A Chemical Engineer Goes to the Horse Races

Jacob H. Lashover
August 28, 2014
Baton Rouge, Louisiana
Abstract
This paper presents a mathematical model for the optimal wagers one should
make on a horse race which uses the pari-mutuel1 wagering system. It is an
extension of Rufus Isaacs model which he developed in 1953 while working for
the RAND Corporation (Isaacs R., 1953). Isaacs solution, using the calculus of
the Newton-Rhapson method, produced wagers that violated the models key
assumption that the size of his wagers would not significantly affect the odds. It
also produced optimal wagers that lost money even though the selected horse
won the race. The purpose of this paper is to demonstrate that the mathematical
method, Monte Carlo Marching (MCM), can be used to improve the solution of the
Isaacs and other nonlinear models (Lashover, 2012). MCM was previously
developed for chemical engineering calculations to guarantee convergence and to
avoid taking the complicated calculus derivatives required for Newton-Rhapson
solutions. While Isaacs showed his considerable prowess with calculus in solving
his nonlinear model, it is doubtful that even he could have expressed and used
calculus to solve the more rigorous model easily developed and solved with MCM.
Using linear programming type constraints, MCM permitted improving the rigor of
Isaacs model by eliminating winning wagers that lost money on multiple bets,
and placing limits on the size of wagers to prevent them from significantly
affecting the horses odds. A constraint was also added to prevent the bettor from
tapping outa situation which Isaacs did not address. While Isaacs model
failed to prove profitable over a series of 540 races, the MCM model produced a
profit of $177,314 over the same races with a 242% ROI (return on investment)
and 27% winners. The longest losing streak was 9 with wealth drawdown of
$6171 on $100,000 of initial wealth. This demonstrated the efficacy of MCM for
solving nonlinear equations other than those encountered in chemical engineering.
1

See Appendix A for explanation of pari-mutuel.

8/28/2014 JHL

Page 1

Keywords: Pari-mutuel, Monte Carlo Marching, optimal wagers, simulation,

expected value, non-linear models

Introduction
Rufus P. Isaacs, while a mathematician at the RAND Corporation in 19451955, worked with Richard Bellman (Dynamic Programming), John Nash2, and
other prominent mathematicians to advance the field of mathematical optimization
and control theory. RAND (Research ANd Development) is a think tank which
was formed by the U. S. Army Air Force shortly after the end of World War II (as
publicly stated) to focus on global policy issues. Located in Santa Monica,
California, its first mission was to solve such problems as The minimum time
interception problem for fighter aircraft where due to the increased speed
of aircraft, nonlinear terms no longer could be neglected (Pesch, 2009). Many
prominent mathematicians were brought on board, and much of their work
involved the defense and avionics industries, and was classified. After leaving
RAND in 1955, Isaacs became famous as a game theorist with his publication of
Differential Games a decade later in 1965 (Isaacs, Differential Games, 1965).
Earlier, in 1953, he had published Optimal Horse Race Bets where he used the
Newton-Rhapson method to solve the nonlinear equation of his model (Isaacs, R.
1953).
The purpose of this paper is to demonstrate that the mathematical method,
Monte Carlo Marching (MCM), previously developed for chemical engineering
calculations to guarantee convergence and to avoid taking the complicated calculus
derivatives required for Newton-Rhapson solutions, can be used to improve the
solution of the Isaacs and other nonlinear models (Lashover, 2012). Use of MCM
permitted modification of Isaacs calculus solution which violated his key
assumption and failed to produce a profit. MCMs use of linear programming type
constraints eliminated winning wagers that lost money on multiple winning bets
and prevented the size of optimal wagers from changing the horses odds and thus
invalidating the model. Another constraint prevented the size of total wagers from
leading to Gamblers Ruin (running out of money) which Isaacs did not consider.
Using MCM and identical probability data, the Isaacs model was converted from a
2

Played by Russell Crowe in the movie, A Beautiful Mind.

8/28/2014 JHL

Page 2

non-profitable system to one which won 27% of its races with a 241% ROI (return
on investment). The longest losing streak was nine with wealth drawdown of
$6171 on $100,000 of initial wealth. Profit was $177, 314.

Development of Isaacs model

To mathematically describe the pari-mutuel model of Rufus Isaacs, we assign
bettors B1, B2, .. Bn handicapping a horse race involving m horses H1, H2, ..Hm.
We acquire an n x m matrix by assigning a subjective probability matrix {pi j},
where pi j designates the probability, in the opinion of Bi, that Hj will win the race.
A sum bi > 0 is then wagered by Bi in a manner that maximizes his mathematical
expectation. That is, Bi observes the pari-mutuel probabilities 1, 2, .. m,
related to the subjective probabilities by the track take, and indicated by the track
tote board as odds, that horses H1, H2, ..Hm , respectively, might win the race.
The real or objective probability of the horse winning the race, j, can be
determined by mathematical analysis of the key predictor variables. He then
follows a strategy of distributing the amount bi among those horses Hj for which
the ratio i/j, the expected value, Ei, is a maximum.
Isaacs assumes that the sum bi is small with respect to the total amount
wagered by the n bettors on the race and therefore does not significantly change
the pari-mutuel probabilities or odds. He further assumes that each column of the
matrix {pi j} contains at least one entry, otherwise, if the jth column consists of all
zeroes, no bettor has selected horse Hj and it can be eliminated.
The pari-mutuel system is described by the following three conditions:
1) j=1m i j = bi for all horses where i j is the sum wagered by Bi on Hj. (1)
2) i=1n i j = kj for all bettors where k is the proportionality constant
relating the amount bet on each horse to its pari-mutuel probability.

(2)

3) Ei = i/j > 1 meaning that each Bi bets only on horses with expected
value > 1.
(3)
Pari-mutuel probabilities for the Win are determined by the proportionality
constant:
k = (1 K) i = 1n bi
8/28/2014 JHL

(4)
Page 3

where K is the track take so that

j = i =1n i j / [1 K]

n
i = 1 bi

(5)

The odds on each horse are calculated as (1 j) / j to 1

(6)

He now has enough information to construct the model to be optimized:

First, he let the total wealth bet on the jth horse, kj, be partitioned into the
sum, sj wagered by the subjective crowd and the amount tj be that contributed by
the bettor. The bettors profit can be represented by the function F(t1, t2, ..,tm)
from an investment spread over the m horses as follows:
F(t) = (1 K)[j=1m (sj + tj )] j=1m j tj / (sj + tj ) -

m
j=1

(7)

Now it is desired to select values of tj >= 0 that will maximize F(t) where the
maximal F(t) possesses a positive value. From this point on, Rufus Isaacs, a pure
mathematician, used calculus to solve the problem. First, he showed that F has a
positive maximum if3
1 <= j <= m j /

sj > 1 / (1 K) sj

(8)

Isaacs then used the Newton-Rhapson method and began by taking the derivative
of F with respect to t and setting it equal to zero to find the maximum yields.4
=0

F/t = (1 K) j=1m j tj/ (sj + tj) + (1 K) i si / (si + ti)^2 j=1m(sj + tavg) 1

(9)

The quantity i si / (si + ti avg)^2 exhibits the same value for all i such that ti avg > 0.
Defining this value by 1/^2 we have ti = (i si)^0.5 - si

(10)

A maximal form of F never occurs for all ti avg > 0, i.e. the bettor never bets on all
horses. Rather, there may exist some number of horses, r, whose real

Isaacs did not take into account that a positive maximum could be achieved with wagers on some winning horses
not producing an overall profit for the race. Colloquially, this is called betting against yourself and happens
frequently when bettors make numerous bets on a race without adequate analysis.
4
Taking the derivative of the model function(s) is not necessary when using Monte Carlo Marching.

8/28/2014 JHL

Page 4

expectation is greater than the subjective expectation realized from the actual
wealth invested on their behalf by the crowd.
Therefore we look for ti = k (i si)^0.5 - si For i = k, k+1,m, and
k^2 = (1 K) j=1k-1 sj [1 (1 K) j=k m j ]^-1.

(11)
(12)

Table 1. Rufus Isaacs Synthesized Horse Race

Horse
1
2
3
4
5
6
7
Total

Real Prob. $ Win Pool Crowd Prob. Calc. Odds Tote Odds
0.4
0.18
0.12
0.1
0.1
0.05
0.05

35000
10000
9000
8000
7000
5000
1000
75000

0.4666
0.1334
0.1200
0.1067
0.0933
0.0666
0.0133
0.9999

1.1432
6.496
7.333
8.372
9.718
14.015
74.188

1
6
7
8
10
14
74

Prob.
-0.0666
0.0466
0
-0.0067
0.0067
-0.0166
0.0367

Horses 2, 5, and 7 have real, positive expectations which are greater than
their subjective expectations, i.e. their expected values, Ei, are calculated as
E2 = 0.18/0.1334 = 1.35
E5 = 0.1/0.0933 = 1.07
E7 = 0.05/0.0133 = 3.76
And, ^2 = {[0.85(35,000 + 9,000 + 8,000 + 5,000)] / [1 0.85(0.18 + 0.10
+ 0.05]} ^ 0.5 = 259.5 when K = 0.15.
Thus the optimal wagers on each horse are calculated as:
t2 = 259.5 x [0.18 x 10,000]^0.5 - $10,000 = $1009.65
t5 = 259.5 x [0.10 x 7,000]^0.5 - $7,000 = < 0
t7 =259.5 x [0.05 x $1,000]^0.5 - $1,000 = $834.94

8/28/2014 JHL

Page 5

One can easily see that the wager on Horse 7 will violate the assumption that
the wagers do not affect the tote odds. The total of $834.94 wagered by Isaacs
solution on Horse 7 will almost double the total of $1,000 by the crowd on Horse
7. The new total wager, s7 , of $1,834.94 will reduce the odds from 74 to 40 to
one, thus reducing the pay-off should Horse 7 win, and invalidating the calculation
of F. Further, no consideration is given to the bettors wealth, and the size of these
wagers could easily lead to tapping out or gamblers ruin5 over a series of
races. Checking for wagers which would not produce an overall profit for the race
was also neglected. Finally, a small point, but the win bet is an integer problem, as
whole dollars, not dollars and cents, are the only acceptable win wagers.

Development of the Monte Carlo Marching model

The Monte Carlo Marching (MCM) method presented here provides robust
convergence performance without derivatives and is insensitive to initial estimates
of the variables. Application of MCM was demonstrated in an earlier paper on
several non-linear models including Isaac Newtons cubic polynomial, y^3 2y -5
= 0, (Cajori, 1911) which Newton used to first demonstrate his Newton-Raphson
procedure. There have been no convergence failures with MCM. (Lashover,
2012)
To prove that Monte Carlo Marching can be applied to solution of general
non-linear systems of equations, Isaacs synthetic case was solved, as presented,
using MCM. The results were a $1,034 wager on #2 and an $839 wager on #7 as
compared to $1,010 and $835, respectively by Isaacs. The maximum F calculated
using MCM was 719.7185 whereas Isaacs maximum F was 719.6997. Isaacs
model lost money over 540 randomly chosen races6. Appropriate constraints were
then introduced into the MCM model and a net profit of $177,314 was achieved
5

It can be shown that for a series of 200 races, with a probability of losing and winning of 0.5, that the longest run
of wins or losses is expected to be 7. The 95% confidence limits with p = 0.5 are 3 so one would expect the
longest runs of wins or losses to be 7 3 or between 4 and 10. A bettor with $100 to wager over 10 races must
therefore not wager more than $10 per race when betting on odds-on favorites. When the probability of losing is
higher, the losing streaks can be expected to be longer, i.e. with p(losing) = 0.6, streaks of 9 are expected
confidence limits. (Schilling, 2012)
6
The races were the most recent in the authors data bank of over 13,000 races with the last race at Santa Anita in
November, 2013. Objective probabilities were obtained from the authors probability model which is not shown
here.

8/28/2014 JHL

Page 6

over the same 540 randomly selected races using identical probability values. The
ROI was 242% with a win percentage of 27%. The longest losing streak was nine
with wealth drawdown of $6171.
Marching, Binary-chop, Interval-halving or Bi-section is a simple
convergence procedure which guarantees a solution to bounded variables. Two
examples are discussed below.
Example No. 1

Consider the solution of Equation 6 above for the probability, j, when the
tote board odds are known. These odds on each horse are calculated as (1 j) / j
to 1. This non-linear equation is not solvable using algebra. The following
simple QB647 program using MCM easily solves for j. The probability, P, varies
from 0 to 1. When solving for O, odds, using Marching, the lower limit, PL is
set to zero and the upper limit, PU, is set equal to 1.0. The initial value of P is
calculated as P = (PL + PU)/ 2 = 0.50. If the first iteration produces a negative
difference, DEL, between the actual value of O, odds, and the calculated odds, OO,
PL is set equal to P. PU is left at 1.0. If the difference, DEL, is positive, PU is set
equal to P, and PL is left at 0. The next value of P to be tried is again calculated as
P = (PL + PU)/2 and the process is repeated. So, if P is too low, the lower limit,
PL, is raised to P as no solution exists below P. Similarly, if P is too high, values
above PU can be eliminated. When the correct probability is calculated there
will be no difference (within tolerance) between the actual and calculated odds.
DEL eventually becomes very small, and below the acceptable tolerance, TOL.
This method always converges and eliminates one-half of the range of feasible
solutions after each iteration. This can be represented algebraically by the function
f(x) = R / (Xn) where R = the range of solutions, X = 2, and n = 1, 2, 3.n, the
number of the iteration. The limit of this function as n approaches infinity is 0
where the entire feasible range of solutions has been examined. It does not
produce an exact solution like algebra or calculus can, however it will be shown

QB64 is a modern version of Microsoft QUICKBASIC which Microsoft no longer supports. It is an open source
programming language that will run on 32 or 64 bit machines with Windows XP, VISTA, Windows 7, Linux and MAC
OSX. It is compatible with VBA and has many features such as stereo sound, graphics loading and transformations,
TCP/IP internet capabilities, devices (joysticks), screen capture, TTF fonts, UNICODE and IME input, and clipboard
access.

8/28/2014 JHL

Page 7

that this convergence procedure permits use of constraints which are not easily
included when using algebra or calculus.
The QB64 statements demonstrating this method follow:
REM SOLVE NONLINEAR ODDS/PROBABILITY EQUATION
INPUT " TOTE BOARD ODDS= ", O
LP = 0.0: UP = 1.0: TOL = 0.0001
2 P = (LP + UP) / 2.0
OO = (1 - P) / P
DEL = O - OO
IF ABS(DEL) <= TOL THEN PRINT: PRINT " P= "; P: GOTO 9 'SOLVED
IF DEL < 0 THEN LP = P: GOTO 2
IF DEL > 0 THEN UP = P: GOTO 2
9 END
Computer output demonstrating the calculation of probability for tote odds
of 2 follows:

Table 2. Convert Tote Odds to Probability

DATE= 06-20-2014 TIME= 17:43:04
ODDS FOR WHICH PROBABILITY IS NEEDED= 2
LP=0.0000 UP=1.0000 P=0.5000 DEL= 1.0000 CYCLE= 1
LP=0.0000 UP=0.5000 P=0.2500 DEL=-1.0000 CYCLE= 2
LP=0.2500 UP=0.5000 P=0.3750 DEL= 0.3333 CYCLE= 3
LP=0.2500 UP=0.3750 P=0.3125 DEL=-0.2000 CYCLE= 4

8/28/2014 JHL

Page 8

LP=0.3125 UP=0.3750 P=0.3438 DEL= 0.0909 CYCLE= 5

LP=0.3125 UP=0.3438 P=0.3281 DEL=-0.0476 CYCLE= 6
LP=0.3281 UP=0.3438 P=0.3359 DEL= 0.0233 CYCLE= 7
LP=0.3281 UP=0.3359 P=0.3320 DEL=-0.0118 CYCLE= 8
LP=0.3320 UP=0.3359 P=0.3340 DEL= 0.0058 CYCLE= 9
LP=0.3320 UP=0.3340 P=0.3330 DEL=-0.0029 CYCLE= 10
LP=0.3330 UP=0.3340 P=0.3335 DEL= 0.0015 CYCLE= 11
LP=0.3330 UP=0.3335 P=0.3333 DEL=-0.0007 CYCLE= 12
LP=0.3333 UP=0.3335 P=0.3334 DEL= 0.0004 CYCLE= 13
LP=0.3333 UP=0.3334 P=0.3333 DEL=-0.0002 CYCLE= 14
LP=0.3333 UP=0.3334 P=0.3333 DEL= 0.0001 CYCLE= 15
P= .3333435
Example No. 2

(See Lashover, J. H. (2012). Monte Carlo Marching. Baton Rouge, Louisiana:

"Academia.edu" for a more detailed explanation of this example.)

Consider the chemical engineering calculation of the fraction of vapor

flashed, V, from a liquid stream which varies from 0 to 100 % or from 0 to 1.0 as a
fraction. The concentration of the liquid stream (feed) is known as are its
temperature and pressure, however the concentrations of both the vapor and liquid
streams formed are unknown. To complicate matters further, the composition of
the vapor stream formed is related to the concentration of the liquid stream formed
by a complex, non-linear relationship, i.e. the composition of the vapor stream is
dependent on the yet unknown liquid stream. The mass of the feed must also equal
the total mass of the resulting liquid and vapor streams (conservation of mass).
There have been many solutions presented, including the authors using MCM, but
others fail frequently, especially when V is close to 0 or 1.
When solving for V using Marching, the lower limit, VL is set to zero and
the upper limit, VU, is set equal to 1.0. The liquid and vapor concentrations

8/28/2014 JHL

Page 9

formed are also bounded between 0 and 1 when expressed as mole fractions and
the total molar concentration of each stream must also sum to unity. An overall
mass balance must be satisfied and calculation of the total pressure must equal the
pressure specified. When one mole of feed is flashed and the total pressure is
equal to one atmosphere, all variables can have a lower limit of 0 and an upper
limit of 1. Since the vapor composition is dependent on the liquid composition, the
liquid compositions become the unknown independent variables. With three
components, there are then four unknown independent variables when V is
included.
In MCM, the simple marching technique shown previously is applied to the
multiple variables. These unknown variables are essentially improved from their
lower boundary values toward their higher boundary values by analyzing linear
regions which are inverse multiples of powers of two. This procedure emulates
the guaranteed conversion procedure of Marching and satisfies the mathematical
theories of hyper-rectangles. (Faloutsos, 1903) Remember, after the first iteration
of marching, the region left to study has decreased by or 1/ (2^1). After the
second iteration, the region left to study has decreased to or 1/ (2^2), and so
forth. The feasible region of solutions for V, after being reduced to by searching
either 0.5 to 1.0 or 0 to 0.5 is next reduced to by searching either 0.75 to 1.0 or 0
to 0.25. By normalizing all independent variables to 0 to 1 or 0 to 100, no one
variable can overly influence the objective function.
Finally, the objective function is the sum of the convex constraint equations
which are crafted as the absolute differences of the constraint values from the
desired constraint values at each iteration, i.e., for the V calculation, G1 = ABS(
X 1.0) would be one constraint, G2 = ABS(Y 1.0) would be the second, G3
= ABS(P 1.0) would be the third, and G4 = ABS(Z V*Y (1 V)*X).
Thus G, profit in most Simplex calculations, would equal to G1 + G2 + G3 +
G48 and comprise the convex objective function which would be minimized in the
calculation of V. The justification for the addition of these constraints is similar
to that used in Maximum Likelihood methods. (Banbura, "Maximum Likelihood
Estimation of Factor Models on Data Sets with Arbitrary) The objective function
8

V is the fraction of vapor flashed, X and Y are the liquid and vapor mole fractions, respectively, and P is the total
system pressure in kPa. The sum of the X's and Y's must equal 1. G4 is the overall mass balance.

8/28/2014 JHL

Page 10

becomes convex by being built of convex functions and can thus be optimized.
The feasible solutions are bounded so as to insure that they are realistic parts of the
model set. Further proof of the linearity (ability to add constraints) of the
simultaneous marching schema used in this work can be found in Analysis of the
n-dimensional quadtree decomposition for arbitrary hyper-rectangles by Christos
Faloutsos et.al. at http://drum.lib.umd.edu/bitstream/1903/678/2/CS-TR-3381.pdf
(Faloutsos, 1903).
In the horse race model, F replaces G and is maximized. When maximizing,
bets which produce an F below the present best are rejected. Also, in the Isaacs'
model Eq. 7 is the objective function, and the constraints on bet size, etc. are tested
separately.

Solving the Isaacs Model using MCM

The same nomenclature used for the Isaacs model is used here and begins
with the solution of Eq. 7. Before the solution is started, races with less than 10
entries and with total win pool wagers less than $8,000 at three minutes to Post
(all horses are in the starting gate) are screened out. See Appendix A:
Background for the horse racing MCM model for the origin of these and other
constraints.
>BETA is the maximum percentage of the bettors wealth which can be
risked on any one race. This is to avoid tapping out or gamblers ruin. The
author has used a conservative 5% as opposed to Quirins 8% where his
simulations required a winning percent of 30 along with an ROI of 25% to avoid
tapping out. Quirins simulations also showed that betting 8+ to 10% of ones
bankroll per race led to a 50-50 chance of making a profit or tapping out (William
L. Quirin, 1979).
>BTEFF is the maximum percentage of the win pool dollars that can be
wagered on any horse to avoid significantly changing the tote board odds. The
author has used 5% here. At 5% the bet on #7 in Isaacs sample race would be
limited to $50 (0.05 x $1,000) as opposed to his $838 wager which cut the odds
from 74 to 40 (46 %).9 The $50 wager cut the odds from 74 to 70 (5.4 %). The
9

This precipitous drop in horse #7s odds invalidates the model as the horse will pay significantly less than the odds
used in the calculations.

8/28/2014 JHL

Page 11

wager on #2 was reduced to $500 (0.05 x $10,000) and the odds were reduced
from 6.4 to 6.1 (4.7%) as compared to 6.4 to 5.9 (9.3%) for the Isaacs solution.
These calculations show only the effects of the optimal bets and do not consider
any dollars wagered by the crowd.10
In each iteration, a number of feasible solutions are calculated for the
marching region and the highest F with its ts (bets on each horse) is stored as a
local maximum, M. The number of solutions examined can easily be increased to
insure location of a global stationary point with little penalty in computing time.
The values of the variables used for each solution are obtained by random (Monte
Carlo) selection of values in the iteration ranges. These random values are first
tested to insure that they are within the current boundaries of the region to be
tested. If not they are rejected and a new set of values are determined.
It should be noted that the solution to the Isaacs model was quickly
determined with a relatively coarse grid of solutions. First, even though there were
7 horses in the race, the final solution only had to consider those horses with
expected values above 1.0. This reduced the number of possible wagers from 7 to
3. Taking a helicopter view of the model, it appears optimal to bet the maximum
that one can on the horses with the highest expected values. However, as
mentioned previously, this can lead to wagers on some horses that are not
profitable if they win. Further, the length of losing streaks encountered will be
smallest when betting on the horses with the highest probability of winning. As
one can intuit, the ideal case will be that of the race favorite having the highest
expected value. In fact, this often occurs during the last race at the track when
gamblers who are losing for the day bet long shots to try and recover their losses
letting favorites go off at profitable odds. Favorites win about 1/3 of their races,
but usually show a negative profit due to being over bet, i.e., they are underlays.
The exception here being favorites with odds less than or equal to 1 to 1 as
predicted by the favorite-long shot bias. These favorites are frequently overlays.

10

Many bettors or punters as they are called in England handicap the handicappers*, i.e. they watch for
significant wagers on a horse as seen by the drop in his odds on the tote board. They assume that these bettors
have inside information and therefore also wager heavily on the same horse further driving down the odds.

8/28/2014 JHL

Page 12

Long losing streaks provide a serious psychological dilemma to gamblers.

With each successive loss, the bettor sees his wealth becoming smaller and begins
to doubt whether he is properly executing his betting system. The modified Isaacs
model had a maximum losing streak of 6 during which the initial wealth was
reduced by about 15%. Higher percentages would provide serious stress to most
bettors.
Another issue which was learned from stock market technical analysts is
that no more than 25% of the profit from any investment system should come from
any one investment or wager. Such a large profit from just one investment might
be the only reason that the system was successful and may not occur again in the
future. The technical analysts also caution that only of the database should be
used for optimization simulations, while saving the last of the database to test
the system. Using the total database for optimization and then testing the system
on the total database may mean that the method will only work on that database.
This fallacious procedure produces what they call a back-optimized system.
As has been shown above, it takes significant expertise to modify an
academic model to succeed in the real world. Much of the optimization work in
the literature features very simplistic examples which are not verified using actual
data from real situations. As an unknown statistician once lamented, Numbers are
like people, if you torture them long enough eventually they will tell you anything
that you want to hear. MCM can solve simultaneous nonlinear equations, and its
facility for constraints has shown it to be adaptable for work on real world
problems which require verifiable solutions (Lashover, 2012).

QB64 Computer code for the modified Isaacs model

The MCM strategy is to create a large W-dimensional11 hyper-rectangle
inside the limits of feasible solutions, always storing the best answers so far, and
re-centering the rectangle about those best answers. Then reduce the Wdimensional rectangle in width in each dimension and repeat the process. This is
done until the W-dimensional rectangle moves around and surrounds the solution,
and then further narrows and finds it when the maximum value no longer changes.

11

W is the number of horses.

8/28/2014 JHL

Page 13

REM PARIMUTUEL BETTING BY JACK LASHOVER 032914

CLS
DIM A%(20), B%(20), N%(20), LL%(20), UL%(20), S(20), T%(20), P(20)
DIM O(20), E(20), PIE(20) % indicates integer variable
DEFSNG A-Z
DEFINT I-K
RACEFIL$ = " SYNTHETIC RACE"
REM INPUT DATA FOR PARIMUTUEL CALCS. USING ISAACS
SYNTHETIC DATA"
PRINT: PRINT RACEFIL$
W = 7 NUMBER OF HORSES
TT = 0.15 'TRACK TAKE
REM DURING ACTUAL WAGERING OR SIMULATIONS, THERE IS A
SUBROUTINE HERE THAT SCRAPES THE WAGER AMOUNTS FROM AN
INTERNET SITE AND ANOTHER WHICH CALCULATES THE
(HOPEFULLY) REAL PROBABILITY OF EACH HORSE WINNING. THE
FOLLOWING NUMBERS ARE FROM ISAACS EXAMPLE PROBLEM:
P(1) = 0.40 'TRUE PROBABILITIES OF ENTRY WINNING FROM ISAACS
P(2) = 0.18
P(3) = 0.12
P(4) = 0.10
P(5) = 0.10
P(6) = 0.05
P(7) = 0.05
S(1) = 35000 'AMOUNT BET BY CROWD ON EACH ENTRY FROM ISAACS
S(2) = 10000
S(3) = 9000
S(4) = 8000
S(5) = 7000
S(6) = 5000
S(7) = 1000
TWP = $75,000 REM TOTAL WIN WAGERS BY CROWD
REM CALCULATE PROBABILITIES & ODDS FROM AMOUNTS WAGERED
8/28/2014 JHL

Page 14

FOR I = 1 TO W
PIE(I) = S(I) / ((1 - TT) * TWP)
O(I) = (1 - PIE(I)) / PIE(I)
NEXT I
REM CALCULATE EXPECTED VALUES
FOR I = 1 TO W
E(I) = P(I) / PIE(I)
NEXT I
REM THE FOLLOWING CODE IS THEN USED TO MANIPULATE THE
HYPER-RECTANGLES TOWARDS CONVERGENCE TO A SOLUTION:
REM CALCULATE FIRST GUESSES AT BETS
FOR I = 1 TO W
B%(I) = 2 MINIMUM BET ACCEPTED BY TRACK
N%(I) = 2000 MAXIMUM EXPECTED BET
A%(I) = (B%(I) + N%(I)) / 2 FIRST GUESS BY MARCHING
NEXT I
RANDOMIZE (TIMER)' USE OF A CONSTANT IN PLACE OF 'TIMER'
PERMITS OBTAINING SAME RANDOM NUMBERS EACH RUN
M=0
BTEFF = 0.05 'BETS CANNOT EXCEED 0.05 * S(K) TO AVOID CHANGING
ODDS
BETA = 0.05 'TOTAL BETS CANNOT EXCEED 0.05 * TOTAL WEALTH TO
AVOID TAPPING OUT
WINWEALTH = 100000
REM ASSUME 5000 (5 x 1000) SOLUTIONS IS ENOUGH TO FIND GLOBAL
MAXIMUM
1 J = 1 TO 5 Exponent of 2 which reduces the size of the hyper-rectangle
I = 1 TO 1000 Size of grid which determines how many solutions are checked
K = 1 to W Number of horses
8/28/2014 JHL

Page 15

IF E(K) < 1.0 THEN A%(K) =0 : T%(K) = 0: GOTO 70 'SKIP THIS HORSE
IF A%(K) N%(K) / 2 ^ J < 0 THEN GOTO 10
GOTO 20
10 LL%(K) = B%(K) $2 MIN BET
GOTO 30
20 LL%(K) = A%(K) N%(K) / 2 ^ J
30 IF A%(K) + N%(K) / 2 ^ J > N%(K) THEN GOTO 40
GOTO 50
40 UL%(K) = N%(K) LL%(K)
GOTO 60
50 UL%(K) = A%(K) + N%(K) / 2 ^ J LL%(K)
60 T%(K) = LL%(K) + INT(RND * UL%(K) + 0.5)
IF T%(K) > (BTEFF * S(K)) THEN T%(K) = INT(BTEFF * S(K) + 0.5)
'NOT USED BY ISAACS METHOD
70 NEXT K
These T%(K)s are used to calculate F (Profit) in Isaacs Eq. 7. If F is larger
than the previous best F then the T%(K)s are converted to A%(I)s which are now
the best bets calculated so far, and this F becomes the new maximum. These
calculations are repeated with as fine a grid as necessary until the Fs or maximum
ceases to change. By observing how many new Fs are obtained per iteration of J,
the size of the hyper-rectangle, one can use a finer grid by increasing J and I.
REM CHECK TOTAL AMOUNT OF ALL BETS & PROFIT
TOTALBETS = 0
FOR II = 1 TO W
TOTALBETS = TOTALBETS + T%(II)
NEXT II
IF TOTALBETS > (BETA * WINWEALTH) THEN PRINT: PRINT "
MAXIMUM TOTAL BETS EXCEED GAMBLER'S RUIN.": GOTO 1 NOT
8/28/2014 JHL

Page 16

USED BY RUFUS ISAACS METHOD

REM CHECK PROFITABILITY OF BETS
FOR II = 1 TO W
IF T%(II) <= 0 THEN GOTO 2
IF T%(II) * (O(II) + 1) < TOTALBETS THEN GOTO 1
2 NEXT II
REM CALCULATE FUNCTION (EQ. 7)TO BE MAXIMIZED
REM BREAK UP FUNCTION INTO PARTS FOR EASIER CALCULATION
PART1 = 0: PART2 = 0: PART3 = 0
FOR JJ = 1 TO W
IF E(JJ) <= 1 THEN T%(JJ) = 0
PART1 = PART1 + (1 - TT) * (S(JJ) + T%(JJ))
PART2 = PART2 + (P(JJ) * T%(JJ)) / (S(JJ) + T%(JJ))
PART3 = PART3 + T%(JJ)
F = (PART1 * PART2) - PART3
NEXT JJ
PRINT: PRINT " F= "; F; PART1; PART2; PART3
IF F > M THEN GOTO 21 NEW MAXIMUM FOUND
GOTO 22NEW MAXIMUM NOT FOUND
21 FOR II = 1 TO W
A%(II) = T%(II)
NEXT II
M=F
22 NEXT I
NEXT J
CLS: PRINT: PRINT " FINAL RESULTS "
FOR I = 1 TO W
IF T%(I) = 0 THEN GOTO 25
PRINT: PRINT BET ON HORSE= ; I; = ; T%(I)
25 NEXT I
PRINT: PRINT " TOTAL BETS= "; TOTALBETS
PRINT " MAXIMUM F= "; M
END
8/28/2014 JHL

Page 17

Table 3. Typical Race Odds Data from Internet

Santa Anita
Race 9

Time: 6:20 11/02/2013

2 mins. To Post

No. M/L12 Odds

Win

Place

Show

15

24

43581

18397

13225

30

83

13361

5309

5529

20

32

33370

14910

13578

176178

64529

39410

30

67

16567

6030

5717

30

75

14782

5133

5707

8/5

431320

92080

66800

216684

75385

53215

9/2

121328

47828

29521

10

124802

56633

37321

11

12

84867

25352

17906

12

17

59848

24161

16607

1336688

435747

304536

Pool Totals:

12

M/L stands for Morning Line which are the odds proposed by the track handicapper before the race.

8/28/2014 JHL

Page 18

Table 4. PARBET RESULTS

WELCOME TO PARBET 05-26-2014
BEGINNING RACE= 13000 ENDING RACE= 13540
BASE BET= 2 BETA= .05 TRP= .79 TRACK TAKE= .2113
BTEFF= .05 TWP MIN= $8,000.
ALL INITIAL WEALTHS=$100,000.
RACE FILES 13000-13540
AFTER RACENO= 13540 TOT WIN$ BET= 73,383 TOT WIN PAY= 250,697
ROI= 242%
WIN BETCOUNT= 66 WINCOUNT= 18 WPCT= 27
WIN PROFIT$= 177,314 WIN WEALTH$= 277,314
CURRENT LOSING STREAK= 0 DRAWDOWN= 0
LONGEST LOSING STREAK= 9 LARGEST DRAWDOWN$= 6171

13

Note that Isaacs used 0.15 which should lead to more favorable profits, i.e. more of the win pool is available to
bettors.

8/28/2014 JHL

Page 19

Bibliography
Banbura, M. a. ("Maximum Likelihood Estimation of Factor Models on Data Sets with Arbitrary, Pattern
of Missing Data"). Social Science Research Network . Retrieved September 1, 2012, from SSRN
Electronic Library (Working Paper Series No. 1189/May 2010,:
http://ssrn.com/abstract_id=1598302
Cajori, F. (1911). Historical Note on the Newton-Raphson Method of Approximation. The American
Mathematical Monthly, Vol. 18, No. 2, 29-32.
Cuir, A. D. (1990). Superhorse: The Secret to Profitable Handicapping. Brooklyn, New York: NMS
Publishing.
Faloutsos, C. (1903). Analysis of the n-dimensional quadtree decomposition for arbitrary hyperrectangles. Retrieved September 15, 2012, from
http://drum.lib.umd.edu/bitstream/1903/678/2/CS-TR-3381.pdf
Hartle, T. R. (1998). Editor. Technical Analysis of Stocks and Commodities.
Hope, B. (2014, April 3). Fast Traders Fight Over Secret Code. The Wall Street Journal, pp. 1A-12A.
Isaacs, R. (1953). Optimal Horse Race Bets. American Math Monthly, 60, No. 5, 310-315.
Isaacs, R. (1965). Differential Games. New York: John Wiley and Sons.
J. L. Kelly, J. (1956). A New Interpretation of Information Rate. Bell System Technical Journal, 917-926.
Lashover, J. H. (2012). Monte Carlo Marching. Baton Rouge, Louisiana: Academia.edu.
Lewis, M. (2014, April 7). Flash Boys. The Wall Street Journal, p. 1A.
Pesch, H. J. (2009). The Maximum Principle of optimal control: A history of ingenious ideas and missed
opportunities. Control and Cybernetics, Vol. 38, No. 4A.
Press, G. B. (2011). 150 Professional Horserace Handicapping Systems. Gamblers Book Club Press.
Schilling, M. F. (2012). The Surprising Predictability of Long Runs. Mathematics Magazine 85, No. 2, 141149.
Thorp, E. O. (1975). Portfolio Choice and the Kelly Criterion. Stochastic Optimization Models in Finance,
599-619.
William L. Quirin, P. (1979). Winning at the Races, Computer Discoveries in Thoroughbred Handicapping.
New York: William Morrow & Company, Inc.
Ziemba, W. T. (1981). Efficiency of the Market for Racetrack Betting. Management Science

8/28/2014 JHL

Page 20

APPENDIX A:

Background for the horse racing MCM model

Pari-Mutuel Wagering
Pari-Mutuel wagering, now used at all U. S. horserace tracks, is based on the
average of probabilities, hence called subjective, assigned to a field of horses by
the crowd of bettors. The total amount bet on each horse by all bettors is shown on
a tote board next to the horses odds to win which are based on the percentage of
the wagers on each horse to the total wagers on all horses (win pool). The method
was invented in France by Pierre Oller in 1865 and insures a profit to the track
operator by claiming a fixed percentage (track take) of the win pool before payoffs of the remainder are distributed among the bettors holding tickets on the
winning horses. An additional 1-2% loss is sustained by each winner due to the
breakage rule where pay-offs are rounded down to the next lower 5, 10 or 20
cent portion of the decimal portion of the pay-off. With 10% breakage, a pay-off
of $6.37 for a $2 bet would be rounded down to $6.30. Track take is usually 15 to
17%. Some off-shore betting sites offer lower track takes, but offer more risk as to
the safety of accounts and the reliability of pay-offs. At this time, online wagering
on horse racing is legal in the U.S., but off-shore wagering on horse races via the
internet is illegal in the U.S.
Place and Show wagers are not addressed in this work, nor are exotic bets
like Quinellas, Trifectas, Superfectas, and Daily Doubles. Exactas, which require
picking the winner and placer in perfect order, use the same probabilities as win
bets, and, like the futures market are related to common stocks, so exacta bets are
related to win bets. Quinellas and Daily Doubles are similarly related to win bets
but are not run in every race as exactas are. Like some stock market traders look
for stock trends in the futures markets, some gamblers look for hidden trends in
the exacta pools that might lead to a horse which is under bet to win (Hope, 2014).
A Chemical Engineering approach to wagering at horse races
How would chemical engineers wager at the track? One would suspect that
they would apply science and mathematics in a systematic and logical way just as
they solve engineering problems. No doubt this would involve use of a computer
and the building of a data base. What follows is the chemical engineering authors
method for wagering on horse races. The following stratagey and related
8/28/2014 JHL

Page 21

methodology was used to extend Isaacs work to an actual wagering system. It is

organized into three major tasks as follows:
1) Finding a financially safe wagering site which pays track odds with
minimum transaction costs--primarily track take. Real time odds must be
available in a digital format that can be converted to computer input. A high speed
computer, with programs that scrape the latest odds from an internet tote board
such as PHONEBET14 and quickly calculate the optimal wagers, is needed since
about 40% of all wagers are placed in the last minute before a race starts (Ziemba,
1981). Getting the most accurate odds therefore requires betting as late as
possible.
2) Technical analysis and handicapping methods must provide superior win
probabilities for the entries.
This requires large data bases and data mining
15
computer programs . William Quirin, a mathematician and university professor,
offers one of the best and most comprehensive studies of the effects of the major
variables which he weighted to scientifically determine the probabilities of horses
winning races (William L. Quirin, 1979). He used the statistical technique called
multiple regression analysis to correlate win probabilities with the predictor
variables. He also discussed money management techniques16 and concluded
through simulations that when winning 25% of the time with a 30% ROI, no more
than 8% of ones wealth should be invested in any one race to avoid tapping out.
Both flat betting (same amount for each wager) and progressive betting
(percentage of ones present wealth) were considered. Quirins system produced
an ROI of 8% over 1000 races when all races were bet.
There are speed differences between tracks depending on the type of surfaces used.
14

The web site is www.parxracing.com and is run by Philadelphia Park race track. EXCEL or parsing of the data
can be used to import the odds to a spreadsheet where a macro can perform the necessary calculations.
15
The author has a data base of over 13,000 races run at major North American tracks. For past odds information,
the Daily Racing Form (www.drf.com) provides PDF results charts of races at most North American race tracks.
Data on major predictor variables comes from The Daily Racing Form, programs printed prior to each racing day.
The Equibase Company provides detailed historical information back to the year 2000 and includes video race
replays. The web site is www.equibase.com which also offers the hobbyist a chance to own a virtual stable. The
TrackMaster Pocket Handicapper was designed for iPhones and Android Smartphones for handicapping on the
go. As advertised, Lets you go deep inside the vast amounts of exclusive ratings and statistics of each horse
race. The TVG internet wagering site at www.tvg.com has an iPhone App that permits wagering from an iPhone
at 150 major racetracks.
16
Called asset allocation by stock brokers.

8/28/2014 JHL

Page 22

West coast tracks generally have faster times than eastern tracks in the U.S., but
new synthetic surfaces like AWT, all weather turf, are now being installed which
make handicapping more interesting. There are also turf (on the grass) races and
those on earthen (dirt) or synthetic surfaces. Whether the track is dry (fast), muddy
(slow), or drying is significant. Some handicappers have listed as many as 70
different variables to be considered. The statistics involved are called nonstationary as opposed to the fixed sizes of vessels and pipes (stationary) upon
which chemical engineering calculations are usually based.
3) Mathematical models which calculate the optimal wagers, if any, for each
race use methods similar to those used for allocating resources for stock market
investments. Technical analysis of market parameters provide excellent formats
for determining the performance of various methods (Hartle, 1998) when
simulating the models on thousands of races.
Hundreds of different methods for calculating optimal wagers have been
published (Press, 2011). One of the most scientific methods was developed by
William Ziemba, an economist who calls himself Dr. Z. Dr. Z noted in his
prolific publications that his work showed that the racetrack market was efficient
for all bets except place and show (Ziemba, 1981). He defined a market as being
efficient if current security prices fully reflect all available relevant
information. If this is the case, experts should not be able to achieve higher than
average returns with regularity. He and his colleagues developed elaborate
formulas to determine the amounts that should be wagered on favorable place and
show bets which he found to be the only inefficient (bettable) wagers.
Dr. Zs optimal betting schemes feature adjustment of probabilities for the
favorite-longshot bias coupled with calculation of the expected value of a wager.
As calculated by Isaacs in Equation 3 above, Ei is the real (objective) win
probability divided by the crowd (subjective) probability. The expected value of a
wager on a horse can also be calculated as the probability of success (objective)
times the odds + 1. If this value, Ei17, is above 1.0, and the probability is correct,
17

For odds-on favorites which pay even money or $1 for every dollar bet, the objective probability found from
thousands of races is 0.53. The expected value would be calculated as 0.53 x (odds + 1) or 0.53 x 2 = 1.06, an
overlay. For a $2 wager, the bettor would receive $2 for odds of 1 to 1 plus his original bet for a total of $4.00.

8/28/2014 JHL

Page 23

then the wager will produce a profit. This wager is called an overlay. When the
Ei is below 1.0, the wager is called an underlay. Dr. Z recommends not wagering
unless the Ei is above 1.15 to provide for margin of error, especially if the win pool
is small and the odds change significantly in the last minutes of wagering. Another
key feature is Dr. Zs use of the Kelly criterion (J. L. Kelly, 1956) to finally decide
his optimal wagers. Kelly worked for Bell Labs and was trying to minimize the
amount of cable required for telephone calls. His model has been found to parallel
methods useful for calculating optimal horse race wagers.

Investing at the racetrack

As noted above, the racetrack shares many of the characteristics of the
securities markets and the tools used by market technical analysts are also useful
for horse race investing or wagering. This author agrees that all races are not
inefficient or suitable for profitable wagering. Being a third generation
handicapper and in market parlance, a technical analyst, a list of constraints was
determined from What if? analyses in thousands of simulations:
The race must have at least 10, preferably 12 entrants. More entries give
more advantage to the high speed computer over manual handicapping.18
Remember, with the pari-mutuel system, you are competing against all of the other
bettors.
There must be at least a total of $10,000 in the win pool with one minute to
post time (all horses in the starting gate). Here again, more is better as there will
be more stable odds upon which to base calculations. This is another assumption,
not mentioned by Isaacs, that all of the odds must not change significantly before
the bets are placed. We are looking to place our bets last. The triple crown
races, including the Kentucky Derby, and the Breeders Cup races have millions of
dollars in the pools hours before post time so calculations can begin earlier.19 As
18

The Jockey Club (www.jockey.com), which was incorporated in New York in 1894, is an organization dedicated to
the improvement of Thoroughbred breeding and racing. Their primary responsibility is the maintenance of the
American Stud Book, which tracks horse lineage and approves names for all American horses. They publish an
annual statistics book called The Jockey Club Fact Book which shows data through 2013 in the 2014 version. It
shows that the 2013 average field for the 48,580 races in North America was 7.85 horses. The highest ever field
was 9.07 in 1950. This shows that there will sometimes be long waits for fields of 10 or higher.
19

Jockey Club data show that the gross purses for North America in 2013 were $1,235,300,000 or an average of
$25,428 for 48,580 total races. The 2013 total pari-mutuel handle was $1,087,600,000 including worldwide

8/28/2014 JHL

Page 24

opposed to the fast traders of the stock market who seek to make their
investments millionths of a second before other traders, we are seeking to make
our investments at the last possible moment when the odds are less likely to change
significantly (Hope, 2014).
Similar to Isaacs, to be wagered upon, the expected value of the horse must
be > 1.0. Our probabilities must be more accurate than the crowds for this to hold.
This value can easily be changed to 1.15 or some other more conservative number.
Regardless of the outcome, all race wagers have limited liability, i.e. in
about two minutes one will know whether he has won or lost and one can only lose
the amount of the wager. This unequivocal outcome and an associated rate of
return within a finite time frame provide an objective benchmark to measure the
quality of an investment. In England, the winning pay-offs are called dividends.
If a horse or entry is so heavily bet that his odds go below 10 cents on a
dollar, this is called a minus pool and the track must pay a minimum of 10 cents
for each dollar wagered on this heavy favorite if he wins. There are no pay-offs for
securities whose values go to zero. Income tax is withheld using IRS form W2-G
for individual race winnings over a certain amount at the track or OTB (off-track
betting parlor). Internet accounts also provide the IRS with information on your
profits. This is another factor to consider in calculating your ROI.
Due to what is called the favorite-longshot bias the expected returns
decline as risk increases, i.e. longshots lose proportionately more than favorites do
according to odds probabilities. In the stock market, one expects additional reward
for more risk.
.

commingling of wagers on North American tracks. There was $35,080,700 in the win pool at Churchill Downs for
the $2,000,000 purse Kentucky Derby won by California Chrome on May 2, 2014. Interestingly, only about 10.9%
of 2013 bets were placed at the tracks, with the balance coming from off-track wagering. Note that this only
considers legal wagering.

8/28/2014 JHL

Page 25