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

net/publication/231610304

Predicting the outcome of roulette

Article  in  Chaos (Woodbury, N.Y.) · September 2012

DOI: 10.1063/1.4753920 · Source: PubMed

CITATIONS READS
9 6,977

2 authors:

Michael Small Chi Kong Tse

University of Western Australia The Hong Kong Polytechnic University
366 PUBLICATIONS   5,820 CITATIONS    754 PUBLICATIONS   16,337 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Prevention of Infectious diseases by public vaccination and individual protection View project

Animal motion and collective behaviour View project

All content following this page was uploaded by Chi Kong Tse on 01 June 2014.

The user has requested enhancement of the downloaded file.

Predicting the outcome of roulette
Michael Small and Chi Kong Tse

Citation: Chaos 22, 033150 (2012); doi: 10.1063/1.4753920

View online: http://dx.doi.org/10.1063/1.4753920
View Table of Contents: http://chaos.aip.org/resource/1/CHAOEH/v22/i3
Published by the American Institute of Physics.

Related Articles
Stability of discrete breathers in nonlinear Klein-Gordon type lattices with pure anharmonic couplings
J. Math. Phys. 53, 102701 (2012)
Secondary nontwist phenomena in area-preserving maps
Chaos 22, 033142 (2012)
Delay induced bifurcation of dominant transition pathways
Chaos 22, 033141 (2012)
Clocking convergence to a stable limit cycle of a periodically driven nonlinear pendulum
Chaos 22, 033138 (2012)
Characterizing the dynamics of higher dimensional nonintegrable conservative systems
Chaos 22, 033137 (2012)

Additional information on Chaos

Journal Homepage: http://chaos.aip.org/
Journal Information: http://chaos.aip.org/about/about_the_journal
Top downloads: http://chaos.aip.org/features/most_downloaded
Information for Authors: http://chaos.aip.org/authors
 CHAOS 22, 033150 (2012)

Predicting the outcome of roulette

Michael Small1,2,a) and Chi Kong Tse2
1
School of Mathematics and Statistics, The University of Western Australia, Perth, Australia
2
Department of Electronic and Information Engineering, Hong Kong Polytechnic University, Hong Kong
(Received 30 April 2012; accepted 4 September 2012; published online 26 September 2012)
There have been several popular reports of various groups exploiting the deterministic nature of the
game of roulette for profit. Moreover, through its history, the inherent determinism in the game of
roulette has attracted the attention of many luminaries of chaos theory. In this paper, we provide a
short review of that history and then set out to determine to what extent that determinism can really
be exploited for profit. To do this, we provide a very simple model for the motion of a roulette
wheel and ball and demonstrate that knowledge of initial position, velocity, and acceleration is
sufficient to predict the outcome with adequate certainty to achieve a positive expected return. We
describe two physically realizable systems to obtain this knowledge both incognito and in situ. The
first system relies only on a mechanical count of rotation of the ball and the wheel to measure the
relevant parameters. By applying these techniques to a standard casino-grade European roulette
wheel, we demonstrate an expected return of at least 18%, well above the 2.7% expected of a
random bet. With a more sophisticated, albeit more intrusive, system (mounting a digital camera
above the wheel), we demonstrate a range of systematic and statistically significant biases which
can be exploited to provide an improved guess of the outcome. Finally, our analysis demonstrates
that even a very slight slant in the roulette table leads to a very pronounced bias which
could be further exploited to substantially enhance returns. V C 2012 American Institute of Physics.

[http://dx.doi.org/10.1063/1.4753920]

“No one can possibly win at roulette unless he steals part of a vain search for perpetual motion. Alternative stories
money from the table when the croupier isn’t have attributed the origin of the game to the ancient Chinese,
looking” (Attributed to Albert Einstein in Ref. 1) a French monk or an Italian mathematician.2,4 In any case,
the device was introduced to Parisian gamblers in the mid-
Among the various gaming systems, both current eighteenth century to provide a fairer game than those cur-
and historical, roulette is uniquely deterministic. Rela- rently in circulation. By the turn of the century, the game
tively simple laws of motion allow one, in principle, to was popular and wide-spread. Its popularity bolstered by its
forecast the path of the ball on the roulette wheel and to apparent randomness and inherent (perceived) honesty.
its final destination. Perhaps because of this appealing The game of roulette consists of a heavy wheel,
deterministic nature, many notable figures from the early machined and balanced to have very low friction, and
development of chaos theory have leant their hand to designed to spin for a relatively long time with a slowly
exploiting this determinism and undermining the pre- decaying angular velocity. The wheel is spun in one direc-
sumed randomness of the outcome. In this paper, we aim
tion, while a small ball is spun in the opposite direction on
only to establish whether the determinism in this system
the rim of a fixed circularly inclined surface surrounding and
really can be profitably exploited. We find that this is def-
abutting the wheel. As the ball loses momentum, it drops to-
initely possible and propose several systems which could
be used to gain an edge over the house in a game of rou- ward the wheel and eventually will come to rest in one of 37
lette. While none of these systems are optimal, they all numbered pockets arranged around the outer edge of the
demonstrate positive expected return. spinning wheel. Various wagers can be made on which
pocket, or group of pockets, the ball will eventually fall into.
It is accepted practice that, on a successful wager on a single
pocket, the casino will pay 35 to 1. Thus, the expected return
1
I. A HISTORY OF ROULETTE from a single wager on a fair wheel is ð35 þ 1Þ  37
5
þð1Þ  2:7%. In the long-run, the house will, naturally,
The game of roulette has a long, glamorous, inglorious
win. In the eighteenth century, the game was fair and con-
history, and has been connected with several notable men of
sisted of only 36 pockets. Conversely, an American roulette
science. The origin of the game has been attributed,2 perhaps
wheel is even less fair and consists of 38 pockets. We con-
erroneously,1 to the mathematician Blaise Pascal.3 Despite
sider the European, 37 pocket, version as this is of more im-
the roulette wheel becoming a staple of probability theory,
mediate interest to us.6 Figure 1 illustrates the general
the alleged motivation for Pascal’s interest in the device was
structure, as well as the layout of pockets, on a standard
not solely to torment undergraduate students, but rather as
European roulette wheel.
Despite many proposed “systems,” there are only two
a)
Electronic mail: [email protected]. profitable ways to play roulette.7 One can either exploit

1054-1500/2012/22(3)/033150/9/$30.00 22, 033150-1 C 2012 American Institute of Physics

V
033150-2 M. Small and C. K. Tse Chaos 22, 033150 (2012)

FIG. 1. The European roulette wheel. In the

left panel, one can see a portion of the rotat-
ing roulette wheel and surrounding fixed
track. The ball has come to rest in the
(green) 0 pocket. Although the motion of
the wheel and the ball (in the outer track)
are simple and linear, one can see the addi-
tion of several metal deflectors on the stator
(that is, the fixed frame on which the rotat-
ing wheel sits). The sharp frets between
pockets also introduce strong nonlinearity as
the ball slows and bounces between pockets.
The panel on the right depicts the arrange-
ment of the number 0 to 36 and the coloring
red (lighter) and black (darker).

an unbalanced wheel, or one can exploit the inherently weakness had also been reported in Time magazine in 1951.
deterministic nature of the spin of both ball and wheel. Casi- In this case, the report described various syndicates of gam-
nos will do their utmost to avoid the first type of exploit. The blers exploiting determinism in the roulette wheel in the
second exploit is possible because placing wagers on the out- Argentinean casino Mar del Plata during 1948.13
come is traditionally permitted until some time after the ball The second type of exploit is more physical (that is,
and wheel are in motion. That is, one has an opportunity to deterministic) than purely statistical and has consequently
observe the motion of both the ball and the wheel before attracted the attention of several mathematicians, physicists
placing a wager. and engineers. One of the first14 was Henri Poincare3 in his
The archetypal tale of the first type of exploit is that of a seminal work Science and Method.15 While ruminating on
man by the name of Jagger (various sources refer to him as the nature of chance, and that a small change in initial condi-
either William Jaggers or Joseph Jagger, or some permuta- tion can lead to a large change in effect, Poincare illustrated
tion of these). Jagger, an English mechanic and amateur his thinking with the example of a roulette wheel (albeit a
mathematician, observed that slight mechanical imperfection slightly different design from the modern version). He
in a roulette wheel could afford sufficient edge to provide for observed that a tiny change in initial velocity would change
profitable play. According to one incarnation of the tale, in the final resting place of the wheel (in his model there was
1873, he embarked for the casino of Monte Carlo with six no ball) such that the wager on an either black or red (as in a
hired assistants. Once there, he carefully logged the outcome modern wheel, the black and red pockets alternate) would
of each spin of each of six roulette tables over a period of 5 correspondingly win or lose. He concluded by arguing that
weeks.8 Analysis of the data revealed that for each wheel this determinism was not important in the game of roulette
there was a unique but systematic bias. Exploiting these as the variation in initial force was tiny, and for any continu-
weaknesses, he gambled profitably for a week before the ca- ous distribution of initial velocities, the result would be the
sino management shuffled the wheels between tables. This same: effectively random, with equal probability. He was not
bought his winning streak to a sudden halt. However, he concerned with the individual pockets, and he further
soon noted various distinguishing features of the individual assumed that the variation in initial velocity required to pre-
wheels and was able to follow them between tables, again dict the outcome would be immeasurable. It is while describ-
winning consistently. Eventually, the casino resorted to ing the game of roulette that Poincare introduces the concept
redistributing the individual partitions between pockets. A of sensitivity to initial conditions, which is now a corner-
popular account, published in 1925, claims he eventually stone of modern chaos theory.16
came away with winnings of £65 000.8 The success of this A general procedure for predicting the outcome of a rou-
endeavor is one possible inspiration for the musical hall song lette spin, and an assessment of its utility was described by
“The Man Who Broke the Bank at Monte Carlo” although Edward Thorp in a 1969 publication for the Review of the
this is strongly disputed.8 International Statistical Institute.9 In that paper, Thorp
Similar feats have been repeated elsewhere. The noted describes the two basic methods of prediction. He observes
statistician Karl Pearson provided a statistical analysis of (as others have done later) that by minimizing systematic
roulette data, and found it to exhibit substantial systematic bias in the wheel, the casinos achieve a degree of mechanical
bias. However, it appears that his analysis was based on perfection that can then be exploited using deterministic pre-
flawed data from unscrupulous scribes9 (apparently he had diction schemes—efforts to minimize exploitation of statisti-
hired rather lazy journalists to collect the data). cal anomalies makes deterministic modeling methods easier.
In 1947, irregularities were found, and exploited, by two Thorp describes two deterministic prediction schemes (or
students, Albert Hibbs and Roy Walford, from Chicago rather two variants on the same scheme). If the roulette
University,10,11 Following this line of attack, Ethier provides wheel is not perfectly level (a tilt of 0:2 was apparently suf-
a statistical framework by which one can test for irregular- ficient—we verified that this is indeed more than sufficient)
ities in the observed outcome of a roulette wheel.12 A similar then there is effectively a large region of the frame from
033150-3 M. Small and C. K. Tse Chaos 22, 033150 (2012)

which the ball will not fall onto the spinning wheel. By wheel is discussed in Strzalko et al.25 Since our preliminary
studying Las Vegas wheels, he observes this condition is met publication,24 private communication with several individu-
in approximately one third of wheels. He claims that in such als indicates that these methods have now progressed to the
cases it is possible to garner a expected return of þ15%, point of at least four instances of independent in situ field
which increased to þ44% with the aid of a “pocket-sized” trials.
computer. Some time later, Thorp revealed that his collabo-
rator in this endeavor was Claude Shannon,17 the founding II. A MODEL FOR ROULETTE
father of information theory.18
In his 1967 book,2 the mathematician Richard A. We now describe our basic model of the motion of the
Epstein describes his earlier (undated) experiments with a roulette wheel and ball. Let ðr; hÞ denote the position of the
private roulette wheel. By measuring the angular velocity of ball in polar co-ordinates, and let u denote the angular posi-
the ball relative to the wheel, he was able to predict correctly tion of the wheel (say, the angular position of the centre of
the half of the wheel into which the ball would fall. Impor- the green 0 pocket). We will model the ball as a single point
tantly, he noted that the initial velocity (momentum) of the and so let rrim be the farthest radial position of that point
ball was not critical. Moreover, the problem is simply one of (i.e., the radial position of the centre of the ball when the ball
predicting when the ball will leave the outer (fixed) rim as is spinning with high velocity in the rim of the wheel). Simi-
this will always occur at a fixed velocity. However, a lack of larly, let rdefl be the radial distance to the location of the
sufficient computational resources meant that his experi- metal deflectors on the stator. For now, we will assume that
drdefl
ments were not done in real time, and certainly not attempted dh ¼ 0 (that is, there are deflectors evenly distributed
within a casino. around the stator at constant radius rdefl < r). The extension
Subsequent to, and inspired by, the work of Thorp and to the more precise case is obvious, but, as we will see, not
Shannon, another widely described attempt to beat the casi- necessary. Moreover, it is messy. Finally, we suppose that
nos of Las Vegas was made in 1977–1978 by Doyne Farmer, the incline of the stator to the horizontal is a constant a. This
Norman Packard, and colleagues.1 It is supposed that situation, together with a balance of forces is depicted in
Thorp’s 1969 paper had let the cat out of the bag regarding Figure 2. We will first consider the ideal case of a level table,
profitable betting on roulette. However, despite the assertions and then in section II B show how this condition is in fact
of Bass,1 Thorp’s paper9 is not mathematically detailed critical.
(there is in fact no equations given in the description of rou-
lette). Thorp is sufficiently detailed to leave the reader in no A. Level table
doubt that the scheme could work, but also vague enough so For a given initial motion of ball ðr; h; h; _ €hÞ
t¼0 and
that one could not replicate his effort without considerable wheel ðu; u;_ u € Þt¼0 , our aim is to determine the time tdefl at
knowledge and skill. Farmer, Packard, and colleagues imple- which r ¼ rdefl . After launch, the motion of the ball will pass
mented the system on a 6502 microprocessor hidden in a through two distinct states which we further divide into four
shoe, and proceeded to apply their method to the various cases: (i) with sufficient momentum it will remain in the rim,
casinos of the Las Vegas Strip. The exploits of this group are constrained by the fixed edge of the stator; (ii) at some point
described in detail in Bass.1 The same group of physicists the momentum drops and the ball leaves the rim; (iii) the
went on to apply their skills to the study of chaotic dynami- ball will gradually loose momentum while travelling on the
cal systems19 and also for profitable trading on the financial stator as h_ drops, so will r; and (iv) eventually r ¼ rdefl at
markets.20 In Farmer and Sidorowich’s landmark paper on some time tdefl . At time t ¼ tdefl , we assume that the ball hits
predicting chaotic time series21 the authors attribute the in- a deflector on the stator and drops onto the (still spinning)
spiration for that work to their earlier efforts to beat the wheel. Of course, the deflectors are discrete and located only
game of roulette. at specific points around the edge of the wheel. While it is
Less exalted individuals have also been employing sim- possible, and fairly straightforward to incorporate the exact
ilar schemes, in some cases fairly recently. In 2004, the position (and more importantly, the orientation) of each de-
BBC carried the report of three gamblers22 arrested by flector, we have not done this. Instead, we model the deflec-
police after winning £1 300 000 at the Ritz Casino in Lon- tors at a constant radial distance around the entire rim. The
don. The trio had apparently been using a laser scanner and exact position of the wheel when the ball reaches the deflec-
their mobile phones to predict the likely resting place of the tors will be random but will depend only on uðtdefl Þ—i.e.,
ball. Happily, for the trio but not the casino, they were depending on where the actual deflectors are when the ball
judged to have broken no laws and allowed to keep their first comes within range, the radial distance until the ball
winnings.23 The scheme we describe in Sec. II and imple- actually deflects will be uniformly distributed on the interval
ment in Sec. III is certainly compatible with the equipment ½0; 2p=Ndefl , where Ndefl is the number of deflectors.
and results reported in this case. In Sec. IV, we conclude
with some remarks concerning the practicality of applying
1. Ball rotates in the rim
these methods in a modern casino, and what steps casinos
could take (or perhaps have taken) to circumvent these While traveling in the rim r is constant and the ball has
exploits. A preliminary version of these results was pre- _ Hence, the radial acceleration of the ball
angular velocity h.
2
sented at a conference in Macau.24 An independent and is ac ¼ r ¼ r ðrhÞ2 ¼ rh_ , where v is the speed of the ball.
v2 1 _
much more detailed model of dynamics of the roulette During this period of motion, we suppose that r is constant
033150-4 M. Small and C. K. Tse Chaos 22, 033150 (2012)

FIG. 2. The dynamic model of ball and wheel. On the

left, we show a top view of the roulette wheel (shaded
region) and the stator (outer circles). The ball is mov-
ing on the stator with instantaneous position ðr; hÞ
while the wheel is rotating with angular velocity u_
(note that the direction of the arrows here are for illus-
tration only, the analysis in the text assume the same
convention, clockwise positive, for both ball and
wheel). The deflectors on the stator are modelled as a
circle, concentric with the wheel, of radius rdefl . On
the right, we show a cross section and examination of
the forces acting on the ball in the incline plane of the
stator. The angle a is the incline of the stator, m is
the mass of the ball, ac is the radial acceleration of
the ball, and g is gravity.

and that h decays only due to constant rolling friction: hence eventually reaches the various deflectors at r ¼ rdefl . The
r_ ¼ 0 and €
h¼€ hð0Þ, a constant. This phase of motion will angular velocity continues to be governed by
continue provided the centripetal force of the ball on the rim
exceeds the force of gravity mac cos a > mgsin a (m is the _ ¼ hð0Þ
hðtÞ _ þ €hð0Þt;
mass of the ball). Hence, at this stage
but now that
2 g
h_ > tan a: (1) 2
r rh_ < g tan a

the radial position is going to gradually decrease too. The

2. Ball leaves the rim
difference between the force of gravity mgsina and the
2
2
Gradually the speed on the ball decays until eventually (lesser) centripetal force mr h_ cos a provides inward acceler-
h_ ¼ gr tana. Given the initial acceleration €hð0Þ, velocity ation of the ball
_
hð0Þ, and position hð0Þ, it is trivial to compute the time at
2
which the ball leaves the rim, trim to be r€ ¼ rh_ cos a  g sin a: (3)
 rffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 _ g Integrating Eq. (3) yields the position of the ball on the
trim ¼  hð0Þ  tan a : (2)
€hð0Þ r stator.

To do so, we assume that the angular acceleration is constant 4. Ball reaches the deflectors
and so the angular velocity at any time is given by hðtÞ _
_
¼ hð0Þ þ€ hð0Þt and substitute into Eq. (1). That is, we are Finally, we find the time t ¼ tdefl for which r(t), computed
assuming that the force acting on the ball is independent of as the definite second integral of Eq. (3), is equal to rdefl . We
velocity—this is a simplifying assumption for the naive can then compute the instantaneous angular position of the ball
model we describe here, more sophisticated alternatives are _
hðtdefl Þ ¼ hð0Þ þ hð0Þt 1€ 2
defl þ 2 hð0Þtdefl and the wheel uðtdefl Þ
1€
possible, but in all cases this will involve the estimation of ¼ uð0Þ þ uð0Þt
_ defl þ 2 u ð0Þt2defl to give the salient value
additional parameters. The position at which the ball leaves
the rim is given by c ¼ jhðtdefl Þ  uðtdefl Þj2p (4)
 
 _ 2 
ðgr tan aÞ  hð0Þ denoting the angular location on the wheel directly below

hð0Þ þ  the point at which the ball strikes a deflector. Assuming the
 2€hð0Þ 
2p constant distribution of deflectors around the rim, some (still
to be estimated) distribution of resting place of the ball will
where j  j2p denotes modulo 2p.
depend only on that value c. Note that, although we have
_ €hÞ and ðu; u;
described ðh; h; € Þt¼0 separately, it is possi-
_ u
t¼0
3. Ball rotates freely on the stator
ble to adopt the rotating frame of reference of the wheel and
After leaving the rim, the ball will continue (in practice, treat h  u as a single variable. The analysis is equivalent,
for only a short while) to rotate freely on the stator until it estimating the required parameters may become simpler.
033150-5 M. Small and C. K. Tse Chaos 22, 033150 (2012)

We note that for a level table, each spin of the ball alters
only the time spent in the rim, the ball will leave the rim of
the stator with exactly the same velocity h_ each time. The
descent from this point to the deflectors will therefore be
identical. There will, in fact, be some characteristic duration
which could be easily computed for a given table. Doing this
would circumvent the need to integrate Eq. (3).

B. The crooked table

Suppose, now that the table is not perfectly level. This is
the situation discussed and exploited by Thorp.9 Without
loss of generality (it is only an affine change of co-ordinates
for any other orientation) suppose that the table is tilted by
an angle d such that the origin u ¼ 0 is the lowest point on
the rim. Just as with the case of a level table, the time which
the ball spends in the rim is variable and the time at which it
leaves the rim depends on a stability criterion similar to
Eq. (1). But now that the table is not level, that equilibrium
FIG. 3. The case of the crooked table. The blue curve denotes the stability
becomes criterion (6), while the red solid line is the (approximate) trajectory of the ball
2
with hðt1 Þ þ 2p ¼ hðt2 Þ indicating two successive times of complete revolu-
rh_ ¼ g tanða þ d cos hÞ: (5) tions. The point at which the ball leaves the rim will therefore be the first
intersection of this stability criterion and the trajectory. This will necessarily
be in the region to the left of the point at which the ball’s trajectory is tangent
If d ¼ 0 then it is clear that the distribution of angular posi- to Eq. (6), and this is highlighted in the figure as a green solid. Typically a
tions for which this condition is first met will be uniform. crooked table will only be slightly crooked and hence this region will be close
Suppose instead that d > 0, then there is now a range of criti- to h ¼ 0 but biased toward the approaching ball. The width of that region
2 depends on hðt _ 2 Þ, which in turn can be determined from Eq. (6).
_ 1 Þ  hðt
cal angular velocities h_ 2 ½g tanða  dÞ; g tanða þ dÞ.
crit r r
2
Once h_ < gr tanða þ dÞ, the position at which the ball leaves follows a uniform distribution over significantly less than
_
the rim will be dictated by the point of intersection in ðh; hÞ- half the wheel circumference. In this situation, the problem
space of of predicting the final resting place is significantly simplified
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi to the problem of predicting the position of the wheel at the
_h ¼ g tanða þ d cos hÞ (6)
time the ball leaves the rim.
r We will pursue this particular case no further here. The
situation (5) may be considered as a generalisation of the
and the ball trajectory as a function of t (modulo 2p) ideal d ¼ 0 case. This generalisation makes the task of pre-
diction significantly easier, but we will continue to work
_ ¼ hð0Þ
hðtÞ _ þ €hð0Þt; (7) under the assumption that the casino will be doing its utmost
to avoid the problems of an improperly levelled wheel.
_ 1
hðtÞ ¼ hð0Þ þ hð0Þt þ €hð0Þt2 : (8) Moreover, this generalisation is messy, but otherwise unin-
2 teresting. In Sec. III, we consider the problem of implement-
If the angular velocity of the ball is large enough, then the ing a prediction scheme for a perfectly level wheel.
ball will leave the rim at some point on the half circle prior
to the low point (u ¼ 0). Moreover, suppose that in one rev- III. EXPERIMENTAL RESULTS
olution (i.e., hðt1 Þ þ 2p ¼ hðt2 Þ), the velocity changes by
In Sec. II, we introduced the basic mathematical model
_ 1 Þ  hðt
hðt _ 2 Þ. Furthermore, suppose that this is the first
which we utilize for the prediction of the trajectory of the
2
revolution during which h_ < gr tanða þ dÞ (that is, hðt_ 1 Þ2 ball within the roulette wheel. We ignore (or rather treat as
 gr tanða þ dÞ but hðt_ 2 Þ < g tanða þ dÞ). Then, the point
2 essentially stochastic) the trajectory of the ball after hitting
r
_ the deflectors—charting the distribution of final outcome
at which the ball will leave the rim will (in ðh; hÞ-space)
from deflector to individual pocket in the roulette wheel is a
be the intersection of Eq. (6) and
tractable probabilistic problem, and one for which we will
  sketch a solution later. However, the details are perhaps only
_ 1 Þ  1 hðt
h_ ¼ hðt _ 2 Þ  hðt
_ 1 Þ h: (9)
2p of interest to the professional gambler and not to most physi-
cists. Hence, we are reduced to predicting the location of the
The situation is depicted in Figure 3. One can expect for a wheel and the ball when the ball first reaches one of the
tilted roulette wheel, the ball will systematically favor leav- deflectors. The model described in Sec. II is sufficient to
ing the rim on one half of the wheel. Moreover, to a good achieve this—provided one has adequate measurements of
approximation, the point at which the ball will leave the rim the physical dimensions of the wheel and all initial positions,
033150-6 M. Small and C. K. Tse Chaos 22, 033150 (2012)

velocities, and accelerations (as a further approximation we

assume deceleration of both the ball and wheel to be constant
over the interval which we predict).
Hence, the problem of prediction is essentially two-
fold. First, the various velocities must be estimated accu-
rately. Given these estimates, it is a trivial problem to then
determine the point at which the ball will intersect with one
of the deflectors on the stator. Second, one must then have
an estimate of the scatter imposed on the ball by both the
deflectors and possible collision with the individual frets. To
apply this method in situ, one has the further complication
of estimating the parameters r, rdefl ; rrim ; a, and possibly d
without attracting undue attention. We will ignore this addi-
tional complication as it is essentially a problem of data col-
lection and statistical estimation. Rather, we will assume
that these quantities can be reliably estimated and restrict
our attention to the problem of prediction of the motion.
To estimate the relevant positions, velocities, and accelera- FIG. 4. Hand-measurement of ball and wheel velocity for prediction. From
tions ðh; h;_ €h; u; u;
_ u€ Þt¼0 (or perhaps just ðh  u; h_  u;
_ two spins of the wheel, and 20 successive spins of the ball we logged the
€ time (in seconds) T(i) for successive passes past a given point (T(i) against
hu € Þt¼0 ), we employ two distinct techniques. T(i þ 1)). The measurements T(i) and T(i þ 1) are the timings of successive
In Secs. III A–III C, we describe these methods. In revolutions—direct measurements of the angular velocity observed over one
Sec. III A, we introduce a manual measurement scheme, and complete rotation. To provide the simplest and most direct indication that
handheld measurements of this quantity are accurate, we indicate in this fig-
in Sec. III B, we describe our implementation of a more so-
ure a deterministic relationship between these quantities. From this relation-
phisticated digital system. The purpose of Sec. III A is to ship, one can determine the angular deceleration. The red (slightly higher)
demonstrate that a rather simple “clicker” type of device— points depict these times for the wheel, the blue (lower) points are for the
along the lines of that utilized by the Doyne Farmer, Norman ball. A single trial of both ball and wheel is randomly highlighted with
crosses (superimposed). The inset is an enlargement of the detail in the lower
Packard, and collaborators1—can be employed to make suf- left corner. Both the noise and the determinism of this method are evident.
ficiently accurate measurements. Nonetheless, this system is In particular, the wheel velocity is relatively easy to calculate and decays
far from optimal: we conduct only limited experiments with slowly, in contrast the ball decays faster and is more difficult to measure.
this apparatus: sufficient to demonstrate that, in principle,
the method could work. In Sec. III B, we describe a more so-
phone.26–28 Our results, depicted in Fig. 4 illustrate that the
phisticated system. This system relies on a digital camera
measurements, although noisy, are feasible. The noise intro-
mounted directly above a roulette wheel and is therefore
duced in this measurement is probably largely due to the lack
unlikely to be employed in practice (although alternative,
of physical hand-eye co-ordination of the first author. Figure
more subtle, devices could be imagined). Nonetheless, our
4 serves only to demonstrate that, from measurements of suc-
aim here is to demonstrate how well this system could work
cessive revolutions T(i) and T(i þ 1) the relationship between
in an optimal environment.
T(i) and T(i þ 1) can be predicted with a fairly high degree of
Of course, the degree to which the model in Sec. II is
certainty (over several trials and with several different initial
able to provide a useful prediction will depend critically on
conditions). As expected, the dependence of T(i þ 1) on T(i)
how well the parameters are estimated. Sensitivity analysis
is sub-linear. Hence, derivation of velocity and acceleration
shows that the predicted outcome (Eq. (4)) depends only lin-
from these measurements should be relatively straightfor-
early or quadratically (in the case of physical parameters of
ward. Simple experiments with this configuration indicate
the wheel) on our initial estimates. More important however,
that it is possible to accurately predict the correct half of the
and more difficult to estimate, is to what extent each of these
wheel in which the ball will come to rest.
parameters can be reliably estimated. For this reason, we first
Using these (admittedly noisy) measurements, we were
take a strictly experimental approach and show that even
able to successfully predict the half of the wheel in which the
with the various imperfections inherent in experimental mea-
ball would stop in 13 of 22 trials (random with p < 0:15),
surement, and in our model, sufficiently accurate predictions
yielding an expected return of 36=18  13=22  1 ¼ þ18%.
are realizable. Later, in Sec. III C, we provide a brief compu-
This trial run included predicting the precise location in
tational analysis of how model prediction will be affected by
which the ball landed on three occasions (random with
uncertainty in each of the parameters.
p < 0:02). Quoted p-values are computed against the null hy-
pothesis of a binomial distribution of n trials with probability
A. A manual implementation
of success determined by the fraction of the total circumfer-
Our first approach is to simply record the time at which ence corresponding to the target range—i.e., the probability p
ball and wheel pass a fixed point. This is a simple approach of landing by chance in one of the target pockets
(probably that used in the early attempts to beat the wheels of
Las Vegas) and is trivial to implement on a laptop computer, number of target pockets
p¼ :
personal digital assistant, embedded system, or even a mobile 37
033150-7 M. Small and C. K. Tse Chaos 22, 033150 (2012)

B. Automated digital image capture

Alternatively, we employ a digital camera mounted

directly above the wheel to accurately and instantaneously
measure the various physical parameters. This second
approach is obviously a little more difficult to implement in-
cognito. Here, we are more interested in determining how
much of an edge can be achieved under ideal conditions,
rather than the various implementation issues associated
with realizing this scheme for personal gain. In all our trials,
we use a regulation casino-grade roulette wheel (a 32”
“President Revolution” roulette wheel manufactured by Mat-
sui Gaming Machine Co. Ltd., Tokyo). The wheel has 37
numbered slots (1 to 36 and 0) in the configuration shown in
Figure 1 and has a radius of 820 mm (spindle to rim). For the
purposes of data collection, we employ a Prosilica EC650C
IEEE-1394 digital camera (1/3” CCD, 659  493 pixels at
90 frames per second). Data collection software was written
and coded in Cþþ using the OpenCV library.
The camera provides approximately (slightly less due to
issues with data transfer) 90 images per second of the posi-
tion of the roulette wheel and the ball. Artifacts in the image
FIG. 5. Predicting roulette. The plot depicts the results of 700 trials of our
due to lighting had to be managed and filtered. From the re-
automated image recognition software used to predict the outcome of inde-
sultant image, the position of the wheel was easily deter- pendent spins of a roulette wheel. What we plot here is a histogram in polar
mined by locating the only green pocket (“0”) in the wheel, coordinates of the difference between the predicted and the actual outcome
and the position of the ball was located by differencing suc- (the “Target” location, at the 12 o’clock position in this diagram, indicating
that the prediction was correct). The length of each of the 37 black bars
cessive frames (searching for the ball shape or color was not denote the frequency with which predicted and actual outcome differed by
sufficient due to the reflective surface of the wheel and ambi- exactly the corresponding angle. Dotted, dotted-dashed, and solid (red) lines
ent lighting conditions). depict the corresponding 99.9%, 99%, and 90% confidence intervals using
From these time series of Cartesian coordinates for the the corresponding two-tailed binomial distribution. Motion forward (i.e.,
ball continues to move in the same direction) is clockwise, motion back-
position of both the wheel (green “0” pocket) and ball, we wards is anti-clockwise. From the 37 possible results, there are 2 instances
computed the centre of rotation and hence derived angular outside the 99% confidence interval. There are 7 instances outside the 90%
position time series. Polynomial fits to these angular position confidence interval.
data (modulo 2p) provided estimates of angular velocity and
acceleration (deceleration). From this data, we found that, higher than and less than (respectively) that expected by
for out apparatus, the acceleration terms where very close to chance: outside the 99% confidence interval. Hence, the pre-
being constant over the observation time period—and hence dicted target pocket is a good indicator of eventual outcome
modeling the forces acting on the ball as constant provided a and those pockets immediate prior to the target pocket
reasonable approximation. With these parameters, we (which the ball would need to bounce backwards to reach)
directly applied the model of Sec. II to predict the point at are less likely. Finally, and rather speculatively, there is a
which the ball came into contact with the deflectors. relatively higher chance (although marginally significant) of
Figure 5 illustrates the results from 700 trials of the pre- the ball landing in one of the subsequent pockets—hence,
diction algorithm on independent rolls of a fair and level rou- suggesting that the best strategy may be to bet on the section
lette wheel. The scatter plot of Fig. 5 provides only a crude of the wheel following the actual predicted destination.
estimation of variance over the entire region of the wheel for
a given prediction. A determined gambler could certainly
C. Parameter uncertainity and measurement noise
extend this analysis with a more substantial data set relating
to her particular wheel of interest. We only aim to show that The performance of the model described above will
certain non-random characteristics in the distribution of rest- depend on the accuracy of estimates of each of the model pa-
ing place will emerge and that these can then be used to fur- rameters: a (the inclination of the wheel rim), rrim (the radius
ther refine prediction. of the wheel rim), and rdefl (the location of the deflectors).
Nonetheless, several things are clear from Fig. 5. First, Using the data collected in Sec. III B, we systematically vary
for most of the wheel, the probability of the ball landing in a each of the parameters by a factor  ð0:9 <  < 1:1Þ so that,
particular pocket—relative to the predicted destination— for example, we estimate the ball position with an inclination
does not differ significantly from chance: observed popula- of a. Figure 6(a) depicts the results, and, as expected, de-
tions in 30 of 37 pockets is within the 90% confidence inter- pendence on each of these parameters is linear (and negative
val for a random process. Two particular pockets—the target in the case of rrim ). Moreover, the location of the deflectors
pocket itself and a pocket approximately one-quarter of the rdefl is not critical, whereas the correct estimation of the other
wheel prior to the target pocket—occur with frequencies two parameters is. Nonetheless, this would not pose a
033150-8 M. Small and C. K. Tse Chaos 22, 033150 (2012)

FIG. 6. Parameter uncertainty. We explore the

effect of error in the model parameters on the out-
come by varying the three physical parameters of
the wheel (a) and introducing uncertainty in the
measurement of timing events used to obtain
estimates of velocity and deceleration (b). In (a), we
depict the effect of perturbing the estimated
values of a (green—affine, increasing steeply) rrim
(red—affine, decreasing) and rdefl (blue—affine and
increasing slowly) from 90% to 110% of the true
value. In (b), we add Gaussian noise of magnitude
between 0.5% and 10% the variance of the true
measurements to initial estimates of all positions
and velocities. Horizontal dotted lines in both plots
depict error corresponding to one whole pocket in
the wheel. The vertical axis is in radians and covers
6 p2—half the wheel. In the upper panel, least varia-
tion in outcome is observed with errors in estima-
tion of rdefl .

significant problem for prediction as in all case the variation We would like to draw two simple conclusions from this
in these parameters introduces a systematic bias which could work. First, deterministic predictions of the outcome of a
easily be corrected for, or even used to estimate the true game of roulette can be made, and can probably be done
value. in situ. Hence, the tales of various exploits in this arena are
What is more striking is the effect of measurement noise likely to be based on fact. Second, the margin for profit is
depicted in Fig. 6(b). We add Gaussian noise to each timing quite slim. Minor manipulation with the frictional resistance
measurement (each frame, recording at 90 frames per sec- or level of the wheel and/or the manner in which the croupier
ond) over the duration of the observation period (25 frames) actually plays the ball (the force with which the ball is rolled
used to estimate initial velocity and deceleration of the ball and the effect, for example, of axial spin of the ball) have
and velocity of the wheel. The added noise has an effect of not been explored here and would likely affect the results
increasing the variation in the predicted resting place of the significantly. Hence, for the casino the news is mostly
ball (since the noise is unbiased) and the strength of this good—minor adjustments will ameliorate the advantage of
effect is linear with the level of noise. As independent noise the physicist-gambler. For the gambler, one can rest assured
realizations are added to 50 measurements (25 each for the that the game is on some level predictable and therefore
ball and wheel), this is a substantial amount of error—even inherently honest.
at a fairly low amplitude. Nonetheless, the final results are Of course, the model we have used here is extremely
still within 2–3 pockets of the original prediction for noise of simple. In Strzalko et al.,25 much more sophisticated model-
up to 2% on every scalar observation. ing methodologies have been independently developed and
presented. Certainly, since the entire system is a physical dy-
IV. EXPLOITS AND COUNTER-MEASURES namical system, computational modeling of the entire system
may provide an even greater advantage.25 Nonetheless, the
The essence of the method presented here is to predict
methods presented in this paper would certainly be within
the location of the ball and wheel at the point when the ball
the capabilities of a 1970s “shoe-computer.”
will first come into contact with the deflectors. Hence, we
only require knowledge of initial conditions of each aspect
of the system (or more concisely, their relative positions,
ACKNOWLEDGMENTS
velocities and accelerations). In addition to this, certain pa-
rameters derived from the physical dimensions of the wheel The first author would like to thank Marius Gerber for
are required—these could either be estimated directly, or introducing him to the dynamical systems aspects of the
inferred from observational trajectory data. Finally, we note game of roulette. Funding for this project, including the rou-
that while anecdotal evidence suggests that (the height of lette wheel, was provided by the Hong Kong Polytechnic
the) frets plays an important role in the final resting place of University. The labors of final year project students, Yung
the ball, this does not enter into our model of the more deter- Chun Ting and Chung Kin Shing, in performing many of the
ministic phase of the system dynamics. It will affect the dis- mechanical simulations describe herein are gratefully
tribution of final resting places—and hence this is going to acknowledged. M.S. is supported by an Australian Research
depend rather sensitively on a particular wheel. Council Future Fellowship (FT110100896).
033150-9 M. Small and C. K. Tse Chaos 22, 033150 (2012)

1 16
T. A. Bass, The Newtonian Casino (Penguin, London, 1990). J. P. Crutchfield, J. Doyne Farmer, N. H. Packard, and R. S. Shaw,
2
R. A. Epstein, The Theory of Gambling and Statistical Logic (Academic, “Chaos,” Sci. Am. 255, 46–57 (1986).
17
New York, 1967). E. O. Thorp, The Mathematics of Gambling (Gambling Times, 1985).
3 18
E. T. Bell, Men of Mathematics (Simon and Schuster, New York, 1937). C. E. Shannon, “A mathematical theory of communication,” Bell Syst.
4
The Italian mathematician, confusingly, was named Don Pasquale2, a sur- Techn. J. 27, 379–423, 623–656 (1948).
19
name phonetically similar to Pascal. Moreover, as Don Pasquale is also N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, “Geometry
the name of a 19th century opera buff, this attribution is possibly fanciful. from a time series,” Phys. Rev. Lett. 45, 712–716 (1980).
5 20
F. Downton and R. L. Holder, “Banker’s games and the gambling act T. A. Bass, The Predictors, edited by A. Lane (Penguin, London, 1999).
21
1968,” J. R. Stat. Soc. Ser. A 135, 336–364 (1972). J. D. Farmer and J. J. Sidorowich, “Predicting chaotic time series,” Phys.
6
B. Okuley and F. King-Poole, Gamblers Guide to Macao (South China Rev. Lett. 59, 845–848 (1987).
22
Morning Post, Hong Kong, 1979). BBC Online, “Arrests follow £1m roulette win,” BBC News March 22,
7
Three, if one has sufficient finances to assume the role of the house. 2004.
8 23
C. Kingston, The Romance of Monte Carlo (John Lane The Bodley Head BBC Online, “Laser scam” gamblers to keep £1m,” BBC News December
Ltd., London, 1925). 5, 2004.
9 24
E. O. Thorp, “Optimal gambling systems for favorable games,” Rev. Int. M. Small and C. K. Tse, “Feasible implementation of a prediction algo-
Stat. Inst. 37, 273–293 (1969). rithm for the game of roulette,” in Asia-Pacific Conference on Circuits
10
Life Magazine Publication, “How to Win $6,500,” Life, 46, December 8, and Systems (IEEE, 2008).
25
1947. J. Strzalko, J. Grabski, P. Perlikowksi, A. Stefanski, and T. Kapitaniak,
11
Alternatively, and apparently erroneously, reported to be from Californian Dynamics of gambling, Vol. 792 of Lecture Notes in Physics (Springer,
Institute of Technology in Ref. 2. 2009).
12 26
S. N. Ethier, “Testing for favorable numbers on a roulette wheel,” J. Am. Implementation on a “shoe-computer” should be relatively straightforward
Stat. Assoc. 77, 660–665 (1982). too.
13 27
Time Magazine Publication, “Argentina—Bank breakers,” Time 135, 34, C. T. Yung, “Predicting roulette,” Final Year Project Report, Hong Kong
February 12, 1951. Polytechnic University, Department of Electronic and Information Engi-
14
The first, to the best of our knowledge. neering, April 2011.
15 28
H. Poincare, Science and Method (Nelson, London, 1914). English transla- K. S. Chung, “Predicting roulette II: Implementation,” Final Year Project
tion by Francis Maitland, preface by Bertrand Russell. Facsimile reprint in Report, Hong Kong Polytechnic University, Department of Electronic and
1996 by Routledge/Thoemmes, London. Information Engineering, April 2010.

View publication stats