Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

download Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

of 18

Transcript of Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    1/18

    O R I G I N A L P A P E R

    Using Neural Networks to Model the Behavior

    and Decisions of Gamblers, in Particular,Cyber-Gamblers

    Victor K. Y. Chan

    Springer Science+Business Media, LLC 2009

    Abstract This article describes the use of neural networks (a type of artificial intelligence)

    and an empirical data sample of, inter alia, the amounts of bets laid and the winnings/losses

    made in successive games by a number of cyber-gamblers to longitudinally model

    gamblers behavior and decisions as to such bet amounts and the temporal trajectory of

    winnings/losses. The data was collected by videoing Texas Holdem gamblers at a cyber-

    gambling website. Six persistent gamblers were identified, totaling 675 games. The

    neural networks on average were able to predict bet amounts and cumulative winnings/losses in successive games accurately to three decimal places of the dollar. A more

    important conclusion is that the influence of a gamblers skills, strategies, and personality on

    his/her successive bet amounts and cumulative winnings/losses is almost totally reflected by

    the pattern(s) of his/her winnings/losses in the few initial games and his/her gambling

    account balance. This partially invalidates gamblers illusions and fallacies that they can

    outperform others or even bankers. For government policy-makers, gambling industry

    operators, economists, sociologists, psychiatrists, and psychologists, this article provides

    models for gamblers behavior and decisions. It also explores and exemplifies the usefulness

    of neural networks and artificial intelligence at large in the research on gambling.

    Keywords Neural network Gamblers/Cyber-gamblers behavior and decisions Successive bet amounts Cumulative winnings/losses Gamblers fallacies

    Introduction

    Gambling, or gaming as it is also known, is a human activity that can be traced back

    through prehistory and a multitude of gambling forms are apparent in many different

    cultures (Dewar 2001). Gambling is often claimed to have positive effects, in that not onlyis it a lucrative and profitable industry worldwide in its own right, worth billions of US

    V. K. Y. Chan (&)

    School of Business, Macao Polytechnic Institute, Rua de Luis Gonzaga Gomes, Macao,

    Peoples Republic of China

    e-mail: [email protected]

    123

    J Gambl Stud

    DOI 10.1007/s10899-009-9139-7

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    2/18

    dollars per annum (American Gaming Association 2008b; Moss et al. 2003), but it also

    fuels tourism (MacLaurin and MacLaurin 2003), employs a considerable amount of labor

    at different skill levels (Gill et al. 2006; MacLaurin and MacLaurin 2003; Siegel and

    Anders 1999), and is a major source of government revenue in many economies (Gill et al.

    2006; Layton and Worthington 1999; Siegel and Anders 1999). Unfortunately, gambling iscoupled with a series of negative socio-economic problems. For example, it financially

    afflicts low-income individuals (Layton and Worthington 1999). Furthermore, there is

    evidence that certain forms of gambling create pathological gambling (alternatively known

    as compulsive gambling, addictive gambling, disordered gambling, or problem gambling)

    through addiction, lead to suicide, attract criminal elements, foster corruption (Feigelman

    et al. 2006; Layton and Worthington 1999; Friedman et al. 1989) and facilitate money

    laundering (Hugel and Kelly 2002).

    The total impact and incidence of both the positive contributions and the negative

    problems have been intensified by the advent of cyber-gambling on the internet in the mid-

    1990s, given that cyber-gambling makes gambling substantially more prevalent and per-

    vasive in society by essentially relaxing the constraints on gambling times and venues (Gill

    et al. 2006). In particular, with regard to alleviating the related problems, whilst traditional

    terrestrial gambling is subject to strict legal and governmental regulations enforced within

    the associated jurisdictional boundaries, cyber-gambling renders these restrictions rela-

    tively unenforceable because of the absence of boundaries in cyberspace. In summary,

    cyber-gambling scales up both the contribution and the problems of gambling. As a matter

    of fact, cyber-gambling has been growing since its inception in 1995 (American Gaming

    Association 2008a; Dewar 2001) and the magnitude of its annual turnover worldwide is

    now tens of billions of US dollars (American Gaming Association 2008a; Dewar 2001)with nearly 23 million people having gambled at well over 2,000 cyber-gambling websites

    as of 2005 (American Gaming Association 2008a).

    Given such a conjunction, namely that the gambling industry is already a multi-billion

    dollar industry and still rapidly expanding, in particular, as a result of the speedy prolif-

    eration of cyber-gambling websites and cyber-gamblers in the last decade, how to strike a

    balance between gamblings contributions and problems has been undeniably controversial

    both socio-economically and politically (Smith 2004; Layton and Worthington 1999).

    Needless to say, both the contributions and the problems of gambling depend on gamblers

    behavior and decisions, in particular, their bet amounts and their temporal trajectory of

    cumulative winnings/losses. Specifically, on the one hand, gamblings contributions to thegambling industrys profit, the industrys employment and government revenue are directly

    derived from gamblers bets and losses. On the other hand, gamblings problems with

    regard to the gamblers financial affliction, pathological gambling, and possible suicide

    stem outright from the gamblers cumulative losses. Therefore, gamblers bet amounts and

    their temporal trajectory of cumulative winnings/losses become one of the determinants of

    the aforementioned gambling-related controversy. In the face of this controversy, models

    for gamblers bet amounts and their temporal trajectory of cumulative winnings/losses can

    assist the work of related government policy-makers, gambling industry operators, econ-

    omists, sociologists, psychiatrists, and psychologists. In particular, such models helpgovernment policy-makers in reformulating related government policies and putting for-

    ward new legislation to regulate the gambling industry in a bid to secure sufficient gov-

    ernment revenue while at the same time containing gambling-related problems (American

    Gaming Association 2008a; Smith 2004; Hugel and Kelly 2002; Dewar 2001; Jones et al.

    2000; Siegel and Anders 1999). Such models also enable gambling industry operators and

    their employees to reposition and redesign their games so as to optimize the profit and thus

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    3/18

    guarantee employment (Moss et al. 2003). At the same time, economists and sociologists

    can better estimate the socio-economic value of the gaming industry based on the infer-

    ences provided by such models regarding bet amounts and cumulative winnings/losses,

    whereas psychiatrists and psychologists can better chart the behavior and decisions of

    gamblers according to such models so as to improve the treatment and rehabilitation ofpathological gamblers (Ladouceur et al. 2007; Gill et al. 2006; Nelson et al. 2006; Slutske

    et al. 2003, 2005; Breen and Zimmerman 2002; Layton and Worthington 1999).

    This article documents research on aforesaid models based on neural networks or, more

    exactly, artificial neural networks, which are a type of artificial intelligence emulating

    biological neural networks (Nikolaev and Iba 2006; Ananda Rao and Srinivas 2003; Arbib

    2003), used to model the aforementioned bet amounts and temporal trajectories of

    cumulative winnings/losses. Such neural networks were and needed to be trained by an

    empirical data sample on the amounts of bets laid and the winnings/losses made in suc-

    cessive games by each of a random sample of cyber-gamblers. The data sample was

    collected from a legally licensed, well-established cyber-gambling website (sometimes

    referred to as an e-casino) where international gamblers played Texas Holdem among other

    poker games. This research aimed at letting the neural networks construct the following

    two models:

    M1: the model for successive bet amounts, which longitudinally models and thus

    predicts the bet amounts in the successive game laid by each individual Texas Holdem

    gambler based on his/her winnings/losses in a number of immediately preceding games

    and his/her current gambling account balance.

    M2: the model for the temporal trajectory of cumulative winnings/losses, which

    longitudinally models the temporal trajectory of cumulative winnings/losses of eachindividual Texas Holdem gambler based on his/her cumulative winnings/losses in a

    number of immediately preceding games.

    Despite the fact that Texas Holdem games are more than games of chance in that the

    outcomes depend on gamblers intellectual decisions, which, believe it or not, in turn are

    generally thought to depend on individual gamblers skills and even strategies (Harrington

    and Robertie 2004), it was to the authors surprise that the neural network for M1 upon

    training turned out to be able to predict a gamblers bet amounts in successive games

    accurately to more than three decimal places of the dollar on average for each of the six

    gamblers in our data sample across the board. More importantly, the neural network for M2upon training was also able to track the temporal trajectory of a gamblers cumulative

    winnings/losses, i.e., successively predict the gamblers cumulative winnings/losses, with a

    similar accuracy again for each of the six gamblers in our data sample across the board.

    Undoubtedly, considering the structure of the two neural networks, which will be elabo-

    rated on later in the article, the above findings empirically strongly suggest that the

    influence of a gamblers skills, strategies, and personality on his/her successive bet

    amounts and cumulative winnings/losses is almost totally reflected by the pattern(s) of his/

    her winnings/losses in the first few initial games and his/her gambling account balance.

    These conclusions at least partially invalidate gamblers various illusions and fallacies

    (Clotfelter and Cook 1993), as detailed in the subsection Economic/Psychological

    Models below, that they can outperform other gamblers or even bankers (e.g., casinos,

    cyber-gambling websites, etc.) For gamblers, the implication is that it is not worthwhile to

    excessively and endlessly experiment with their skills, strategies, and personalities in an

    attempt to outperform other gamblers or even bankers. For casinos and cyber-gambling

    websites, the implications are that their profits can be better maximized through

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    4/18

    emphasizing the provision of particular games and redesigning their games based on the

    different predicted temporal trajectories of typical gamblers cumulative winnings/losses in

    different games and/or games of different designs.

    In addition, from a broader point of view, such neural networks can serve related

    government policy-makers, gambling industry operators, economists, sociologists, psy-chiatrists, and psychologists in the manners discussed above. Such application of neural

    networks also explores and exemplifies the usefulness of neural networks and artificial

    intelligence at large (Hu et al. 2007; Kaklauskas and Zavadskas 2007; Schocken and Ariav

    1994), apart from traditional parametric statistics, in the context of gambling research and

    socio-economically and politically controversial issues in general.

    The remainder of this article proceeds as follows. The section Literature discusses the

    literature relevant to this articles research. The section Methodology develops the

    neural networks underlying this research. The section Data delineates the empirical data

    sample to be fed into the neural networks for training and testing, followed by the section

    Results which interprets the results. The implications of the results are further elaborated

    in the section Discussion and Conclusions.

    Literature

    Previous research into gamblers behavior and decisions can be subsumed under two

    categories, which the author refers to as economic/psychological models and sociological

    studies, respectively, and which are briefly outlined in the following subsections Eco-

    nomic/Psychological Models and Sociological Studies. The subsequent subsectionWhat Gaps Does This Research Fill highlights the gaps in the foregoing previous

    research that this article seeks to fill.

    Economic/Psychological Models

    Traditional economics assumes that human beings are rational agents with a view to

    maximizing their advantage. In view of the huge expenses and profit of (Ladouceur 2004)

    as well as government tax on casinos and gambling websites (American Gaming Asso-

    ciation 2008c), it is crystal clear that the total losses of all gamblers on gambling at a

    casino or a gambling website must substantially exceed their total winnings. In otherwords, gambling does not maximize gamblers advantage, so traditional economics per se

    falls short of being able to explain gamblers decision to gamble in the first place.

    Behavioral economics, however, as a relatively newer branch of economics (with Nobel

    Prize laureate D. Kahneman as one exponent), assumes bounded-rationality of humans

    with one consequence being that judgment under uncertainty often relies on heuristics

    which are sometimes prone to biases (Kahneman 2003; Kahneman et al. 1982). Gamblers

    decisions to gamble may be an exemplification of such judgments. Although not explicitly

    articulated, Delfabbro et al. (2006) carried out empirical research to prove the existence of

    particular gamblers mentalities that could be regarded as being implied by behavioraleconomists bounded-rationality assumption. Further to heuristics and biases, behavioral

    economics theorized decisions under risk by the prospect theory (Kahneman 2003;

    Kahneman and Tversky 1979), which critiqued and was an alternative to the rather dated

    expected utility theory of traditional economics. The prospect theory introduced the con-

    cept of decision weights, which took the place of probabilities in the expected utility

    theory. Decision weights are generally lower than the corresponding probabilities, except

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    5/18

    in the range of low probabilities. In other words, low probabilities, say, of wins in gambles

    are over-weighted in human decisions under risk, leading to decisions to take gambles or

    the like.

    A further direct consequence of bounded-rationality, heuristics, and biases is the various

    illusions and fallacies of gamblers, which can be categorized as follows:

    the misconceived dependence between successive gamble outcomes, which the famous

    term gamblers fallacy is coined to describe (Clotfelter and Cook 1993; Ladouceur

    2004), e.g., a win being reckoned to be very likely after a series of losses, a certain

    recently occurred outcome being believed to be unlikely to recur in the near future,

    illusions of control, e.g., opining that the gambler himself/herself has more control over

    a gambles outcome than other competing gamblers or even the banker (Ladouceur

    2004), and

    superstitions, e.g., thinking that some rituals or other unrelated happenings bear on the

    outcomes of gambles (Ladouceur 2004).All these three categories of illusions and fallacies of gamblers end up with gamblers

    believing that their own skills, strategies, and rituals can beat competing gamblers or even

    the bankers.

    Sociological Studies

    There has been research into the macroscopic, cross-sectional analysis of socio-economic

    factors determining gamblers gambling propensity (Gill et al. 2006; Delfabbro et al. 2006;

    Layton and Worthington 1999). Although there is also research that aims to take a lon-

    gitudinal perspective, their actual contribution concerns comparing and contrasting gam-

    blers gambling propensity before and after a certain event, such as the opening of a casino

    (Jacques and Ladouceur 2006), in a macroscopic manner. The author is aware of no

    research specifically performing longitudinal or time-series analysis to microscopically

    model gamblers behavior and decisions as to their successive bet amounts. Again, whilst

    there has been research into the macroscopic, cross-sectional analysis of how different

    demographic factors or the like of individual gamblers impinge on their trajectories of

    developing pathologic gambling (Nelson et al. 2006) and also research into the classifi-

    cation of gamblers based on the trajectories of their gambling propensity over age (Vitaro

    et al. 2004), the author is aware of no research microscopically modeling temporal tra-jectories of gamblers cumulative winnings/losses.

    What Gaps Does This Research Fill?

    It is more than indisputable that both the economic/psychological models and the socio-

    logical studies depicted above have been substantively contributing to gambling research.

    The economic/psychological models are primarily analytical in the sense of explaining the

    motivation behind gambling by means of economic/psychological theories. In contrast, the

    sociological studies are cross-sectional, macroscopic, or for the purpose of gamblers

    classification and are consequently macroscopically modeling and predicting a gamblers

    gambling propensity based on some socio-economic and demographic factors or the like.

    So far, there is no research directly modeling gamblers behavior and decisions. Nor are

    there any models that are quantitatively predictive in this aspect.

    In summary, there has been no previous research on longitudinally modeling or pre-

    dicting gamblers behavior and decisions microscopically and quantitatively as to their

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    6/18

    successive bet amounts. Nor is there any previous research on longitudinally modeling or

    predicting the temporal trajectory of gamblers cumulative winnings/losses. This research

    aims to fill this gap by developing such models.

    In addition, as there are no grounds to assume linearity or a particular type of non-

    linearity (e.g., exponential) when developing these models, traditional parametric model-ing techniques, which make such a presumption, are regarded as insufficient for these

    models. As such, this research employed the non-parametric modeling technique of neural

    networks, which is a type of artificial intelligence, on the grounds of their black-box

    (sometimes referred to as model-free) nature, which does not presume linearity or any

    particular type of non-linearity. Therefore, another contribution of this research is its

    application of non-parametric modeling and artificial intelligence in this area as opposed to

    traditional parametric modeling techniques, such as least squares regression, which are

    widely used in the aforesaid previous research.

    However, for the author, the most far-reaching contribution of this research is that the

    various illusions and fallacies of gamblers, as enumerated in the subsection Economic/

    Psychological Models above, can be partially invalidated by the models constructed in

    this article. This research also explores and exemplifies the usefulness of neural networks

    and artificial intelligence at large (Hu et al. 2007; Kaklauskas and Zavadskas 2007;

    Schocken and Ariav 1994), apart from traditional parametric statistics, in the context of

    gambling research and socio-economically and politically controversial issues.

    Methodology

    Like neural networks employed in forecasts for financial product pricing (Bennell and

    Sutcliffe 2004), backpropagation neural networks are employed in this research. The

    reason for this approach is that backpropagation neural networks are the most popular

    neural network paradigm (Walczak 2001) and have long been demonstrated as universal

    approximators (Hornik et al. 1989; White 1990). Moreover, past research has indicated that

    backpropagation neural networks have superior performance to other neural network

    paradigms (Walczak2001).

    Shown in Fig. 1 below is the structure of the fully connected backpropagation neural

    network adopted in this research for M1: the model for successive bet amounts. Also in

    Fig. 1 are the definitions of the related variables. Simply speaking, this network tries topredict the normalized bet amount betik to be laid by a particular gambler k in the next

    game i based on his/her normalized winnings/losses wini-1,k, wini-2,k,, wini-n,k in his/

    her n immediately preceding games and his/her current gambling account balance bal_sik.

    Such a predicted, normalized bet amount is denoted by betik in Fig. 1.

    Whereas readers can refer to relevant literature on neural networks for details (Nikolaev

    and Iba 2006; Ananda Rao and Srinivas 2003; Arbib 2003), the basic use of the neural

    network in Fig. 1 is as in the following steps:

    1. Collect and prepare an empirical data sample consisting ofwini-1,k, wini-2,k,, wini-

    n,k, bal_sik and betik for game i = n ? 1, n ? 2, and for each gambler k.2. Out of the data sample in step 1, select a sub-sample for training the network. Here, a

    sub-sample means wini-1,k, wini-2,k,, wini-n,k, bal_sik, and betik for some particular

    set of selected (i, k)s.

    3. Train the network by inputting wini-1,k, wini-2,k,, wini-n,k and bal_sik into it for

    each (i, k) in the training sub-sample of step 2, calculate the mean square error (MSE)

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    7/18

    between the predicted output betik and betik over all (i, k)s in the training sub-sample,

    and adjust the network parameters intrinsic to the nodes in the network (Nikolaev and

    Iba 2006; Ananda Rao and Srinivas 2003; Arbib 2003) so as to minimize MSE which

    is defined by

    MSE 1NTR

    X

    i;k2trainingsub

    -sample

    betik betik2 and NTR the size of the training sub-sample:

    4. Repeat step 3 above by a large number of episodes until the network converges, i.e.,

    until its parameters stabilize.

    5. Out of the data sample in step 1, select the remaining sub-sample (i.e., observations

    not in the training sub-sample) for testing the network.

    6. Test the network by inputting wini-1,k, wini-2,k,, wini-n,k, and bal_sik into it for each

    (i, k) in the testing sub-sample of step 5, and calculating the mean magnitude of

    relative error (MMRE) (Foss et al. 2003; Conte et al. 1986) between BETik (i.e., the

    predicted output betik after denormalization) and BETik (i.e., betik after denormaliza-

    tion) over all (i, k)s in the testing sub-sample to measure how accurately the network

    predicts BETik where

    MMRE 1NTE

    X

    i;k2testingsub-sample

    BETik BETikBETik

    and NTE the size of the testing sub-sample: 1

    kiwin ,3

    kiwin ,2

    kiwin ,1

    iksbal _

    .

    .

    .

    .

    .

    .

    .

    ikteb

    wheremeans a neuron;

    k = the gambler number;i = the game number, starting from n +

    1 for each gambler k;

    ikwin = the normalized (see later

    paragraphs of this Section)ikWIN ;

    iksbal _ = the normalized ikSBAL _ ;

    ikteb = the normalized ikTEB ;

    ikWIN = the winnings ofgambler k in his/her

    game i if positive and the loss inhis/her game i if negative;

    ikSBAL _ = the gambling account balance of

    gambler kat the start of his/her

    game i;

    ikTEB = an estimate of

    ikBET as predicted

    by this neural network;

    ikBET = the total bet amount of gambler k in

    his/her game i, covering all roundsof bets in the game

    Note 1; and

    n = the number of each gamblerspreceding games winnings/lossesto be included as inputs into thisneural network.

    input layer

    hidden layer

    output layer

    kniwin ,

    Note 1: A Texas Holdem game consistsof several rounds of betting.

    Fig. 1 The structure of the neural network for M1

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    8/18

    The smaller the MMRE, the more accurately the network predicts.

    Readers must have noticed the word normalized in Fig. 1. Normalization of the

    inputs and outputs of the neural network here is necessary to guarantee that the values of all

    the inputs and outputs are confined to a certain range, say, [0, 1], to which the network

    would be tuned during training in steps 3 and 4 above, and thus no extrapolation wouldoccur afterwards (Nahvi and Esfahanian 2005; Drew 2004). Normalization of, say, WINikto the range [0, 1] is done by the formula below:

    winik WINik WINminWINmax WINmin

    where WINmin = the minimum ofWINikover all (i, k)s in the whole data sample of step 1;

    and WINmax = the maximum of WINik over all (i, k)s in the whole data sample.

    Conversely, denormalization is done by WINik= winik (WINmax - WINmin) ? WINmin.

    Likewise, normalization and denormalization are done analogously for other variables.

    Also, whilst the numbers of input layer neurons and output layer neurons in Fig. 1 are

    obviously n ? 1 and 1, respectively, readers may notice that the number of hidden layer

    neurons therein is free to choose. On the advice of Morphet (2004) and Azoff (1994), the

    number of hidden layer neurons was set to

    5 if the number of input layer neurons is less than 10, and

    n1

    2

    1 otherwise where xb c = the largest integer less than x.Another kind of leeway is that a certain transfer function (Nikolaev and Iba 2006;

    Ananda Rao and Srinivas 2003; Arbib 2003) is attached to each hidden and output layer

    neuron, and such a transfer function is again free to choose, usually, among the pure-linearfunction, the log-sigmoid function, the sigmoid function, etc. Upon experiments on these

    various functions, the sigmoid function was found to give rise to the highest prediction

    accuracy or the smallest MMRE (as defined in (1)) for our data sample, so the sigmoid

    function was chosen for all the hidden and output layer neurons. Likewise, the value n is

    discretionary. Again the criterion for choosing its value was to render the highest pre-

    diction accuracy or the smallest MMRE for our data sample. The section Results below

    will show the MMREfor each value ofn within a reasonable range. By the same token, the

    network inputs wini-1,k, wini-2,k,, wini-n,k, and bal_sik in Fig. 1 were not arbitrarily

    determined at the very beginning of the research, but were the result of an extended

    searching process of trial and error. In fact, beti-1,k, beti-2,k,, bal_si-1,k, bal_si-2,k,,etc., as network inputs were all tried but wini-1,k, wini-2,k,, wini-n,k, and bal_sik pre-

    vailed in the end based on their highest prediction accuracy or the smallest MMREfor our

    data sample.

    Figure 2 below shows the structure of the fully connected backpropagation neural

    network adopted in this research for M2: the model for the temporal trajectory of cumu-

    lative winnings/losses. Also there are the definitions of the variables in addition to those

    defined in Fig. 1. Simply speaking, this network tries to predict the normalized, cumulative

    winnings/loss cwinik of a particular gambler k immediately after the next game i based

    on his/her normalized, cumulative winnings/losses cwini-1,k

    , cwini-2,k

    ,, cwini-m,k

    immediately after each of his/her m immediately preceding games and his/her random walk

    indicator RWIi-1,k immediately after the last game i - 1. Such a predicted, normalized,

    cumulative winnings/loss is denoted by cwinik in Fig. 2. Again, of course, the network

    inputs cwini-1,k, cwini-2,k,, cwini-m,k, and RWIi-1,k in Fig. 2 were not determined at the

    very beginning but were the result of an extended searching process of trial and error

    analogous to that for Fig. 1.

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    9/18

    The random walk indicator is recommended as a predictor in the field of finance for

    predicting quantities of a cumulative nature. In view of cwinik in Fig. 2 being cumulative

    albeit unrelated to finance, the random walk indicator RWIi-1,k of gambler k immediately

    after his/her game i - 1 is included as an input to the neural network and is define by

    RWIi1;k

    Dcwini1;kD

    cwinkffiffiffiffiffiffiffiffiffiffii 1p

    2

    where cwini-1,k= cwini-1,k- cwin0,k; cwin0,k= the normalized CWIN0,k; CWIN0,k=

    the cumulative winnings of gambler k immediately after his/her game 0 (i.e., at the very

    beginning before his/her game 1) if positive or his/her cumulative loss immediately after

    his/her game 0 if negative : 0 by definition; Dcwink PNkp1

    cwinp;kcwinp1;kj jNk

    ; and Nk= the

    total number of consecutive games played by gambler k.

    The basic use of the neural network in Fig. 2 is analogous to that in Fig. 1 except that

    the inputs and outputs are different accordingly. Consequently, the MMREto measure how

    accurately the network in Fig. 2 predicts cwinik becomes

    MMRE 1NTE

    X

    i;k2testingsub-sample

    cwinik cwinikcwinik

    : 3

    kicwin ,3

    kicwin ,2

    kicwin ,1

    kiRWI ,1

    .

    .

    .

    .

    .

    .

    .

    iknicw

    wheremeans a neuron;

    k = the gambler number;i = the game number, starting from m + 1

    for each gambler k;

    ikcwin = the normalized ikCWIN ;

    kiRWI ,1 = the random walk indicator of gambler

    k immediately after his/her gamei - 1, as elaborated in laterparagraphs of this Section;

    iknicw = the normalized

    ikNICW ;

    ikCWIN = =

    i

    p

    pkWIN1

    = the cumulative

    winnings of gambler k immediately

    after his/her game i if positive orhis/her cumulative lossimmediately after his/her game i ifnegative;

    ikNICW = an estimate of ikCWIN as predicted

    by this neural network; andm = the number of the gamblers

    cumulative winnings/loss to beincluded as inputs into this neuralnetwork.

    input layer

    hidden layer

    output layer

    kmicwin ,

    Fig. 2 The structure of the neural network for M2

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    10/18

    Data

    Like all backpropagation neural networks, those used in this research required an empirical

    data sample for training and later testing of prediction accuracy. Such a data sample was

    collected from a legally licensed, well-established cyber-gambling website www.betxxxx.com, part of the websites name being concealed by the xxxx for anonymity. There,

    international gamblers uniquely identified by their registered identifiers played poker games

    like Texas Holdem, Omaha High, Omaha Hi-Lo, and Seven Card Stud round the clock. This

    website was chosen because it was one of the very few cyber-gambling websites where the

    unique identifiers of all gamblers in an ongoing game together with the gamblers origins

    (i.e., cities or countries), their successive bet amounts, their decisions (e.g., to fold, raise,

    etc.), and their gambling account balances were all displayed. This research focused on

    Texas Holdem because of the top popularity of this family of games in major gambling

    markets (Clark 2006) so as to make our research the most representative of general gam-

    blers behavior and decisions.

    The collection of the data sample involved two stages, in the first of which ongoing

    Texas Holdem games at the aforementioned website were videoed using a screen capture

    software package. In the second stage, such videos were replayed to identify gamblers who

    played a long enough series of consecutive games for the purpose of our longitudinal

    modeling. It was found that most gamblers played for a short while, whereas only a small

    proportion of them played as many as around 100 consecutive games, which seemed to be

    the practical maximum. In fact, from the long videos we managed to identify only six such

    persistent gamblers, totaling 675 games. For each of these persistent gamblers, his/her

    bet amounts, gambling account balances, etc., throughout his/her series of games werenoted during the replays. As an example, Table 1 is an abridged copy of such data for one

    of the persistent gamblers whose identifier was Trevor38. Subsequently, Table 2 shows

    the data derived from Table 1 for the construction of M1: the model for successive bet

    amounts with the number of preceding games winnings/losses n = 6 as an example, the

    data here being before normalization (see the section Methodology above) for easier

    visualization. Similarly, Table 3 shows the data derived from Table 1 for the construction

    of M2: the model for the temporal trajectory of cumulative winnings/losses with the

    number of preceding games cumulative winnings/losses m = 6 as an example, again the

    data here being before normalization (see the section Methodology above) for easier

    visualization.

    Results

    M1: The Model for Successive Bet Amounts

    Table 4 shows the MMREs (as defined in Eq. 1) of M1 with n = 6, 7,, 20. They are all in

    the order of 10-2 and thus are rather small, especially when n = 6, 7, 8, 13, or 14,

    indicating rather accurate prediction ofBETik by M1. In fact, their small magnitudes imply

    that the mean magnitude of the error in the prediction of BETik is only around 13% of the

    magnitude of BETik. It is noteworthy that analogous to any other neural networks for

    prediction purposes, the neural network for M1 only tries to predict the main trend of a

    gamblers normalized bet amount betik by picking up the pattern(s) (presumably non-

    linear) as reflected in the predictors wini-1,k, wini-2,k,, wini-n,k, and bal_sik, which

    concern the gamblers normalized winnings/losses in his/her n immediately preceding

    J Gambl Stud

    123

    http://www.betxxxx.com/http://www.betxxxx.com/http://www.betxxxx.com/http://www.betxxxx.com/
  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    11/18

    games and his/her current gambling account balance. Nevertheless, the network is notcapable of tracking any extraordinary bet amount laid by the gambler in a particular game,

    say, due to his/her exceptionally strong hand of cards in that game when he/she would tend

    to bet exceptionally heavily or vice versa. Therefore, it is true that occasionally the

    magnitude of relative error MRE BETikBETikBETik

    for some game i played by gambler kin our

    testing sub-sample was as large as over 1 or 2. The foregoing MMREs in the order of 10-2

    Table 1 Data collected for gambler Trevor38 whose gambler number k= 1

    Game (i) BAL_Si1 BAL_Mi1 BETi1 =

    BAL_Si1 - BAL_Mi1

    BAL_Ei1 = BAL_Si?1,1 WINi1 =

    BAL_Ei1 - BAL_Si1

    1 90.16 89.16 1.00 89.16 -1.002 89.16 87.66 1.50 87.66 -1.50

    3 87.66 87.66 0.00 87.66 0.00

    4 87.66 87.41 0.25 87.41 -0.25

    5 87.41 86.66 0.75 88.08 ?0.67

    6 88.08 86.83 1.25 86.83 -1.25

    7 86.83 86.83 0.00 86.83 0.00

    8 86.83 86.83 0.00 86.83 0.00

    9 86.83 85.83 1.00 85.83 -1.00

    10 85.83 85.33 0.50 85.33 -0.50

    11 85.33 83.83 1.50 88.03 ?2.70

    12 88.03 87.78 0.25 87.78 -0.25

    13 87.78 87.53 0.25 87.53 -0.25

    14 87.53 86.78 0.75 86.78 -0.75

    15 86.78 86.78 0.00 86.78 0.00

    16 86.78 86.53 0.25 86.53 -0.25

    17 86.53 85.53 1.00 85.53 -1.00

    18 85.53 85.03 0.50 85.03 -0.50

    19 85.03 84.78 0.25 84.78 -0.25

    20 84.78 82.78 2.00 89.38 ?4.60

    : : : : : :

    : : : : : :

    BAL_Si1 = Trevor38s gambling account balance at the start of his game i

    BAL_Mi1 = Trevor38s gambling account balance in the middle of his game i after laying all rounds of

    bets in the game

    BETi1 = Trevor38s total bet amount in his game i, covering all rounds of bets

    BAL_Ei1 = Trevor38s gambling account balance at the end of his game i, which is equal to BAL_Mi1plus the received winnings in game i, if any, noting that here the received winnings refers to the money

    received upon winning the gameWINi1 = Trevor38s winnings in his game i if positive and the loss in game i if negative, noting that: (1)

    here the winnings differ from the received winnings mentioned in the definition of BAL_Ei1 above in that, in

    a game, the winnings (WINi1) is the received winnings net of the total bet (BETi1), and (2) here the loss must

    be equal to BETi1 upon losing the game

    Note: The currency for the games played by Trevor38 as listed in this table happened to be the Great

    Britain pound (GBP), so all monetary variables in this table are in GBP. Nevertheless, the international

    gambling website www.betxxxx.com underlying this research has games in various other currencies, e.g.,

    US dollars (USD), euros (EUR), etc. It is obvious that the validity of this research is not confined to any

    particular currency

    J Gambl Stud

    123

    http://www.betxxxx.com/http://www.betxxxx.com/
  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    12/18

    are just the means of the MREs averaged over all the games played by all the gamblers in

    our testing sub-sample. In other words, the networks prediction tracks the overall trend of

    any gamblers successive bet amounts well even if occasionally there are considerable

    prediction errors (or residuals if deliberately using the terminologies of statistical

    regression, which is a more popularly used prediction method than neural networks).

    Similar things are, for example, true of financial market prediction models which are,

    however, regarded as acceptably useful and accurate if they are able to track the overalltrend of the market well, even if they fail severely during periods of financial turmoil.

    In addition, it is worthwhile to point out that only one single trained neural network for

    M1 was devised by using our training sub-sample covering all the six gamblers, and the

    high prediction accuracy above was achieved by applying this identically trained neural

    network to our testing sub-sample, which also included all the six gamblers. This implies

    that the relationship between the main trend of betik and the pattern(s) in the predictors

    wini-1,k, wini-2,k,, wini-n,k, and bal_sik as represented intrinsically by the post-training

    parameters inside this trained neural network, is almost the same for all the six gamblers in

    our data sample. In other words, if there is any difference in the main trend ofbetik between

    the six gamblers, such idiosyncrasy is reflected not predominately by the aforesaid

    relationship but almost totally by the difference in the pattern(s) of wini-1,k,

    wini-2,k,, wini-n,k, and bal_sik between these gamblers. The influence of a gamblers

    skills, strategies, and personality on his/her successive bet amounts is almost totally

    reflected by the pattern(s) of his/her winnings/losses in the several immediately preceding

    games and his/her gambling account balance.

    Table 2 Data derived from Table 1 for the construction of M1 with, say, n = 6

    Game (i) Using game i of Trevor38 to either train or test M1

    Upon normalization, these are inputs Predicted output,beti1; upon

    denormalization,

    estimatesWINi-6,1 WINi-5,1 WINi-4,1 WINi-3,1 WINi-2,1 WINi-1,1 BAL_Si1 BETi1

    7 -1.00 -1.50 0.00 -0.25 ?0.67 -1.25 86.83 0.00

    8 -1.50 0.00 -0.25 ?0.67 -1.25 0.00 86.83 0.00

    9 0.00 -0.25 ?0.67 -1.25 0.00 0.00 86.83 1.00

    10 -0.25 ?0.67 -1.25 0.00 0.00 -1.00 85.83 0.50

    11 ?0.67 -1.25 0.00 0.00 -1.00 -0.50 85.33 1.50

    12 -1.25 0.00 0.00 -1.00 -0.50 ?2.70 88.03 0.25

    13 0.00 0.00 -1.00 -0.50 ?2.70 -0.25 87.78 0.25

    14 0.00 -1.00 -0.50 ?2.70 -0.25 -0.25 87.53 0.75

    15 -1.00 -0.50 ?2.70 -0.25 -0.25 -0.75 86.78 0.00

    16 -0.50 ?2.70 -0.25 -0.25 -0.75 0.00 86.78 0.25

    17 ?2.70 -0.25 -0.25 -0.75 0.00 -0.25 86.53 1.00

    18 -0.25 -0.25 -0.75 0.00 -0.25 -1.00 85.53 0.50

    19 -0.25 -0.75 0.00 -0.25 -1.00 -0.50 85.03 0.25

    20 -0.75 0.00 -0.25 -1.00 -0.50 -0.25 84.78 2.00

    : : : : : : : : :

    : : : : : : : : :

    Note: The variables in the headings of this table are as defined in Table 1

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    13/18

    Table3

    Dataderive

    dfrom

    Table1fortheconstructionofM2with,say,m=

    6

    Game(i)

    UsinggameiofTrevor38toeither

    trainortestM2

    Uponnormalization,

    theseareinputs

    Predictedoutput,

    CW^INi1;

    upondenormalization,estimates

    CWIN

    i-6,

    1

    CWIN

    i-5,

    1

    CW

    INi-4,

    1

    CWIN

    i-3,

    1

    C

    WIN

    i-2,

    1

    CWIN

    i-1,

    1

    RWIi

    -1,

    1

    CWIN

    i1

    7

    -1.0

    0

    -2.5

    0

    -2

    .50

    -2.7

    5

    -

    2.0

    8

    -3.3

    3

    *

    -3.3

    3

    8

    -2.5

    0

    -2.5

    0

    -2

    .75

    -2.0

    8

    -

    3.3

    3

    -3.3

    3

    *

    -3.3

    3

    9

    -2.5

    0

    -2.7

    5

    -2

    .08

    -3.3

    3

    -

    3.3

    3

    -3.3

    3

    *

    -4.3

    3

    10

    -2.7

    5

    -2.0

    8

    -3

    .33

    -3.3

    3

    -

    3.3

    3

    -4.3

    3

    *

    -4.8

    3

    11

    -2.0

    8

    -3.3

    3

    -3

    .33

    -3.3

    3

    -

    4.3

    3

    -4.8

    3

    *

    -2.1

    3

    12

    -3.3

    3

    -3.3

    3

    -3

    .33

    -4.3

    3

    -

    4.8

    3

    -2.1

    3

    *

    -2.3

    8

    13

    -3.3

    3

    -3.3

    3

    -4

    .33

    -4.8

    3

    -

    2.1

    3

    -2.3

    8

    *

    -2.6

    3

    14

    -3.3

    3

    -4.3

    3

    -4

    .83

    -2.1

    3

    -

    2.3

    8

    -2.6

    3

    *

    -3.3

    8

    15

    -4.3

    3

    -4.8

    3

    -2

    .13

    -2.3

    8

    -

    2.6

    3

    -3.3

    8

    *

    -3.3

    8

    16

    -4.8

    3

    -2.1

    3

    -2

    .38

    -2.6

    3

    -

    3.3

    8

    -3.3

    8

    *

    -3.6

    3

    17

    -2.1

    3

    -2.3

    8

    -2

    .63

    -3.3

    8

    -

    3.3

    8

    -3.6

    3

    *

    -4.6

    3

    18

    -2.3

    8

    -2.6

    3

    -3

    .38

    -3.3

    8

    -

    3.6

    3

    -4.6

    3

    *

    -5.1

    3

    19

    -2.6

    3

    -3.3

    8

    -3

    .38

    -3.6

    3

    -

    4.6

    3

    -5.1

    3

    *

    -5.3

    8

    20

    -3.3

    8

    -3.3

    8

    -3

    .63

    -4.6

    3

    -

    5.1

    3

    -5.3

    8

    *

    -0.7

    8

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    :

    CWINi1

    Pi p1

    WINp1

    =

    Trevor38scumulativewinn

    ingsimmediatelyafterhisgameiifpositiveorhiscumulativelossimmediatelyafterhisgameiifnegative;and

    RWIi

    -1,

    1

    =

    Trevor3

    8srandomwalkindicator(seethesectionMethodologyabove)immediatelyafterhisgamei-

    1.

    Itsvalueforeachgameisomittedhereforthe

    sakeofsimplicity,giventhatitisnotmeaningfultoinc

    ludeithereinthatitcannotbede

    rivedtriviallyfromTable1butca

    nonlybecalculatedusingthecom

    plicatedEq.

    2

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    14/18

    M2: The Model for the Temporal Trajectory of Cumulative Winnings/Losses

    Table 5 is an M2 counterpart of Table 4 where the MMREs here are as defined in Eq. 3.

    The MMREs are all in the order of 10-3

    or 10-2

    and thus are rather small, especially whenm = 6, 7, 8, or 13, indicating quite an accurate prediction of CWINik by M2. In fact, their

    small magnitudes imply that the mean magnitude of the error in the prediction ofCWINik is

    only around 0.84% of the magnitude of CWINik. When interpreting such MMREs, readers

    are reminded to refer to the noteworthy point stated in the case of M1 above.

    In addition, it is worthwhile to point out that only one single trained neural network for

    M2 was devised by using our training sub-sample covering all the six gamblers, and the

    high prediction accuracy above was achieved by applying this identically trained neural

    network to our testing sub-sample, which also included all the six gamblers. This implies

    that the relationship between the main trend of CWINik and the pattern(s) in the predictors

    cwini-1,k, cwini-2,k,, cwini-m,k, and RWIi-1,k, as represented intrinsically by the post-training parameters inside this trained neural network, is almost the same for all the six

    gamblers in our data sample. In other words, if there is any difference in the main trend of

    cwinik between the six gamblers, such idiosyncrasy is reflected not predominately by the

    aforementioned relationship but almost totally by the difference in the pattern(s) of

    cwini-1,k, cwini-2,k,, cwini-m,k, and RWIi-1,k between these gamblers. That is to say, the

    influence of a gamblers skills, strategies, and personality on his/her cumulative winnings/

    losses is almost totally reflected by the pattern(s) of his/her cumulative winnings/losses in

    the several immediately preceding games.

    Discussion and Conclusions

    Summarizing all its regulations, a Texas Holdem game involves only one type of behavior

    or decisions of a participating gambler, which is the amount of bet to be laid in the game

    (Harrington and Robertie 2004). Even the decisions to fold and to raise are just particular

    Table 5 MMREs for M2 with m = 6, 7,, 20

    n 6 7 8 9 10 11 12 13

    MMRE 0.0079 0.0082 0.0083 0.0103 0.0116 0.012 0.015 0.0096

    14 15 16 17 18 19 20

    MMRE 0.0188 0.0191 0.0143 0.0204 0.0414 0.0346 0.0251

    Bold values represent the best results

    Table 4 MMREs for M1 with n = 6, 7,, 20

    n 6 7 8 9 10 11 12 13

    MMRE 0.01 0.0103 0.0138 0.0171 0.026 0.0208 0.0174 0.014

    14 15 16 17 18 19 20

    MMRE 0.013 0.0241 0.0226 0.0227 0.0235 0.0183 0.0322

    Bold values represent the best results

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    15/18

    cases to freeze the current bet amount and to increase the current bet amount, respectively.

    If two gamblers lay the same amount of bet in each identical game (with the same hand,

    etc.), then it can be said that the behavior or decisions of the two gamblers are identical or,

    equivalently, the skills, strategies, and personalities of the two gamblers are identical as far

    as Texas Holdem games are concerned. The findings from the section Results above forM1: The Model for Successive Bet Amounts are that all gamblers successive bet

    amounts in the 675 Texas Holdem games of our data sample rather strictly obey the

    longitudinal model M1 constructed in this research. Given that M1 relates a gamblers bet

    amount in a game with his/her gambling account balance at the beginning of the game and

    the winnings/losses in his/her few (n) immediately preceding games, it can be concluded

    that if some gamblers in our data sample have the same initial gambling account balance

    and the same winnings/losses in their few (n) initial games, such gamblers will give rise to

    almost the same subsequent time-series of bet amounts as predicted by M1. In other words,

    such gamblers will have almost identical behavior and decisions, or equivalently, they have

    almost identical skills, strategies, and personalities as far as Texam Holdem games are

    concerned. Hence, such gamblers will lead to almost the same gambling winnings/losses in

    the long run. Even if there is a difference, it will be negligible as attested by the small

    MMREin Table 4. In general, in respect of the aforesaid gamblers behavior and decisions,

    the only substantial idiosyncrasy of an individual gambler lies in his/her gambling account

    balance and the pattern(s) of the winnings/losses in his/her few (n) initial games, through

    which his/her skills, strategies, and personality are practically fully reflected.

    Even if a more result-oriented viewpoint is taken to focus only on the cumulative

    winnings/loss just as most real-world gamblers do, a similar conclusion can be drawn. The

    findings from the section Results above for M2: The Model for the Temporal Trajectoryof Cumulative Winnings/Losses are that all gamblers temporal trajectories of cumulative

    winnings/losses in the 675 Texas Holdem games of our data sample quite strictly obey the

    longitudinal model M2 constructed in this research. Given that M2 relates a gamblers

    cumulative winnings/loss immediately after a game with his/her cumulative winnings/loss

    immediately after each of his/her few (m) immediately preceding games, it can be con-

    cluded that if some gamblers in our data sample have the same initial cumulative winnings/

    loss immediately after each of their few (m) initial games and thus have the same random

    walk indicators, such gamblers will give rise to almost the same subsequent time-series of

    cumulative winnings/losses as predicted by M2. Even if there is a difference, it will be

    negligible as attested by the small MMREin Table 5. In general, in respect of the foregoingcumulative winnings/losses, the only substantial idiosyncrasy of an individual gambler lies

    in the pattern(s) of the cumulative winnings/losses in his/her few (m) initial games, through

    which his/her skills, strategies, and personality are practically fully reflected.

    A combined further conclusion from the last two paragraphs is that a gamblers skills,

    strategies, and personality are practically fully reflected in the few (n or m) initial games.

    Beyond these few initial games, both his/her behavior and decisions (in terms of the bet

    amounts in successive games) and his/her cumulative winnings/losses are fairly accurately

    predictable by M1 or M2. In other words, on the one hand, gamblers behavior, decisions,

    and cumulative winnings/losses have substantive commonality among different gamblersin that the inherent relationship between the inputs and outputs of M1 and that between

    those of M2 are almost the same among different gamblers. On the other hand, the only

    idiosyncrasies of individual gamblers are exhibited by the pattern(s) of the winnings/losses

    in each gamblers few (n or m) initial games, through which the influence of his/her skills,

    strategies, and personalities on his/her behavior, decisions, and cumulative winnings/losses

    is practically fully reflected. Therefore, this conclusion partially invalidates gamblers

    J Gambl Stud

    123

  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    16/18

    various illusions and fallacies (Clotfelter and Cook 1993), as detailed in the subsection

    Economic/Psychological Models above, that they can outperform other gamblers or

    even bankers through specific skills, strategies, and personalities. Although the conclusion

    here cannot rule out the possibility of individual gamblers idiosyncratic performance

    ascribable to their specific skills, strategies, and personalities, an implication for gamblersis that it is not worthwhile or even relevant to excessively and endlessly experiment beyond

    n or m games with the gamblers skills, strategies, and personalities in an attempt to

    outperform other gamblers or even bankers.

    From casinos and cyber-gambling websites viewpoint, the implications of this research

    are that their profits can be better maximized through emphasizing the provision of par-

    ticular games and redesigning their games based on the different predicted temporal tra-

    jectories of typical gamblers cumulative winnings/losses in different games and/or games

    of different designs as empirically verified by M2.

    From a broader viewpoint, such models can assist the work of related government

    policy-makers, gambling industry operators, economists, sociologists, psychiatrists, and

    psychologists. In particular, through these models, they all can realize and even simulate

    the temporal trajectory of a general gamblers losing (or, less likely, winning) money in

    such a gambling game as Texas Holdem in this paper and thus quantitatively determine

    gamblings political, financial, economic, social, psychiatric, and psychological contribu-

    tions and problems in respect of the individual gambler himself/herself and society at large.

    Since such models can be applied generically to other gambling games by just feeding into

    it empirical data samples of gamblers bet amounts in successive games, etc., in such other

    gambling games, different trajectories for different gambling games can be derived for

    comparison and contrast between their aforesaid contributions and problems.Such models also explore and exemplify the usefulness of neural networks and artificial

    intelligence at large (Hu et al. 2007; Kaklauskas and Zavadskas 2007; Schocken and Ariav

    1994), apart from traditional parametric statistics, in the context of gambling research and

    socio-economically and politically controversial issues in general.

    These models together with its predicted gamblers behavior and decisions can also

    serve well as an educational tool for gamblers, in particular, pathological gamblers

    undergoing rehabilitation, hopefully helping to partially uproot their irrational illusions

    and fallacies, as elaborated on in the subsection Economic/Psychological Models

    above, that they can outperform other gamblers or even bankers through their specific

    skills and strategies.

    Acknowledgments The authors thank Macao Polytechnic Institute for generous financial support under

    Research Grant RP/ESCE-6/2004 and Dr. Jonathan Fearon-Jones for his proof-reading. Thanks are also due

    to Kai Piu Benny Chan and Chin Wa Tam for their assistance in data collection, data analysis, etc.

    References

    American Gaming Association. (2008a). Industry information: Fact sheets: Industry issues: Internet gam-

    bling. http://www.americangaming.org/Industry/factsheets/issues_detail.cfv?id=17. Accessed 27 Dec2008.

    American Gaming Association. (2008b). Industry information: Fact sheets: Statistics: Gaming revenue:

    Current-year data. http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=7.

    Accessed 27 Dec 2008.

    American Gaming Association. (2008c). Industry information: Fact sheets: Statistics: Tax payments

    commercial casinos. http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=10.

    Accessed 27 Dec 2008.

    J Gambl Stud

    123

    http://www.americangaming.org/Industry/factsheets/issues_detail.cfv?id=17http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=7http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=10http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=10http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=7http://www.americangaming.org/Industry/factsheets/issues_detail.cfv?id=17
  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    17/18

    Ananda Rao, M., & Srinivas, J. (2003). Neural networks: Algorithms and applications. Pangbourne, UK:

    Alpha Science International.

    Arbib, M. A. (Ed.). (2003). The handbook of brain theory and neural networks (2nd ed.). Cambridge, MA:

    MIT.

    Azoff, E. M. (1994). Neural network time series forecasting of financial markets. Chichester, UK: Wiley.

    Bennell, J., & Sutcliffe, C. (2004). Black-Scholes versus artificial neural networks in pricing FTSE 100options. Intelligent Systems in Accounting, Finance and Management, 12(4), 243260.

    Breen, R. B., & Zimmerman, M. (2002). Rapid onset of pathological gambling in machine gamblers.

    Journal of Gambling Studies, 18(1), 3143.

    Clark, B. (2006). The dying days of Las Vegas 1-5 stud. Two Plus Two Internet Magazine, 21.

    http://www.twoplustwo.com/magazine/issue21/clark0906.html. Accessed 4 Oct 2006.

    Clotfelter, C. T., & Cook, P. J. (1993). The gamblers fallacy in lottery play. Management Science,

    39(12), 15211525.

    Conte, S., Dunsmore, H., & Shen, V. Y. (1986). Software engineering metrics and models. Menlo Park, CS:

    Benjamin Cummings.

    Delfabbro, P., Lahn, J., & Grabosky, P. (2006). Its not what you know, but how you use it: Statistical

    knowledge and adolescent problem gambling. Journal of Gambling Studies, 22(2), 179193.

    Dewar, L. (2001). Regulating internet gambling: The net tightens on online casinos and bookmakers. ASLIBProceedings, 53(9), 353367.

    Drew, P. J. (2004). Time-scales of neural computation. PhD dissertation, Brandeis University, Waltham,

    MA.

    Feigelman, W., Gorman, B. S., & Lesieur, H. (2006). Examining the relationship between at-risk gambling

    and suicidality in a national representative sample of young adults. The American Association of

    Suicidology, 36(4), 396408.

    Foss, T., Stensrud, E., Kitchenham, B., & Myrtveit, I. (2003). A simulation study of the model evaluation

    criterion MMRE. IEEE Transaction on Software Engineering, 29(11), 985995.

    Friedman, J., Hakim, S., & Weinblatt, J. (1989). Casino gambling as a growth pole strategy and its effect

    on crime. Journal of Regional Science, 29(4), 615623.

    Gill, T., Grande, E. D., & Taylor, A. W. (2006). Factors associated with gamblers: A population-based

    cross-sectional study of South Australian adults. Journal of Gambling Studies, 22(2), 143164.Harrington, D., & Robertie, B. (2004). Harrington on holdem: Expert strategy for no-limit tournaments,

    Volume I: Strategic play. Henderson, NV: Two Plus Two Publishing LLC.

    Hornik, K., Stinchcombe, M., & White, H. (1989). Multilayer feedforward networks are universal

    approximators. Neural Networks, 2(5), 359366.

    Hu, P. J.-H., Wei, C.-P., Cheng, T.-H., & Chen, J.-X. (2007). Predicting adequacy of vancomycin regimens:

    A learning-based classification approach to improving clinical decision making. Decision Support

    Systems, 43(4), 12261241.

    Hugel, P., & Kelly, J. (2002). Internet gambling, credit cards and money laundering. Journal of Money

    Laundering Control, 6(1), 5765.

    Jacques, C., & Ladouceur, R. (2006). A prospective study of the impact of opening a casino on gambling

    behaviours: 2- and 4-year follow-ups. Canadian Journal of Psychiatry, 51(12), 764773.

    Jones, P., Clarke-Hill, C. M., & Hillier, D. (2000). Viewpoint: Back street to side street to high street toe-street: Sporting betting on the internet. International Journal of Retail and Distribution Management,

    28(6), 222227.

    Kahneman, D. (2003). Maps of bounded rationality: Psychology for behavioral economics. American

    Economic Review, 93(5), 14491475.

    Kahneman, D., Slovic, P., & Tversky, A. (Eds.). (1982). Judgment under uncertainty: Heuristics and biases.

    New York, NY: Cambridge University Press.

    Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica,

    47(2), 263291.

    Kaklauskas, A., & Zavadskas, E. K. (2007). Decision support system for innovation with a special emphasis

    on pollution. International Journal of Environment & Pollution, 30(3/4), 518528.

    Ladouceur, R. (2004). Gambling: The hidden addiction. Canadian Journal of Psychiatry, 49(8), 501503.

    Ladouceur, R., Sylvain, C., & Gosselin, P. (2007). Self-exclusion program: A longitudinal evaluation study.Journal of Gambling Studies, 23(1), 8594.

    Layton, A., & Worthington, A. (1999). The impact of socio-economic factors on gambling expenditure.International Journal of Social Economics, 26(1/2/3), 430440.

    MacLaurin, T., & MacLaurin, D. (2003). Casino gaming and tourism in Canada. International Journal of

    Contemporary Hospitality Management, 15(6), 328332.

    J Gambl Stud

    123

    http://www.twoplustwo.com/magazine/issue21/clark0906.htmlhttp://www.twoplustwo.com/magazine/issue21/clark0906.html
  • 8/8/2019 Using Neural Networks to Model the Behavior and Decisions of Gamblers in Particular Cyber Gamblers

    18/18

    Morphet, S. B. (2004). Modeling neural networks via linguistically interpretable fuzzy inference systems.

    PhD Dissertation, Syracuse University, Syracuse, NY.

    Moss, S. E., Ryan, C., & Wagoner, C. B. (2003). An empirical test of Butlers resort product life cycle:

    Forecasting casino winnings. Journal of Travel Research, 41(4), 393399.

    Nahvi, H., & Esfahanian, M. (2005). Fault identification in rotating machinery using artificial neural

    networks. Proceedings of the Institution of Mechanical Engineers: Part C: Journal of MechanicalEngineering Science, 219(2), 141158.

    Nelson, S. E., LaPlante, D. A., LaBrie, R. A., & Shaffer, H. J. (2006). The proxy effect: Gender and

    gambling problem trajectories of Iowa gambling treatment program participants. Journal of Gambling

    Studies, 22(2), 221240.

    Nikolaev, N. Y., & Iba, H. (2006). Adaptive learning of polynomial networks: Genetic programming,

    backpropagation and bayesian methods. New York, NY: Springer.

    Schocken, S., & Ariav, G. (1994). Neural networks for decision support: Problems and opportunities.

    Decision Support Systems, 11(5), 393414.

    Siegel, D., & Anders, G. (1999). Public policy and the displacement effects of casinos: A case study of

    riverboat gambling in Missouri. Journal of Gambling Studies, 15(2), 105121.

    Slutske, W. S., Caspi, A., Moffitt, T. E., & Poulton, R. (2005). Personality and problem gambling: A

    prospective study of a birth cohort of young adults. Archives of General Psychiatry, 62(7), 769775.Slutske, W. S., Jackson, K. M., & Sher, K. J. (2003). The natural history of problem gambling from age 18

    29. Journal of Abnormal Psychology, 112(2), 263274.

    Smith, A. D. (2004). Controversial and emerging issues associated with cybergambling (e-casinos). Online

    Information Review, 28(6), 435443.

    Vitaro, F., Wanner, B., Ladouceur, R., Brendgen, M., & Tremblay, R. E. (2004). Trajectories of gambling

    during adolescence. Journal of Gambling Studies, 20(1), 4769.

    Walczak, S. (2001). An empirical analysis of data requirements for financial forecasting with neural net-

    works. Journal of Management Information Systems, 17(4), 203222.

    White, H. (1990). Connectionist nonparametric regression: Multilayer feedforward networks can learn

    arbitrary mappings. Neural Networks, 3(5), 535549.

    J Gambl Stud