number of runners, which was used as a covariate. A

runner is considered to be placed if it ends up in one of the three first positions in the

race.

As the “going” for AWT is classed differently (standard, standard to fast and

standard to slow) from the turf, the AWT races were initially not considered in the model,

so that the “going” condition could be included. Once the “going” was found to be non

significant, AWT races were included in the study and the “going” excluded.

The type of race was simplified as “flat” (flat, NH flat and stakes) or “jump”

(hurdles and chase), and stakes was also isolated as a factor.

 In total, 4410 races took place during that period, involving the participation

of 47847 horses. From these, races where the favorite horses had not finished or started

were discounted. Also, races where favorites shared placements were not considered to

avoid eventual bias.

We analyzed 3604 racing starts in the five month period. Of these 2918 (81%)

were flat group races (flat, NH flat and stakes) and 686 (19%) were NH group races

(hurdles and chase). Of the flat group races 542 (19%) were run on AWT and 2376

(81%) were run on turf. All 686 NH races were run on turf, making a total of 3062 runs

on turf and 542 on AWT. NH flat or bumpers races accounted for less than 2% of the

total races and therefore were included as flat races, as it did not seem appropriate to

include them in the jump category. The favorite won 1188 races which equates to 33%,

was placed in 1147 races or 32% (overall placement 65%), and was not placed in 1269

(35%). The second favorite won 713 races or 20%, was placed in 1281 races which

equals 36% (overall placement 55%), and was not placed in 1610 races (45%). The

results are summarized in table 1.

Table 1- Number of races taking place and analyzed in the period of analysis, and
placements of first and second favorite horses
Total Number of Races 4410
No of Races Considered 3604
1st favorite 2nd favorite
Number % Number %
Overall placement 2335 65 1994 55
Placed won 1188 33 713 20
Placed not won 1147 32 1281 36
Not placed 1269 35 1610 45
 A backwards stepwise approach was used to fit the model and several non

significant (p>0.05) variables were excluded (type of race, stakes, and track surface). In

the end, two factors (period of the day and handicapping) and the covariate (number of

runners) were found to be significant (p<0.001, p<0.001 and p<0.05 respectively). To

finalize, the interactions between the factors were added to the model but were found to

be non significant (p>0.05). The reduced adjusted model parameters are summarized in

table 2.

Table 2 - Logistic regression model for predicting the placement or not, of favorite
horses. Day period has a positive coefficient for evening races, and these have higher
odds for favorite horses to be place; Handicap has a negative coefficient for
handicapped races, and these have lower odds for favorite horses to be place;
Runners has a negative coefficient, and the number of runners correlate negatively
with the odds for favorite horses to be placed.
Variables in SE (β) p-value 95% CI (β) OR (eβ) 95% CI OR (eβ)
β
the equation
Intercept 4.068 0.1937 <0.001 3.688 4.447 58.414 39.965 85.381

Runners -0.152 0.0141 <0.001 -0.180 -0.125 0.859 0.835 0.883

Handicap -0.951 0.1073 <0.001 -1.161 -0.740 0.386 0.313 0.477

Day period 0.231 0.1113 <0.05 0.013 0.449 1.260 1.013 1.567
SE: standard error; CI: confidence interval; OR: odds ratio. The adjusted final model has
a deviance of 274 and a Akaike's information criterion (AIC) of 638.

Analysis & Conclusion

The results from the current study show that there are certain situations in

thoroughbred racing where the winning result anticipated by the allocation of a favorite

status is more or less likely to be achieved. In fact, 3 of the variables investigated

(number of runners, handicap and period of the day) appear to help predict whether or not

one of the favorites will end up placing.

Several authors refer to the logic of the odds of a favorite placing decreasing

with the number of horses running. Fewer runners mean less competition and less

potential for interference. The importance of interference is supported by G.S. Martin and

coauthors in 1996 in the “Journal of the American Veterinary Association”) who noted

that the winning time tends to increase by about 0.23 seconds for each additional runner.

Another logical result is that the favorites are less likely to place in handicap

races than in non-handicap races. In handicapped races, more weight is added to better

performers, in an effort to give all of the runners about the same chance of winning. In

fact, handicapping increases the competitive capacity of less skilled pairs (horse/rider) by

the addition of weights to be carried by the favorites, to equalize winning odds. The

favorite is, thus, less likely to win in handicapped races, in comparison to non

handicapped races.

Perhaps the most interesting result is that the period of the day appeared to be

an important factor. After controlling for the number of runners and type of the race,

favored horses were more likely to place in evening races than in daytime races. This has

not been mentioned before in the literature, and we do not have an intuitive explanation

for this effect, and therefore this is a subject in need of further investigation.

Figure 1 displays the estimated probability of the number of runners, for each

of the four combinations of the two factors (day period and handicapping). The results

are also summarized in table 3. The probability of the favorite placing decreases with the

number of runners, with the odds ratio decreasing by about 14.1% on average for each
additional horse. The decline is faster in handicapped races than in non-handicapped

races. And for both handicapped and non-handicapped races, the decline is sharper for

day races than for evening races.

Figure 1 – Logistic curves graphing the probability of placement for a favorite

horse, in dependency of the number of runners, period of the day and handicapping.
Cut off point refers to the point where the probability equals 0.5. In the original
data races had between 3 and 29 runners with an average of 11.

1
0.9
0.8
0.7 evening, no handicap; cut
Probability

0.6 off point:28 - 29

0.5 day, no handicap; cut off
0.4 point: 26 - 27
0.3 evening, handicapped;
cut off point: 22 - 23
0.2
0.1 day, handicapped; cut off
point: 20 - 21
0
11
16

26
31
36
41
21

46
1
6

Number of runners

Table 3 - Numbers of runners corresponding to the given probabilities of placement of

a favorite horse, based on the model fitted.

Number of runners
Probability Day Evening Day Evening
Handicapped Handicapped No Handicap No Handicap
0.9 6 7 12 13

0.8 11 12 17 19

0.7 14 16 21 22

0.6 17 19 24 25

0.5 20 22 26 28
 Despite trying to analyze a wide number of variables that may affect

performance, there were still many influences not measured and therefore not

incorporated in the model, which may affect the results. Firstly the race track itself is an

uncontrolled environment and performance studies are difficult to accomplish because of

the hidden influences that arise. Other variables that may impact upon performance

include rider influences such as weight, riding style and use of whip, poor start, wide

positioning through a turn and positioning behind a slower horse in the straight. The

horse itself is also a never ending source of possible variables to be considered in

performance.

The uncountable number of different variables to be taken into consideration

in the discussion of horse performance in races is so large that makes it difficult to isolate

them for a convenient analysis. Also the eventual interaction between these would

obviously result in a biased conclusion, if not considered. Therefore it seems logical to

observe that the non inclusion of hidden variables in horse races outcome prediction

models leads to an increase in unpredictability and obviously to a decrease in the correct

prediction of wining places.

As a final thought, racing possesses a strong element of random chance which

makes it impossible to predict every winner with perfect accuracy. Even so, if you have a

desire to gamble and want to take the chance of betting on a horse, place your money on

a favorite, not handicapped and running in an evening race with a small number of

runners. The question now is: do the bookmakers have this information? If so, the payoff

will reflect this, otherwise it won’t, and you have here an opportunity to make some easy
money. It’s up to you to find out and decide where to put your money, but please don’t

blame us if things go wrong…!

Further Reading

Benter,W. (1994). Computer based horse race handicapping and wagering systems: A

report. In: Hausch, D.B., Lo, V.S.Y., Ziemba, W.T. (Eds.), Efficiency of Racetrack

Betting Markets. Academic Press, London

Brown, L.D.; D’Amato, R.; and Gertner, R. (1994). “Racetrack betting: Do bettors

understand the odds?”. CHANCE 7:17-28.

Duncan, D. (2005). Winning Horse Racing Formulae. Foulsham and Co. Ltd,

Chippenham, UK

Hutson, G. D.; and Haskell M. J. (1997). “Pre-race behavior of horses as a predictor of

race finishing order”. Applied Animal Behavior Science 53(4):231-248.

Martin, G.S.; Strand, E.; and Kearney, M.T. (1996). “Use of statistical models to evaluate

racing performance in Thoroughbreds”. Journal of the American Veterinary Association

209(11):1900-1906.

Mota, M.D.S.; Abrahão, A.R.; and Oliveira H.N. (1998). “Genetic and environmental

parameters for racing time at different distances in Brazilian Thoroughbreds”. Journal of

Animal Breeding Genetics. 122(6):393-399.

 Newton, J.R.; Rogers, K.; Marlin, D.J.; Woods, J.L.N.; and Williams, R.B. (2005). “Risk

Factors for epistaxis on British race courses: evidence for locomotory impact-induced

trauma contributing to the aeteology of exercise-induced pulmonary hemorrhage”.

Equine Veterinary Journal. 37(5):402-411.

Rowe, R.V. (2004) How to Win at Horse Racing. New York: Cardoza Publishing,

Smith, M.A.; and Williams, L.V. (2010). “Forecasting horse racing outcomes: new

evidence on odds bias in UK betting markets”. International Journal of Forecasting.

26(3):543-550.

White, E.M.; Dattero, R.; and Flores, B. (1992). “Combining vector forecasts to predict

Thoroughbred horseracing outcomes”. International Journal of Forecasting. 8(4):595-

611.

Young, L.E.; Rogers, K.; and Wood, J.L.N. (2005). “Left ventricular size and systolic

function in Thoroughbred racehorses and their relationship to race performance”. Journal

of Applied Physiology. 99(4):1278-1285.

View publication stats