Market Efficiency in Racetrack Betting

Author(s): Peter Asch, Burton G. Malkiel and Richard E. Quandt

Source: The Journal of Business, Vol. 57, No. 2 (Apr., 1984), pp. 165-175
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/2352733 .
Accessed: 22/06/2014 22:51

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

.
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].

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to The
Journal of Business.

http://www.jstor.org

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
Peter Asch
Rutgers University

Burton G. Malkiel
Yale University

Richard E. Quandt
Princeton University

Market Efficiency in
Racetrack Betting

I. Introduction Do marketinefficien-
cies in racetrackbet-
Betting at racetracksand investing in the stock ting permitprofitable
markethave several characteristicsin common. strategiesto be devised
In neither situation are future earnings known on the basis of ob-
with certainty. Both situationsare characterized served bettingpat-
terns? Logit analysis of
by a large number of participants and by the marketdata indicates
availabilityof extensive informationand profes- that profitscannot be
sional advice. Horse racing data, therefore, per- earnedin win betting,
mit interesting tests of the efficient market althoughit is possible
theory. to outperformthe aver-
age bettor substan-
Numerousauthorshave analyzedhorse racing tially. Surpris-
data and various conclusions have been con- ingly, profitsnet of
firmed: (1) There is a systematic tendency to transactionscosts on
overbet long-shots and underbet favorites place and show betting
(Rosett 1965, 1971; Snyder 1978; and Ali 1979). appearpossible, al-
thoughthere are rea-
(2) The tendency to overbet long-shots is sons to suspect that
strongest in late races (McGlothlin1956; Asch, such opportunitiesmay
Malkiel, and Quandt 1982). (3) The utility func- not be exploitableon a
tion of bettors is convex, indicating risk love substantialscale. The
(Weitzman 1965; Asch et al. 1982). Figlewsky findingsdo not suggest
irrationalbehaviorby
(1979)tested hypotheses concerningwhetherthe bettors.
betting "market" is efficient with respect to
track odds and handicappers'picks of winning
horses. By fittinga logit model with horses' win-
loss records representingthe dependentvariable
and trackodds and handicappers'picks the inde-
pendent variables,it is shown by likelihoodratio

(Journal of Business, 1984, vol. 57, no. 2)

? 1984 by The University of Chicago. All rights reserved.
0021-9398/84/5702-0003$01.50

165

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
166 Journal of Business

tests that handicappers'picks are substantiallyaccounted for in the

track odds. Finally, Hausch, Ziemba, and Rubinstein(1981)analyzed
betting horses to place (come in first or second) or to show (first,
second, or third)and found that inefficienciesexist in place and show
betting.
The present paper analyzes a new sample of racetrackresults from
AtlanticCity, New Jersey. We investigatehow well the trackodds and
the morning line odds (rather than handicappers'picks) predict the
victors of races. But there is an importantfurther question that re-
quiresattention.If track odds and other publiclyavailableinformation
do very well in predictingwinners, one might want to use such infor-
mationin betting;yet this use of informationwill tend to lower the odds
(and payoffs) of the horses predictedto win.
The pertinent question, then, concerns the efficiency of the race-
trackbettingmarket.If efficiency is perfect, the rate of returnto every
betting strategy based on available informationshould approximate
closely the track "take." For the sample of races analyzed here, the
take is about 18.5%;thus investors who pursue any strategy should
earn a negative returnclose to 18.5%.It is of particularinterestto ask,
however, not whether the racetrackbetting marketis efficient in this
very stringent sense, but whether there exist departuresfrom effi-
ciency sufficiently large to permit profitablebetting strategies. Can
one, in other words, devise strategies based on observable betting
patterns that imply positive rates of return?Only if such unexploited
arbitrageopportunities were being consistently neglected could we
conclude that the racetrackbetting marketwas inefficient.'
A second question is, If one were able to predict winners by using
publicly available information,would such a finding be inconsistent
with the assumption of rational behavior in the racetrack market?
Rosett (1965) has shown that rationalityimposes requirementson the
relationshipbetween returnand the probabilityof winningby exploit-
ing the possibilityof combinationstrategies, such as parlays.He found
specifically that the return to low-win-probabilityhorses appeared
lower than predicted by the rationalityassumption. We comment on
this issue in the concluding section.

1. An analogy may be made to empiricalresearchin the stock market.Departures

from perfect efficiencyhave often surfacedin empiricaltests, such as some findingsof
serial correlationand evidence of inferrableinformationfrom "insider"tradingthat is
not immediatelyreflectedin marketprices. Nevertheless,the conclusionremainsthat, in
general, any departuresfrom efficiency are not large enough to permitan investor to
devise an active strategythat can consistentlybeat a passive buy-and-holdstrategy.In
this sense, the empiricalevidence generallysupportsa findingof efficiencyfor the stock
market.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
Market Efficiency in Racetrack Betting 167

II. Data and Definitions

The Data Set
Observationsare based on the entire 1978thoroughbredracingseason
at the Atlantic City Race Course, which includes 712 races and 5,714
horses. The data are
1. The "morning"line odds for each horse in each race, determined
by the track's professionalhandicapperand printedin the daily racing
program.These are the handicapper'sestimates of the winningproba-
bilities for each horse that confront bettors before the start of each
betting period.
2. Parimutuelodds for each horse in each race at various points
("cycles") duringthe betting period. Twenty-fourcycles are typically
recorded, showing the minute-by-minutecourse of the actual betting
for each race. Only some of the availablecycles were employed in the
analysis.
3. Final parimutuelodds for each horse, given at the final betting
cycle. The odds are determinedby the ratio of the betting pool avail-
able for distributionto the amount bet on each horse. The racetrack
subtractsa percentage or "take" from the total betting pool to cover
taxes, expenses, and profits.In addition,the trackenjoys "breakage,"
the gain frombeing allowed to roundpayoffs downward(to the nearest
10 cents or 20 cents on a $2.00 bet), or to roundodds downward(to the
nearest tenth). The total betting pool minus the take, includingbreak-
age, is paid on the winninghorse.
4. The outcome of each race.
Some Definitions
The following definitionsare employed:B = the total amountbet on
the race; bh = the amount bet on horse h, h = 1, 2, . .. , H, where H is
the numberof horses in the race. It follows that
H

B = bh.
h= 1

Let t = the total take of the track includingbreakage,all expressed

as a percentageof the total bet. The odds of horse h may then be stated
as
O B(1 -t)-bh B(1 - t) 1
bh bh
We define the (bettors' subjective)probabilitythat horse h will win
the race, Ph. It follows then that

h B = (1 t)/(l+Oh)

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
168 Journal of Business

One can also defineimplicitor marginalodds derivedfrombets made

late in the betting period. The bettingperiod is approximately25 min-
utes, the time between races. Define cycle 1 as the betting cycle en-
compassingthe first 17 minutes (the first 2/3) of the bettingperiod, and
cycle 2 as the betting in the last 8 minutes (the last 1/3) of the bet-
ting period. The marginalodds on horse h in betting cycle 1, Oh, 1 =
[BI(I - t)lbh,11 - 1, where B1 is the total amount bet up to the
end of cycle 1, and bh,I is the amountbet on horse h up to the end of
cycle 1. The finalodds on horse 1 (whichare also the odds on the horse
derived from all betting up to and includingcycle 2), are
B2 -t)
(1-
bh, 2

where B2 is the total amountbet throughboth cycles (1 and 2) and bh, 2

is the total amountbet on horse h throughboth cycles (1 and 2).
Define the marginalodds as the odds derivedfromthe bettingduring
a particularcycle, but not including previous cycles. We seek the
marginalodds created by bettors in cycle 2, namely, the bets of the
"late money," or what is sometimes hypothesized to be the "smart
money." The argumentis that bettors with true inside informationwill
prefer not to signal that informationto the public until late in the
betting cycle to minimize "following" behavior on the part of other
track bettors.
The marginalodds on horse h in cycle 1 are simplythe regularodds,
that is:
?h,l = Oh,l

The marginalodds on horse h in cycle 2 can be derivedfromthe odds at

the end of cycles 1 and 2, as follows:
Om, - Oh, 2bh, 2 - Oh, lbh,1
bh, 2 - bh, 1

Finally, definea payoff Whfromhorse h as unity plus the finalodds if

h wins and zero otherwise. A horse that goes off at odds of three to one
pays $4.00 for each dollarbet if the horse wins. A rate of returnR from
betting one dollar on one horse in each of N races is R = (EWh - N)l
N, where 2Wh is the sum of payoffs in the N races.

III. Models and Results

Define Wijas the win function of the ith horse in the jth race, by
Wij = I'xij + Uij,
where g3'is a vector of coefficients, xija vector of observablevariables,
and uij an error term. Wijmay be thought to measure the "winning-

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
Market Efficiency in Racetrack Betting 169

ness" of a horse and to depend on a systematic part (f'x1j) and a

stochastic part. The ith horse wins thejth race if Wij> Wlj, . * *, Wij >
Wn,j,where nj is the numberof horses enteredin thejth race. If the uij
are assumed to have independent extreme value (Weibull) distribu-
tions, this leads to the well-known logit model, giving the probability
that the ith horse wins the jth race as

- exp(pYxij)
>3 exp(I'xkj)
k= 1

Estimates of the j's are obtained by maximizingthe likelihoodfunc-

tion. It is well known that the logit model possesses the independence
of irrelevantalternativesproperty:the relative odds for horses i and k
(Pij/Pkj) depend only on the characteristicsof these horses. As a first
approximationthis appearsto be reasonablefor horse racing.
Table 1 displays some basic results. All independentvariableswere
odds convertedto probabilities.Thus ML representsthe win probabili-
ties impliedby the morningline odds, F the probabilitybased on the
finalodds, and M2 the probabilitybased on the marginalodds for cycle
2 (the last 1/3of the bettingperiod).The variableM2 is includedin some
of the computationsto account for the possibility that the late bettors
(smart money) have better informationthan the aggregateof all bet-
tors. All coefficients are highly significantand, comparingruns 1, 2,
and 3 with 4 and 5, respectively, the likelihoodratio test emphatically
rejects the null hypothesis that the morningline odds or the final odds
or the marginalodds alone provide a satisfactoryexplanation.In par-
ticular,we reject the null hypothesisthat the morningline odds contain
no informationnot alreadyaccountedfor by the trackodds. The proba-
bility of winningis positively and significantlyaffected by the morning
line and final track probabilitiesas well as by the marginalprobabili-
ties. IncludingF and ML as well as M2 causes the standarderrorsto

TABLE 1 Logit Results

Logit Run

Variable 1 2 3 4 5

ML 9.99 ... ... 4.15 4.74

(.54) ... ... (1.02) (.96)
F ... 6.42 ... 4.26 ...
... (.33) ... (.62) ...
M2 ... ... 5.74 ... 3.54
... ... (.30) ... (.54)
Log likelihood -1,280.0 -1,265.0 -1,270.0 -1,257.0 -1,258.0

NOTE.-Asymptotic standard errors in parentheses.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
170 Journal of Business

increase sharply,and no furthersignificantimprovementoccurs in the

log likelihood when both these variablesare included.2
Simulations Using Only Win Bets
Using the results of the logit estimationone can performthe following
simulations.For each race, using the coefficientsin run4 or 5 one may
compute the win probabilityPijfor each horse. We may then consider
the following strategies:
Strategy 1: In each race j(j = 1, . . . , N) bet $2.00 on the horse with
the highest Pi.
Strategy2: In each race bet $2.00 on the horse with the highestPij if
P? ' y, where y is a "filter" coefficient.
Strategy3: In each race bet $2.00 on the horse with the highestPjj, if
the highest Pu is at least 3.5 times greaterthan the second highest Pii.
Strategy3 differsfrom strategy2 in its concern, not for a highPijper se,
but ratherfor selecting horses to bet only if the second choice by the
criterionfunction is sufficientlyinferiorto the first choice.
Runningsuch simulationson the samplefrom which the coefficients
were obtainedwould bias the results. We thereforedividedthe sample
into two halves and reestimated runs 4 and 5 from each of the two
halves.3 We then used the coefficientsfrom each half samplefor simu-
lations on the other half sample.
For each simulationone can compute the rate of returnfor each of
the strategies.These are displayedin table 2. The results show that all
but one of the rates of returnare better than the rate of returnimplied
by the tracktake ( - 0.185) if the marketis perfectlyefficient,as defined
above. Averagingthe rates over the two half samples shows that mar-
ginallybetter results are obtainedwith run5, thatis, when the marginal
odds pertainingto the last 8 minutesof bettingtime are used insteadof
the final odds. For one of the strategies the average rates of return
indicatethat bettingcan almostbe a fairgame. The resultsdemonstrate
clear departuresfromperfect efficiencybut do not suggestthe possibil-
ity of bettingstrategiesthat are profitableafteraccountingfor the track
take.
Simulations Using Place Bets and Show Bets
An alternativeto placing win bets is to bet the selected horse to place
or to show. The selection of the horse in each race was based on the
same strategiesdiscussed beforebut in the present simulationsplace or
show bets were made. The rates of return are shown in table 3. All

2. The standarderrorsincrease because of collinearitybetween F and M2.

3. The coefficientsfrom the half samplesand the results of the likelihoodratio tests
comparingthe half-sampleruns 4 and 5 with half-sampleruns 1, 2, and 3 are similarto
those reportedin table 1.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
Market Efficiency in Racetrack Betting 171

I
'C 00 en00 00 - 00 tf)
\0 or0 "^ "T
Ct t aor0I
N ^ ?o0O~
o14

I I I I I I I I I I

00 - Cl t- t0 00 C 0r
go oo -or -

mo 0D O - 'IC? 00 cq ut

ct 4-> o >m

; . E~~~~~~~~~~~~~~~.
en~~~~~~~~~~~~~~~~~a 0
.0

a~~~~~~~~~~~~~~~~~ t: o.o o4 'A

0
~~~~~~~~~~~C> VI) 11W c )

N
- -0-- - 00--0

2 0~~~~~~~~~~~~~

~~~~4a) cO~n'I n ) W) C9 I
~~~~~~~~~~~~0
CB li~~~~~~~~~~~~~~~~~~~~~
Co0 Ro0
0..~~~~~~~~~~~~~~~~~~~~~~~~.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
172 Journal of Business

\1 e nON ir ) eq r- W) 0
t- N1 Nt \C- 0 N -4

W~~~~~~~r oo
00 OO"'t o\0X tt

t>~~~~~~C
.L,>W)e kf

4 . ? oo en C \ oo, It 'C

4.;

F.~~~~~~~~~~ "N W) ""INOt

eq en oo en 0\ I" "t

D o o o
ooo o o o o o o

Cd CT3

t ~~ ~~~~~~% C)t-t Z C

C3
X~~~~~~~~W W) W) C>W)o
t4

-4
%0 C) 0\ m C) en eq 43 0

4- Dd \8 It *4 t- W) ON 11
m C> C

6) a:13V C

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
Market Efficiency in Racetrack Betting 173

returnsare positive. Betting the horse with the highest win probability
to place yields a higherreturnthan betting it to show. The numberof
horses on which bets are made under strategy2 with y = .5 or .6 and
under strategy 3 is sufficiently small so that one cannot place great
confidence in the resulting returnfigures. But for strategy 1, for ex-
ample, involving356 bets in each subsample,the returnto place betting
rangesfrom 0.143 to 0.176. Finally, one may note that for those strate-
gies which allow a large numberof horses to be bet on (strategy 1 and
strategy2 withy = .4), the results of run 5, which employ the morning
line and marginalodds, are slightly better.
Even simpler strategies can yield positive returns. One could, for
example, predict the winner on the basis of logit runs 1, 2, or 3, em-
ploying a single explanatoryvariable.4The results are broadlycompa-
rablebut somewhatless favorable.Thus, for example, strategy1 yields
an average returnfrom place bets (over the two subsamples)of 0.081
when the morningline alone is used to predictthe winner, 0.076 when
the final odds are used, and 0.101 when the marginalodds are em-
ployed. This still compares favorably with placing win bets on the
favorites, from which the returnis - 0.173.
We were not able to improve on the results of table 3 by employing
more sophisticatedcompound strategies.The winners and runners-up
can, for example, be predictedfrom a generalizationof the logit model
that explicitly takes into account not only which horse won, but which
horses were second and third.5These predictionshave been employed
for combinationstrategiesinvolvingwin, place, and show bets simulta-
neously.6 The most successful of these combinationstrategiesyielded
average returnsof 0.133-0.140, dependingon which explanatoryvari-
ables were used in the generalizedlogit estimation.

IV. Conclusions
The results from win betting strategies show that strategies can be
developed that exceed the returnfrom simply betting the favorite to
win. The results from using place and show bets are surprisingin the
sense that we now find that net profits can be made. Several conjec-
tures may be offered to explainthis finding.It is possible, for example,
that the tendency to overbet long shots is accentuated in place and

4. For strategy 1 this is equivalentto bettingon the favorite.

5. The generalizationfollows from Harville(1973).
6. The most successful strategyrequirespredictingfor each race three highest win
probabilitiesdenoted by P1 P2 _ P3, where 1, 2, and 3 are the indices of the corre-
spondinghorses. Accordingto this strategy,if (i) P1 - P2 > .25 andP2 - P3 > .15, bet
horse I to win, 2 to place, 3 to show; (ii) if P1 - P2 - .25 andP2 - P3 > .15, bet horse 1
to place, 2 to show; (iii)if P1 - P2 > .25, P2 - P3 ' .15, bet I to win and I to place;(iv) if
P1 - P2 ' .125, P2 - P3-C .15, bet I to show.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
174 Journal of Business

show betting. It is also possible that all the informationcontainedin the

win pool may not be efficientlyimpoundedin the place and show pool.
This findingconfirmsthe broad results of Hausch et al. (1981).
Are these results incompatiblewith the assumptionof rationality?
We think not. First, we may ask whether the relationshipbetween
returnand the probabilityof winningpredictedby Rosett (1965)would
still hold today. The reason is that the opportunitiesfor low-probability
parlay-typebets at racetrackshave greatlyincreasedin the last decade
and there is now a plethoraof exactas, trifectas, "pick fours," and so
on, that are more profitablethan parlaysin the sense that they involve
only one tracktake ratherthan n. More important,however, it is quite
possible, as Rosett suggests, that even today the marketfails to pro-
vide enough low-probabilitybetting opportunities,given a substantial
"taste" by the betting public for such bets. This clearly will have the
effect of loweringthe returnon low-probabilitybets and is compatible
with our results. Ourfindingsthat positive returnscan be earneddo not
suggest irrationalbehavior; ratherthey suggest that the taste for cer-
tain nonpecuniarybenefits associated with certaintypes of bets make
positive profits possible for some bettors.
In other words, it is not a paradoxthat bets on differenthorses have
different expected returns (and variance of returns);this is simply a
consequence of bettors' havingdifferentutilityfunctionsand selecting
different points among the available opportunities(characterizedby
differentobjective and subjective winningprobabilities).
It should be noted, in conclusion, that these results cannot be inter-
preted as firm evidence of unexploited arbitrageopportunitiesfor at
least two reasons. First, for any strategythat employs more thanjust
morningline odds, there is a serious difficultyin implementation,since
finalodds are typically not posted untilthirtyor more seconds afterthe
betting windows close. Second, the place and show pools are usually
much smallerthan the win pools. Any attemptto employ the implied
arbitrage strategies may prove infeasible since a relatively small
volume of betting could quickly wipe out the apparentprofitopportu-
nity. Finally, it goes without saying that our simulationsneed to be
repeatedover alternativeandlarge samplesbefore we can have consid-
erable confidence in the results.

References
Ali, Muktar M. 1979. Some evidence of the efficiency of a speculative market.
Econometrica 47:387-92.
Asch, P.; Malkiel, Burton G.; and Quandt, RichardE. 1982. Racetrackbetting and
informedbehavior.Journal of Financial Economics 10:187-94.
Figlewsky, Stephen. 1979. Subjective informationand marketefficiency in a betting
market.Journal of Political Economy 87:75-88.
Harville, D. A. 1973. Assigningprobabilitiesto outcomes of multi-entrycompetition.
Journal of the American Statistical Association 68:312-16.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions
Market Efficiency in Racetrack Betting 175

Hausch, D. B.; Ziemba, W. T.; and Rubinstein,M. 1981. Efficiencyof the marketfor
racetrackbetting. Management Science 27:1435-52.
McGlothlin,W. H. 1956. Stabilityof choices among uncertainalternatives.American
Journal of Psychology 69:604-15.
Rosett, R. N. 1965. Gamblingand rationality.Journal of Political Economy 73:595-607.
Rosett, R. N. 1971. Weak experimentalverificationof the expected utility hypothesis.
Review of Economic Studies 38:481-92.
Snyder,WayneN. 1978.Horse racing:Testingthe efficiencymarketsmodel.Journal of
Finance 22:1109-18.
Weitzman,M. 1965.Utility analysisand groupbehavior:An empiricalstudy.Journal of
Political Economy 73:18-26.

This content downloaded from 91.229.248.204 on Sun, 22 Jun 2014 22:51:26 PM

All use subject to JSTOR Terms and Conditions