The Application of Ranking Probability Models to Racetrack Betting

Author(s): Victor S. Y. Lo, John Bacon-Shone and Kelly Busche

Source: Management Science, Vol. 41, No. 6 (Jun., 1995), pp. 1048-1059
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/2632831
Accessed: 29-07-2016 19:39 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 The Application of Ranking Probability
Models to Racetrack Betting

Victor S. Y. Lo * John Bacon-Shone * Kelly Busche

Mercer Management Consulting, Inc., 33 Hayden Avenue, Lexington Massachusetts 02173
Department of Statistics and Social Sciences, Research Centre, The University of Hong Kong,
Pokfulam Road, Hong Kong
School of Economics, The University of Hong Kong, Pokfulam Road, Hong Kong

H ausch et al. (HZR) (1981) developed a betting system that demonstrated positive profits
at two racetracks. The system assumes running times are distributed exponentially, but
other distributions for running times (Henery 1981 and Stem 1990) have been shown to produce
a better fit in Bacon-Shone et al. (1992a), Lo (1994), and Lo and Bacon-Shone (1994) using
data from Hong Kong, the Meadowlands, and Japan. The better fit is at the cost of severely
increased complexity in computing ranking probabilities, though. In response, Lo and Bacon-
Shone (1992) proposed a simple model of computing ranking probabilities which closely ap-
proximates those based on the Henery and the Stern models and fits the data as well. This
paper couples the Lo and Bacon-Shone model and the HZR system. For data sets from the
United States and Hong Kong, we show improved profit over the HZR system at lower levels
of risk using final betting data assuming zero computational costs. With data from Japan, our
model shows little difference in profits from the HZR system.
(Racetrack Betting; Running Time Distributions; Betting Systems)

1. Introduction where ri P (horse i wins). This is equivalent to saying

A betting system developed by Hausch et al. (1981) that given that i wins, the probability that j finishes
(the HZR system1) searches for possible profits in ra- second is the same as the probability that j wins in an
cetrack place and show2 betting pools. The HZR system n - 1 horse race with horse i removed. Similarly,
approximates winning probabilities by public's win bet
7Xijk P (horse i wins,
fractions (proportions of money bet on different horses
in the win pool) and uses the Harville (1977) model to j finishes second and k finishes third)
predict ranking probabilities. For instance,
7ri j 7rk

= ij P(horse i wins and horse j finishes second) (1-7ri(1 - i-7rj) (2)

7ri (1) Ranking probabilities together with the track's for-

1 - mulas for place and show payoffs allow an expected
return to be calculated on each possible place and show
bet. For those that are sufficiently positive, HZR deter-
' Also called the Dr. Z system where it is more fully developed for
mines optimal wagers.
practitioners in Ziemba and Hausch (1984, 1987). See also Hausch
and Ziemba (1985) for full development and tests, and Skinner's The main advantage of (1) and (2) is that we need
(1989) discussion on the development of this system. the win probabilities of only two or three horses. The
2 We use U.S. terminology: a place bet pays out if the horse is first or win probability ri can be estimated by the win bet frac-
second; a show bet pays out if the horse is first, second, or third. tion; see Ali (1977), Snyder (1978), Asch et al. (1982,

0025-1909/95/4106/1048$01.25
Copyright ? 1995, Institute for Operations Research
1048 MANAGEMENT SCIENCE/VO1. 41, No. 6, June 1995 and the Management Sciences

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

1984, 1986), Ziemba and Hausch (1986), Busche and Stern (1990) assumed independent gamma distri-
Hall (1988), and Bacon-Shone et al. (1992a) for dis- butions with a fixed shape parameter r for the running
cussions and analyses on using the win bet fractions. times. The model is motivated by considering a com-
Hausch et al. (1994) provide a thorough survey of var- petition in which n players, scoring points according to
ious racing efficient market research. independent Poisson processes, are ranked according
The Harville model is implied by assuming that run- to the time they take to score r points. Thus r should
ning times are distributed as independent exponential be an integer in this scenario. The Stern model gener-
random variables with unique parameters for each horse alizes the Harville and Henery models. Suppose Ti
in each race (Dansie 1983). An alternative to the Har- Gamma(r, 6j) independently for some r and 6s. Then,
ville model was proposed by Henery (1981), who as- r = 1 is Harville's exponential model. The Stern model
sumes independent normal distributions for the running also requires a computationally intensive procedure of
times. The Henery model is a similar order to the Henery model for evaluating the
ranking probabilities.
7ri = P(Ti < Tj < min{Tr}) Bacon-Shone et al. (1992b) compared the Harville,
r$i,j
Henery, and Stern (r = 2, 4) models using data from
= 4??(u + -4 ) various betting pools in Hong Kong and Meadowlands
oo
(New Jersey) racetracks. The Henery model predicted
(by likelihood measure and Cox's (1962) test) the rel-
X fI [1 - 4(u + Oj- 0010(u)du,
r$i,j evant ranking probabilities better than the Harville
model or Stern (r = 2, 4). Analysis of the Stern model
and
with higher values of r (10, 20, 30, and 40) in Lo (1994)
7ijk = P(Ti < Tj < Tk < min {Tr}) confirmed the superiority of the Henery model in
r#ki,j,k
Meadowlands and Hong Kong. However, Lo and
Bacon-Shone (1992) showed that the Stern model with
-oo~ [I -ookO(v + (v - +i)o(v)dv
a particular shape parameter fits the Japan data better
than the Henery model.
X ]7 [1 - (D(U + Ok- Or)]O(u)du, In ?2, we consider a simple model for predicting the
r$ijk
ranking probabilities. The HZR system is briefly de-
where Ti = running time of horse i, which is distributed scribed in ?3. We apply our model for predicting the
as N(0i, 1 ), and ?( * ) and 0( - ) are the standard normal ranking probabilities to the HZR system in ?4. Conclu-
cumulative density function and probability density sions are in ?5.
function, respectively. The 0i's are obtained by solving
the equations:

2. A Simple Approximation of the

7ri= P(Ti < min{Tr})
r$i
Henery and the Stern Model
for i = 1, . . ., n (number of horses) Lo and Bacon-Shone (1992) proposed a class of simple
models (hereafter called the discount model) which ap-
proximates closely both the Henery and the Stern mod-
= fi rD(u + Oi - Or)o(u)du,
-? r*i els. The original idea behind this model was to provide
a simple extension of the Harville model. Define the
where the iri's are assumed to be known or estimated Log Odds ratio
by the win bet fractions. The procedure involves two
steps: (1) numerically solve a system of nonlinear LO(i, j I k) LogOdds (horse i beats horse j
equations for the Oi's; and (2) compute the ranking
for kth place I horses i, j do not place
probabilities using the estimated 0i's by numerical in-
tegration. in top k - 1 places).

MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995 1049

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

For the Harville model, this has the simple form Table 1 Parameter Values for the
Discount Model
Log(-ri) - Log( rj), which is independent of k. A rea-
sonable extension is for the log odds ratio to shrink
r Xr
r~~~~~~~~~~~~~~~N A3
toward zero as k increases, on the rationale that the
horse's ability would matter less (be discounted more) 1 (Harville) 1.00 1.00
as the prize money diminishes to zero. We assume this 2 0.93 0.89
can be modelled as 3 0.90 0.84
4 0.88 0.81
LO(i, j I k) = XkLO(i, j I 1), 5 0.87 0.80
6 0.86 0.78
where Xk is a decreasing function of k. In fact, if we
7 0.86 0.77
calculate the logodds ratio under the assumption of
8 0.85 0.76
Stern's model with parameter r, 10 0.84 0.75
20 0.82 0.72
LO(i,1j2)/LO(i,Ij1) and LO(i,jl3)/LO(i, jI1) 30 0.81 0.71
40 0.81 0.70
are empirically very close to their limiting values
oo (Henery) 0.76 0.62
achieved as the mean running times tend to be equal.3
Approximating the above functions by their limiting
values, denoted by X and Xr (dependent on r), re-
spectively, and rearranging, we obtain the discount suggested values of X and X' for different r appear in
model for a relation between the ranking probability Table 1 below. When r = 1, the Stern model reduces to

7rijk and the simple winning probability 7ri: the Harville model. The discount model with X = X
= 1 is also the Harville model. To include both the
Ar r kA
Stern and Henery models, we define that when r = oo,
rijk = i r 7 toij rXA (3) the discount model is an approximate form of the Hen-
ery model.5
where indicates an estimate, and 7ri is estimated by
the win betfraction of horse i. Under the discount Table 1 also shows that the values of X and X are

model, the ranking probability is a function of win odds decreasing in r; as r increases, the Stern model deviates

data on all horses, whereas the Harville model (2) only farther away from the Harville model (i.e., r = 1). The

requires win odds on three horses. decreasing behavior of X and X implies that if the data

Using various data sets, Lo and Bacon-Shone (1992) actually follow the Stern model with parameter r > 1,
the Harville model will systematically overestimate the
showed that the approximated log likelihoods based on
ranking probabilities of finishing second and third for
(3) are very close to the true log likelihoods based on
the Stern model for the events of the horses finishing favorites and underestimate those for longshots (see Lo
1994), and as the true value of r increases, the systematic
in the first three positions. A fixed race size, n = 11, is
bias caused by the Harville model becomes more serious.
adequate4 for the approximation since the limiting val-
To determine an appropriate value of r, we find the
ues do not vary much with the race size. Hence, the
maximum likelihood estimate for r given the win bet
discount model can be considered as an approximation
to the Stern class model with shape parameter r. The
fractions and the finishing order in a set of m races. The
log likelihood for the events of the horses finishing in
the top three positions is
3 The mathematical forms of the limiting values depend on the race
size and r only. They appeared in Lo and Bacon-Shone (1992).
logrj' rI rI yi,s l log lr[123],s, (4)
4 Lo and Bacon-Shone (1992) showed that the limiting values are s=1 i,j,k s=l

quite robust to the race size. The average race sizes in the Hong Kong,
Meadowlands, and Japan data are 10.0, 9.5, and 11.3, respectively.
Their log likelihood measures showed negligible differences between But the true Stem model itself does not necessarily tend to the Henery
parameters with n = 11 and race-size dependent parameters. model when r -- oo .

1050 MANAGEMENT SCIENCE/VOL 41, No. 6, June 1995

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

where s denotes the race number and yijk,S is the indicator ficiently high, the system recommends a bet, with the
variable that horse i wins, j finishes second, and k fin- actual amount to bet described below. The HZR system
ishes third in race s. The second product on the left- uses a rule of thumb to restrict bets to those with an
hand side denotes the multiplication over all combi- estimated expected return per dollar above 1.16 or above
nations of horses i, j, and k, and [123] on the right- 1.20 (HZR 1981, p. 1444).8 If the win pool is nearly
hand side denotes the horses finishing first, second, and efficient and the Harville model is accurate, this selection
third. procedure produces profit from (sufficiently large) mar-
Using this likelihood approach, Lo and Bacon-Shone ket inefficiencies that exist in place and show pools.
(1992) and Lo (1994) concluded that the Henery model
3.2. Optimization of Bet Amounts
fits the data best (of Harville, Henery, and Stern
Bets in the HZR system are made according to the
(r = 2, 4, 10, 20, 30, 40)) in both the Meadowlands
"Kelly" criterion, which maximizes the rate of growth
and Hong Kong. However, in Japan, the Stern model
of capital. (For details, see Kelly 1956; Breiman 1960,
with shape parameter r = 4 fits the data best.
1961; Thorp 1971; Hakansson 1971; Grauer 1981; Al-
The models of Stern and Henery require solving sys-
goet and Cover 1988; and McLean et al. 1992.) This
tems of nonlinear equations and numerical integrations,
entails maximizing the expected logarithm of final
which is computationally intensive.6 In contrast, the
wealth subject to a budget constraint. The effect of ad-
discount model requires no integration nor the solution
ditional bet amounts on the odds is also accounted in
of a system of equations for individual races.
the system.
The data to be utilized are ideally obtained as late as
3. The HZR System possible in the betting period, but in practice some time
Assuming the win market produces accurate estimates, has to be reserved for data input, computations, and
Hausch et al. (HZR) (1981) identified horses underbet wagering. HZR (1981) gave evidence that profits can
in the place and show markets. Using data from Santa be made by using odds from two minutes before the
Anita (Los Angeles) and Exhibition Park (Vancouver), start. We cannot address this question here, though, as
profits of the order of 10% were reported. Additional our data consist only of final odds. Since publications
evidence of profits can be found in Ziemba and Hausch of Ziemba and Hausch (1984, 1986, 1987) and Skinner
(1984, 1987). The details of the HZR system are de- (1989), the HZR system has been widely used, so it
scribed below. may suffer to some extent from overbetting.9"10

3.1. Selection of Horses

' Data in Ziemba and Hausch (1987, Chapter 9), using cutoffs of
The HZR system replaces the winning probabilities by
1.10, 1.14, and 1.20 for various quality races, showed actual profits
the bet fractions and then uses the Harville model7 of 11.2%. The extra percentage edge was suggested to account for
(Equations (1) and (2)) to estimate the ranking prob- errors in approximation including the breakdown of the Harville for-

abilities, rij and 11ijk. When these probabilities indicate mulas for probabilities close to 1.

that the expected return on a place or show bet is suf- 9 HZR suggested an approximation procedure using a set of linear
regressions which we have not used in this paper. This approximation
is used in their handheld calculator, which only takes about 20 seconds
6 Based on 50 races of 14 horses in Hong Kong, the average time for a race to input data and compute the expected return and optimal
required on an IBM 386 machine for solving the system of nonlinear Kelly bet. The other point is that the discount model requires all win
equations associated with the Henery model is 10.2 minutes, with a odds while the Harville model requires input from only those com-
range of 1.1 to 36.5 minutes. Computing the ranking probabilities for binations that have been determined to be possible bets. It may be
all the combinations for three horses in each race takes more than possible that special devices, such as a scanner, would permit a practical
one hour, method of data collection to be devised for the discount model.
'Harville (1973) and HZR (1981) showed that the Harville model 10 Market players betting too much may drive away the inefficiencies.
appeared to overestimate (underestimate) the chances of a second or There are calculations in Hausch and Ziemba (1985) and Ziemba and
third place finish for favorites (longshot). They also showed that this Hausch (1987) concerning how many can play. An example from
effect tended to be cancelled by the favorite-longshot bias of win odds Kentucky Derby in Hausch and Ziemba (1985) showed that 1,316
when computing the probabilities of placing and showing. system bettors were needed to reduce the expected return per dollar

MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995 1051

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

4. Applying the Discount ModeI to receives back his original bet. Then the remaining
amount in the show pool, after the track take, is split
the HZR System
evenly among the show bettors on i, j and k. The payoff
We examine the HZR system with data from four race-
to horse i to show is independent of whether i finishes
tracks: the Meadowlands and Atlantic City in New Jer-
first, second, or third, but it is dependent on which
sey; Hong Kong; and Japan. The first step is to select
horses finish with it. If i is a longshot (i.e., smaller value
the horses to be bet on. The estimated expected return
of Si), the bettor will receive a higher amount.
from a $1 bet to show on horse i is
In the Japanese return formula, if horse i finishes at

E(Xq) = (7rijk + lIjik + 7'jki) Retijk, least third, the return on a show bet on horse i per yen
j #k,i
k #i
is (ignoring the additional amount bet):

where 7Frjk estimated probability that horse i wins, 0.246(S - S[11 - S[21 - S[31)
0.838 +
horse j finishes second, and horse k finishes third (see Si

Equation (3)), and Retijk =show return on horse i per (0.738S + 0.262 2T3=1 S[r] - 0.486Si ) - r=1 S[r]
dollar if horses i, j, and k finish in the first three posi- =1 + 3 Si
tions. For the three racetracks,
(1 - ti)S - r S[r]
Retijk = 1 + (- t)(S + 1) - (1 + Si + Sj + Sk) = 1 + 3~~3Si
Ret~~1k=1+ 3(1 +Si)
where ti = the implied track take, S[r] = show bet on
in the U.S. and Hong Kong; and
the horse which finishes in the rth order, and S = sum
of show bets for all horses. The last expression is the
0.838 + 0.246 (S - Si - SI - Sk
1 + Si traditional return formula if ti is a constant for all i.
Here, ti is not a constant, and
in Japan, where Sj total dollar amount bet to show
on horse j, S Ej Sj, and t track take. 0.262 Er S[rj - 0.486Si
1-ts=0.738+ S X (5)
The track take" is 17% and 18% in Hong Kong and
the Meadowlands (U.S.), respectively. The return for-
i.e., the implied track take (and thus the return) of horse
mula for Japan is different from the other" two, and its
i depends on the show bet of that horse relative to the
implied track take is not a constant. Formulas for place
other winning horses.
bets, E (X '), are similar.
To illustrate the effect of using the Japanese return
The return formulas for the U.S. and Hong Kong are
formula, consider a numerical example. Suppose the
now discussed. If horses i, j, and k finish in the top
proportion of show bets on the first three horses,
three positions, each bettor who bets on i to show first
r=1 S[r]/S = 0.6 and also S[1]/S = 0.5, S[2]/S = 0.01.
Then the implied track take for the favorite [1] is 0.348
with return per dollar bet = 1.03, and the track take for
bet to 1.02, assuming a betting wealth of $1000. It is not known how
the longshot [2] is 0.110 with return per dollar bet
many people actually play this system at major racetracks, but at
= 10.68. If track take is a constant (1 - 0.738 = 0.262),
many of them the number is not small. Hence, profits that are attain-
able in 1993 may be substantially less than in the simulations in HZR then the returns per dollar bet for [1] and [2] are 1.09
(1981) and Ziemba and Hausch (1987). and 5.60, respectively. Thus the return strongly depends
" In addition to the track take, another cost of betting is "Breakage," on whether the horse is favorite or longshot.
which is simply due to rounding or truncation of the returns. In Hong The HZR system indicates a show or place bet on
Kong, returns are rounded to the nearest cents and in the U.S. and
horse i if E(Xq) or E(XV) > a. The cutoff value a is
Japan, returns are truncated to the lower 104 and Y10, respectively.
chosen "subjectively." In HZR (1981, p. 1444), "In-
We have taken this into account in calculating the returns per dollar.
Calculations in Hausch and Ziemba (1985) showed that this is a large
tuition suggests that Santa Anita with its larger betting
value. In particular, their analysis showed that profit dropped by 24.1% pools would have more accurate estimates of the qi
with breakage, using the HZR system. [winning probabilities] than would be obtained at Ex-

1052 MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

hibition Park. Hence positive profits would result from Table 2 Approximate Log Likelihoods for Different Data Sets

lower values of a." HZR use a = 1.20 for Exhibition

Hong Kong Meadowlands Atlantic City Japan
Park and a = 1.16 for Santa Anita.'2 Pool sizes per race
1981-82 1984 1978 1990
in Hong Kong are at least an order of magnitude greater
r (905 races) (705 races) (712 races) (600 races)
than those in the U.S. We follow HZR and use 1.16 for
Hong Kong, and also try a = 1.08, 1.12, and 1.20. We 1 (Harville) -5361.2 -4006.9 -3679.7 -3558.2

vary the HZR system in two ways: 2 -5337.9 -3977.4 -3663.3 -3549.2
3 -5329.7 -3966.1 -3657.4 -3548.1
(1) In addition to HZR's selection rule that expected
4 -5325.8 -3960.3 -3654.4 -3548.0<
return 2 a, we use probability restrictions of the form
5 -5323.5 -3956.7 -3652.7 -3648.2
that P (achieving a profit) 2 f (also arbitrarily chosen) 6 -5322.0 -3954.3 -3651.5 -3648.5
to screen low probability horses. The two restrictions 7 -5320.9 -3952.5 -3650.7 -3648.8
considered are: 8 -5320.2 -3951.2 -3650.1 -3549.1

a. P (i finishes 1st or 2nd) 2 0.10 for place bets, 10 -5319.2 -3949.4 -3749.3 -3549.7
20 -5317.1 -3945.2 -3647.5 -3551.4
and P (i finishes 1st, 2nd or 3rd) 2 0.20 for show bets,
30 -5316.4 -3943.6 -3646.9 -3552.3
b. P (i finishes 1st or 2nd) 2 0.20 for place bets,
40 -5316.0 -3942.7 -3646.6 -3553.9
and P (i finishes 1st, 2nd or 3rd) 2 0.30 for show bets. oo (Henery) -5315.2< -3935.4<< -3644.9<< -3562.0
Note that b is more restrictive than a. A similar idea
was recommended by Ziemba and Hausch (1986, N.B. The value marked with < is the maximum log likelihood among the
others in the same column.
1987), who suggested screening out those horses with
win odds larger than 8-1.
(2) To estimate ranking probabilities 7rijk, we use 4, 5, or 6 are small. In all cases, the change in the ap-
the discount model with an appropriate r which is op- proximate log likelihood is largest when we move from
timal for the particular racetrack. r = 1 to r = 2.
Our data for comparing the discount model and the The betting results for Hong Kong (2291 races), the
Harville model using the HZR system are Hong Kong Meadowlands (705 races), and Japan (983 races) are
1985-1990 (2291 races), the Meadowlands 1984 (705 shown in Tables 3 through 5. The corresponding graphs
races), Atlantic City 1978 (712 races), and Japan 1990- for no probability restriction (the column under "orig-
1991 (1583 races). To avoid using the same data sets inal" in the tables) appear in Figures 1 through 3.13.14
for both estimation of r and analysis of betting strategy, Following HZR (1981), the initial wealth in all cases is
we use out-of-sample Hong Kong data from 1981-1982 U.S. $10,000. Absolute risk aversion (see HZR 1981, p.
(905 races) for estimation in Hong Kong and split the 1445) for the Kelly criterion is 1 /(initial wealth), or
Japanese data into two parts (the first 600 races are almost zero. From the likelihood measures we choose
chosen for estimation). For the U.S. data, the parameter the discount model parameter, r = oo ((Xr, X3) = (0.76,
r estimated in Meadowlands is used with our second 0.62)) in Hong Kong and Meadowlands and r = 4
U.S. data set from Atlantic City. The approximate log
likelihoods, based on the discount model (see Equations
13 In Figures 1 through 3, a = 1.12 is used for Hong Kong and Japan
(3) and (4) and Table 1 in ?2) for the events of the
data, and a = 1.16 is used for Meadowlands data. This is consistent
horses finishing in the top three positions, are reported with a recommendation in Ziemba and Hausch (1987) that higher
in Table 2. From Table 2, the approximate Henery model cutoff value is more appropriate for races with smaller pools.
appears to fit the Hong Kong, Meadowlands, and At- 14 Following HZR ( 1981), the straight lines in the graphs are approx-
lantic City data best. For the Japanese data, r = 4 appears imate wealth histories for random horse betting, where the total bet

to be the best, although the differences among r = 3, amount is the same as that of the Harville model, e.g., total dollar bet
using the Harville model in Figure 4 (the Meadowlands) is $161,525
and track take is 18%. Therefore the final wealth level is 10,000
12 Ziemba and Hausch (1987) recommended 1.10 for the best races - 0.18(161,525) = -$19,074.5 (which should actually be zero in
in the U.S., 1.14 for races at top tracks, and 1.18 for races at other practice). For simplicity, the track take used in Japan is assumed to
tracks. be fixed at 26.2% for the random betting line (see equation (5)).

MANAGEMENT SCIENCE/VOL 41, No. 6, June 1995 1053

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

Table 3 The HZR System and Its Variants in Hong Kong 1985-1989

Harville Discount (r = oo)

Prob. Prob. Prob. Prob.

Original Restr. a Restr. b Original Restr. a Restr. b

us$ us$ us$ us$ us$ us$

a = 1.08

F.W. 0.6 0.7 0.6 41 152 871

T.B. 44626 44848 44514 68660 68097 70657
N.W. 241 240 235 547 318 203
M.C. 0.6 0.7 0.6 29 105 505
%FAV 58 59 62 10 23 37

a = 1.12

F.W. 5 5 5 55 179 806

T.B. 31172 31184 30947 41216 37110 36859
N.W. 141 140 140 344 172 145
M.C. 2 5 5 38 121 292
%FAV 61 62 66 8 20 45

a = 1.16

F.W. 2 35 26 674 1222 3332t

T. B. 34926 35063 33297 55010 38757 34446
N.W. 142 174 155 227 90 50
M.C. 2 4 3 440 651 1672
%FAV 66 68 71 8 23 42

a = 1.20

F.W. 55t 53t 62t 944t 1601t 2569

T.B. 22270 22139 21210 28627 23281 17449
N.W. 124 120 115 143 50 23
M.C. 14 13 16 69 879 1574
%FAV 60 62 65 5 16 32

Notes:

(i) In Table 3, HK$ have been converted at US$ 1 = HK$ 7.8. Similarly, in Table 5, Japanese yens have been
converted at US$ 1 = Y 140.
(ii) The initial wealth for all three data sets: US$ 10,000.
(iii) F.W. = Final Wealth; T.B. = Total amount bet; N.W. = Number of races that have wagers (it does not
count those races with optimal bet < minimum allowed bet); M.C. = Minimum Capital during the betting history;
%FAV = Percentage of most favorite horses chosen by the first step of the HZR system.
(iv) t denotes the row in each column with the highest final wealth.
(v) In Hong Kong, only US style place bets are available for 4-6 horse races, and for races with more horses,
US style show bets are available. In Japan, only place bets are available for 4-7 horse races, and for races with
more horses, show bets are available.
(vi) For the three cases of a = 1.08 with the Harville model in Hong Kong (Table 2), we ended up with wealth
less than the minimum bet amount, hence the betting processes stopped before reaching the end of the whole
period.

1054 MANAGEMENT SCIENCE/VOl. 41, NO. 6, June 1995

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

Table 4 The HZR System and Its Variants in Meadowlands 1984

Harville Discount (r = oo)

Prob. Prob. Prob. Prob.

Original Restr. a Restr. b Original Restr. a Restr. b

us$ us$ us$ us$ us$ us$

a = 1.08

F.W. 7 10 3 14517 15629 14000

T.B. 144746 146024 132089 94431 93402 89873
N.W. 472 471 456 333 301 271
M.C. 4 6 2 7380 7507 7630
%FAV 58 59 60 32 37 44

a = 1.12

F.W. 2007 2076 1358 17388t 18470t 16888t

T.B. 215251 215887 210843 62762 61862 59479
N.W. 354 351 344 242 208 183
M.C. 1268 904 1321 7182 7170 7537
%FAV 57 58 60 29 35 41

a= 1.16

F.W. 4296 4448 3378 15631 16449 14539

T.B. 161525 161733 158735 36466 35555 33763
N.W. 253 248 244 163 132 114
M.C. 3895 4051 3008 8371 8464 8388
%FAV 54 55 56 25 31 36

a = 1.20

F.W. 6397t 6569t 9921 t 14273 15029 13871

T.B. 128097 101888 101348 21153 21153 19980
N.W. 174 170 159 115 88 77
M.C. 3471 3519 4829 8218 8218 8330
%FAV 51 53 52 18 24 29

((X2, X3) = (0.88, 0.81)) in Japan. In addition to final or Meadowlands.'6 The discount model with r = ox
wealth, we also show total amount bet, number of wa- produces positive profits in Meadowlands, while in
gers, minimum capital during the wealth history, and Hong Kong, losses are smaller than those from the Har-
percentage of most favorite horses chosen by the system ville model. Moreover, the percentages of most favorite
in the tables. Intuitively, minimum capital is associated horses chosen by the Harville model are larger than
with the risk involved in the betting method. those of the discount model in all cases in Hong Kong
Tables 3-5 show that there are no profits based on
the HZR system with the Harville model in Hong Kong 15 16 For the Meadowlands results in Figure 2, the pattern of decreasing
wealth history of the Harville model started when race sequence ? 650.
15 The departmental editor suggested that the HZR system will not This was likely because more favorites are selected during this period:
work in Hong Kong because the favorite-longshot bias does not exist 71% of the selected horses are favorites for the Harville model in
there (Busche and Hall 1988 and Busche 1994). these races (but only 54% are favorites for the whole history).

MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995 1055

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

Table 5 The HZR System and Its Variants in Japan 1990-1991

Harville Discount (r = 4)

Prob. Prob. Prob. Prob.

Original Restr. a Restr. b Original Restr. a Restr. b

us$ us$ us$ us$ us$ us$

a = 1.08

F.W. 4860 5935 5901 8922t 11072t 11281t

T.B. 35539 35562 34865 15075 11914 10973
N.W. 53 37 30 61 18 13
M.C. 3403 3638 3850 8217 8970 8970
%FAV 33 37 40 28 42 54

a = 1.12

F.W. 8332 9453 9524 8201 10517 10615

T.B. 13174 12221 11434 9569 7069 6176
N.W. 30 17 13 46 11 7
M.C. 7057 7729 8081 7943 8970 8970
%FAV 33 44 62 30 42 71

a = 1.16

F.W. 9503t 101 66t 10671 t 7592 9371 9867

T.B. 9059 8479 8040 4976 3101 2636
N.W. 18 8 6 30 6 4
M.C. 8543 8641 8641 7592 8721 8970
%FAV 32 50 67 24 33 50

a = 1.20

F.W. 8791 9231 9730 7212 8565 9048

T.B. 4275 3879 3421 3522 2185 1713
N.W. 13 6 4 22 5 3
M.C. 8084 8409 8641 6981 8115 8591
%FAV 29 33 50 17 20 33

and Meadowlands and in 75% of the cases in Japan. capital for larger a than is optimal for the discount
The reason is as follows. Since the Harville model tends model. Thus the inaccuracy of the Harville model ap-

to overestimate the ranking probabilities rij and Tijk pears to be partially compensated for by a higher a
where j and k are favorites, it overbets favorites to place level.17 However, the small number of wagers in Japan
or show. In Japan, the optimal discount model (r = 4) means that results for Japan are not conclusive.
is much closer to the Harville model (r = 1) than the
discount model (r = ox) we use in Hong Kong and
17 Since the Harville model overestimates the ranking probabilities
Meadowlands. Thus the overestimation of favorites in
for the favorites, some favorites will be incorrectly selected. Increasing
Japan using the Harville model is less important. the a level screens out some of these mistakes but when using the
In the Meadowlands and Japan, the Harville model discount model, increasing the a level only limits the number of wagers
performs best in terms of final wealth and minimum rather than removing some bad selections.

1056 MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

Figure 1 Wealth Level Histories in Hong Kong Figure 2 Wealth Level Histories in Meadowlands

Weal Lvel (U.S.$) Harville

46000- alpha=1.12 Wealt Lvel (U.S.$) Discount

Harville 25000 alpha=1.16 ... . Random

35000--Random
Discount

20000
2=~~~~~~~~ -

25000 15000 -

20000-
10000- V
15000

10000 - . .. .. ...... . ....-...-..

5000 ...

280 5b 70 ldO0 12t0 1560 150 20bo 2290 0

10o 20 360 40o 50O 60b *7 0
race sequence
race sequence

Rather than considering the final wealth only, an al-

5. Conclusion
ternative analysis on the betting results tests the differ-
Ziemba and Hausch ( 1987) report that the HZR system
ence in expected returns per race between the Harville
produced positive profits at a number of racetracks in
and discount models. Define the return for race s to be
the U.S. and Canada. The discount model presented in
log(W,/W,-), where W, = the wealth level just after
this paper offers a simple modification to the HZR sys-
race s.18 A common paired-sample t-test is performed
tem in order to estimate the ranking probabilities better,
to test the null hypothesis that the expected returns per
and produces higher returns with both Hong Kong and
race under the two systems are the same. The corre-
Meadowlands data sets. In Japan, the discount model
sponding t-statistics for the optimal alphas from Table
3-5 (the cases marked with t) appear in Table 6.1'
From Table 6, the expected return per race for the
discount model appears to be higher than that for the Figure 3 Wealth Level Histories in Japan

Harville model in Hong Kong and Meadowlands in Wealt Level (U.s.$)

many cases, although it is not significant in a few cases. alpha=1.12
Harville
However, the Japanese data do not show that the dis-
count model has significantly higher expected returns, Discount

although this conclusion is based on a small number of -000...---- Random

races. 9500 -

9000-

1 Like the definition of stock return, ln(W,/W,1-) (W, -Ws_l /Ws_l 8500-
We do not directly apply the test used in HZR (1981, p. 1450-1451),
because their underlying assumptions of equal win probabilities and 8000-

equal returns (7r and w in their paper) for each bet do not appear to
7500-
be realistic enough. However, the idea of our simple test is similar to
theirs. 7000
19 We use only those races which have a bet for at least one of the
systems for our test, as the other races provide no additional infor- 6500. 200 400 600 800
mation. race sequence

MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995 1057

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

Table 6 The t-statistics for the Difference in Expected Ali, M. M., "Probability and Utility Estimates for Racetrack Bettors,"
Returns per Race Using a which Produces J. Political Economy, 84 (1977), 803-815.
Highest Final Wealth Asch, P., B. G. Malkiel, and R. E. Quandt, "Racetrack Betting and
Informed Behavior," J. Financial Economics, 10 (1982), 187-194.
Prob. Prob. , ~~, and , "Market Efficiency in Racetrack Betting," J.
Original Restr. a Restr. b Business, 57, 2 (1984), 165-175.
,and , "Market Efficiency in Racetrack Betting: Further
Hong Kong 1985-89 Evidence and a Correction," J. Business, 59, 1 (1986), 157-160.
Bacon-Shone, J., V. S. Y. Lo, and K. Busche, "Modelling the Winning
No. obs. 233 141 126
Probability," Research Report 10, Department of Statistics, Uni-
t-stat 0.98 1.23 1.6400
versity of Hong Kong,
, and , "Logistic Analyses for Complicated Bets," Re-
Meadowlands 1984
search Report 11, Department of Statistics, University of Hong
No. obs. 259 226 201 Kong,

t-stat 1.480 1.540 0.65 Breiman, L., "Investment Policies for Expanding Businesses Optimal
in a Long-run Sense." Naval Res. Logistics Quarterly, 7 (1960),

Japan 1990-91 647-651.

, "Optimal Gambling Systems for Favorable Games," in Pro-

No. obs. 61 18 13 ceedings of the Fourth Berkeley Symposium, Mathematical Statistics
t-stat -0.48 0.96 0.66 and Probability, 1 (1961), 65-78.
Busche, K., "Efficient Market Results in an Asian Setting," in D. B.
Note: Hausch, V. S. Y. Lo, and W. T. Ziemba, (Eds.), Efficiency of Race-
(1) The t-statistic is associated with the null hypothesis track Betting Markets, Academic Press, 1994, 615-616.
that the expected return per race for the discount model minus and C. D. Hall, "An Exception to the Risk Preference Anomaly,"
that for the Harville model = 0. Since the differences do not J. Business, 61(1988), 337-346.
appear to be normally distributed, we assume asymptotic nor- Cox, D. R., "Further Results on Tests of Separate Families of Hy-
mality of the statistic in all the cases (which of course is not potheses," J. Royal Statistical Society B, 24, 2 (1962), 406-424.
valid for small samples such as the Japanese data). Dansie, B. R., "A Note on Permutation Probabilities," J. Royal Statistical
(2) No. obs. = number of races that have at least one bet Society B, 45, 1 (1983), 22-24.
from any of the two systems. Grauer, R. R., "A Comparison of Growth Optimal and Mean Variance
(3) DD (D) indicates that the null hypothesis rejected in Investment Policies," J. Financial and Quantitative Analysis, 16,
favor of the alternative that the discount model produces higher 1 (1981), 1-21.
returns at the 5% (10%) significance level. Hakansson, N. H., "Capital Growth and the Mean-variance Approach
to Portfolio Selection," J. Financial and Quantitative Analysis, 6
(1971), 517-557.
Harville, D. A., "Assigning Probabilities to the Outcomes of Multi-
provides some improvement for the cases of lower a entry Competitions," J. American Statistical Association, 68 (1973),
level only. In conclusion, the use of the discount model 312-316.

with appropriate r in the HZR system appears to im- Hausch, D. B., V. S. Y. Lo, and W. T. Ziemba (Eds.), Efficiency of

prove the estimation of ranking probabilities and avail- Racetrack Betting Markets, Academic Press (Mathematical Eco-
nomics and Econometrics Series), San Diego, CA, 1994.
able profit.20
and W. T. Ziemba, "Transactions Costs, Extent of Inefficiencies,
Entries and Multiple Wagers in a Racetrack Betting Model," Man-
20 We thank Richard Quandt for his U.S. data and Junji Shiba for his agement Sci., 31 (1985), 381-394.
help in acquiring the Japanese data. We also acknowledge the referees, , and M. Rubinstein, "Efficiency of the Market for Racetrack
the associate editor, and the departmental editor for their helpful Betting," Management Sci., 27 (1981), 1435-1452.
comments on earlier versions of this paper. The results in this paper Henery, R. J., "Permutation Probabilities as Models for Horse Races,"
were presented at the University of Hong Kong, University of British J. Royal Statistical Society B, 43, 1 (1981), 86-91.
Columbia, and the ORSA/TIMS conference in Phoenix (Nov 1993). Kelly, J. L., "A New Interpretation of Information Rate," Bell System
Technical J., 35 (1956), 917-926.
Lo, V. S. Y., "Application of Running Time Distribution Models in
References Japan," in D. B. Hausch, V. S. Y. Lo, and W. T. Ziemba, (Eds.),
Algoet, P. H. and T. M. Cover, "Asymptotic Optimality and Asymp- Efficiency of Racetrack Betting Markets, Academic Press, 1994, 237-
totic Equipartition Properties of Log-optimum Investment," The 247.
Annals of Probability, 16 (1988), 875-898. and Bacon-Shone, J., "An Approximation to Ordering Proba-

1058 MANAGEMENT SCIENCE/VOL 41, No. 6, June 1995

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms
 LO, BACON-SHONE, AND BUSCHE
Ranking Probability Models to Racetrack Betting

bilities of Multi-entry Competitions," Research report 16, De- Stern, H., "Models for Distributions on Permutations," J. American
partment of Statistics, University of Hong Kong, 1992. Statistical Association, 85, 410 (1990), 558-564.
and , "A Comparison Between Two Models for Predicting Thorp, Edward, "Portfolio Choice and the Kelly Criterion," in W. T.
Ordering Probabilities in Multi-entry Competitions," The Statis- Ziemba and R. G. Vickson (Eds.), Stochastic Optimization Models
tician, 43, 2 (1994), 317-327. in Finance, 1971.
McLean, L. C., W. T. Ziemba, and G. Blazenko, "Growth Versus Se- Ziemba, W. T. and D. B. Hausch, Beat the Racetrack, Harcourt Brace
curity in Dynamic Investment Analysis," Management Sci., 38, Jovanovich, 1984.
11(1992), 1562-1585. and , Betting at the Racetrack, Dr. Z. Investments, Inc., Los
Skinner, M., "Easy Money," OMNI, May (1989), 42-49. Angeles, 1986.

Snyder, W. W., "Horse Racing: Testing the Efficient Markets Model," and , Dr. Z's Beat the Racetrack, Revised edition, William
J. Finance, 33 (1978), 1109-1118. Morrow and Co., Inc., New York, 1987.

Accepted by William T. Ziemba, former Departmental Editor; received August 1992. This paper has been with the authors 5 months for 2 revisions.

MANAGEMENT SCIENCE/VOl. 41, No. 6, June 1995 1059

This content downloaded from 138.25.78.25 on Fri, 29 Jul 2016 19:39:35 UTC
All use subject to http://about.jstor.org/terms