A Model For Gamers' Revenue in Casinos
A Model For Gamers' Revenue in Casinos
Jehoshua Eliashberg
Sam K. Hui
Raghuram Iyengar*
*Jehoshua Eliashberg is the Sebastian S. Kresge Professor of Marketing and Professor of Operations and
Information Management at the Wharton School of the University of Pennsylvania, Sam K. Hui an
Assistant Professor in Marketing at Stern School of Business of New York University, and Raghuram
Iyengar is an Assistant Professor in Marketing at the Wharton School of the University of Pennsylvania.
The authors are listed in alphabetical order. Please address all correspondence to Jehoshua Eliashberg
([email protected]). The authors benefitted from comments by Sunil Gupta and Christophe
Van den Bulte.
A Model for Gamers’ Revenues in Casinos
Abstract
We develop a model for predicting gamers’ revenues using data provided by a major casino
operator. Our model captures both gamers’ visit and gambling behavior. By incorporating latent
individual-level skill parameters, our model allows casino managers to tease apart whether “skill” or
“luck” is driving the revenue from each player. It also provides the casino operator with a better
characterization of players (high skill, high rollers etc.) and hence leads to more accurate predictions of
future revenue from each player. We estimate the model using Bayesian methods. We show that
predictions of future player revenues using our model are more accurate than those from nested and
alternative models.
1. Introduction
Consumers engage in various leisure activities. For instance, during 2005-2007, Americans made
3 visits per month on average to a mall with each trip lasting more than 1 hour and involved a spending of
more than $75 (Connolly and Rogoff 2008). Recent estimates indicate that Americans went to the cinema
about 5 times a year on average (Nationmaster 2009) and spent more than $9 Billion in a year (Motion
Picture Association of America 2009). In the gaming industry, which is the focus of this paper, the
average trip frequency to a casino is about 6 visits per year (Harrah’s Report 2006) with a spending of
$90 per trip (Morse and Goss 2007). In 2006, the gaming industry generated more than $90 Billion in
revenues (American Gaming Association 2006). Of this total, casinos accounted for around $59 Billion
with the remaining generated from alternatives such as online gambling, state lotteries and parimutuel
wagering. Recent estimates indicate that online gambling generated about $5.9 Billion, state lotteries
accounted for about $24 Billion while parimutuel wagering generated about $3.5 Billion (American
Gaming Association 2006). Hence, the casino environment is the dominant revenue generator. Despite
the size and significant revenues it generates, so far the casino industry has received little attention by
marketing academics.
The casino industry can be characterized as driven by database marketing. Other such industries,
for instance, are retail and movie exhibition chains. From the perspective of a retail chain, understanding
consumer behavior involves when consumers will visit a store (Morrison and Schmittlein 1981; Helsen
and Schmittlein 1993), what brand they buy and how much (Gupta 1988). Some of these issues are also
of great interest to casino operators. Similar to the behavior of movie goers which has been studied
extensively (see Eliashberg, Alberse and Leenders 2006 for a review), gamers come to the physical casino
facility for excitement and fun (Abt, Smith and Christiansen 1985; Kallick et al. 1979). However, unique
aspects of the casino context which deserve research attention are: i) Unlike the retail chain, in the casino
industry, there are significantly more cross-market outlet visits and ii) a casino operator observes
individual gamer-level data unlike the aggregate level of box-office data that a movie theater owner
observes (Swami et al. 1999) and iii) the active participation of gamers as opposed to their passivity in
1
both the retail and movie contexts. Such active behavior from gamers often requires certain skills that
casino operators may find costly to ignore. In comparison with other forms of gambling and in particular
to online gambling, there is much greater data available related to gamer behavior within a casino
environment (Cotte and Latour 2009). Having access to more data allows researchers to study closely
gaming behavior where gamer skill is essential as well as the opportunity to relate such latent skill to
Casino operators seek to better understand their customer base and to predict their future expected
revenues. Achieving this goal involves three major challenges. First, given the gambling context, the
revenues that each player generates in a trip is highly uncertain. As will be shown later in this paper, this
is due to the significant randomness in gambling outcomes. Observed past revenue from a player does not
provide a reasonable estimate for her future revenue. Thus, in order for a casino operator to predict future
expected revenues accurately, it needs to tease apart the “chance” component from the “skill” component
of gambling outcomes. Doing so will allow a casino manager to identify the low-skill players who are
likely to provide a higher expected revenue for the casino, and target those players accordingly.
Second, a consumer’s decision to visit and gamble in a casino may be dependent on the past
outcome of her previous visit(s). Some previous research in psychology (Clotfelter and Cook 1993;
Guryan and Kearney 2008) and economics (Stigler and Becker 1977; Becker and Murphy 1988) have
shown that the outcomes of past gambles can affect future gambling behavior. Thus, to obtain an accurate
prediction of future revenue, one needs to consider the dynamics of how the outcome of previous trip(s)
affects future visit and wagering behavior. For instance, after a gamer loses money in a trip (thus
generating revenue for the casino in that trip), she may become less enthusiastic about gambling and may
reduce her wager amount on her next trip (thus, the expected revenue will go down on the next trip).
Third, another key managerial question is whether casino visit and wagering behavior are related
to observable demographics information such as age and gender. If so, demographics information can be
used to identify and target high-value players and provide information to help in customer-base
management.
2
In this paper, we address all the aforementioned challenges and develop an integrated model that
jointly captures players’ visit and wagering behavior. At the heart of our model is a latent “skill”
parameter for each player, which affects the quality of the in-play decisions she makes (e.g., whether to
stand/hit in blackjack). This in turn increases or decreases the effective odds (and hence the expected
return) of her wager and changes the casino’s expected revenue. As will be discussed later, our model
allows us to estimate the latent skill for each player through her previous wagering and win/loss patterns.
This provides the casino manager with the opportunity to tease apart whether a certain gambling outcome
is due to skill, or due solely to chance. Furthermore, our model allows the outcome of previous visit to
affect both a consumer’s rate of future visit (thus the waiting time before her next trip) and the amount
that she wagers on the next trip. Finally, we also incorporate some demographics information into our
model, allowing us to examine how, if at all, demographic information is related to casino visit and wager
behavior.
We apply our model to player-level data obtained from a major casino operator, owning
properties in seven different geographical regions in the U.S. The dataset contains information concerning
when players’ visit a casino, their aggregate trip-level wager amount in slots and table games and net
wins/losses. We estimate our model using Bayesian methods. We find that our model provides more
accurate, individual-level prediction of future player revenue than predictions based on observed past
revenue. In addition, we find that some demographics information are indeed related to visit and
gambling behavior.
Our research diverges from previous research in the casino context by focusing on capturing
gambling behavior and predicting revenue on the individual level. Instead of modeling individual-level
revenue, previous studies have modeled industry-level annual revenues using time-series methods (Cargill
and Eadington 1978; Nichols 1998). In a separate stream of research, Baker and Marshall (2005) model
players’ wagered amount and their choice of gambling venues. They do not model how wagers are linked
with wins, and how players’ skill can affect their wins from table games. In addition, they use cross-
3
sectional survey data (rather than player-level longitudinal play data) and hence cannot address
The remainder of this paper is organized as follows. Section 2 outlines our modeling framework.
Section 3 describes the data. Section 4 reports our empirical results. Section 5 discusses managerial
implications in regards to targeting and individual-level revenue prediction. Section 6 concludes with
2. Model
In this section, we develop a model that describes gamers’ visit and wagering behavior. Our
overall modeling framework, similar in spirit to the integrated modeling framework of Gupta (1988)
We begin by specifying a model of visit in Section 2.1, which captures when a gamer will visit a
casino, and which region she visits. As we discussed earlier, these decisions are influenced to a large
extent by the gamer’s gambling wins/losses on her previous trip. Once a gamer enters a casino region, we
model the total amount of wagers she makes, and how she divides her wager between slot and table
games in Section 2.2. We then model the gambling outcomes in slot games and table games separately, in
Section 2.4 and 2.5, respectively. In both models, a key driver of gambling outcomes is the effective
“house advantage” (the expected loss per dollar wagered by the player) that is driven by the odds that the
casino offers, the player’s choice of wager, and the player’s skill level (in the case of table games). The
On slot machines, the revenue for a casino from a player’s wager depends only on the choice of
slot machine. For table games, the net revenue for a casino from a player’s wager depends on the odds
4
Before discussing each part of our model in detail in the subsections that follow, a few caveats of
our modeling approach are noted. We will return to these issues in more detail in Section 6. First, given
that our data (described in Section 3) do not contain information about the casino operator’s promotion
activities1, we do not consider the effect of promotion/targeting on the part of the casino in our model.
Including such information may lead to better parameter estimates and improved prediction performance.
Second, we consider player data from only one national casino operator; because we have no information
on how players behave in other casinos, we do not model the “share of wallet” issue (Du et al. 2007).
Third, given data limitations (described later), we model all table games under one group and slot games
under another group; with more detailed data, it will be interesting to look at which specific table
We begin by modeling when a player visits a casino region and which region she visits. We
closely follow previous work on inter-purchase time models (e.g., Gupta 1991). Figure 2 illustrates the
Let the data collection time period be denoted by time t = 0 to t = T; let player i make Ni trips
during this time, and let the time span between the (k-1)’th visit and the k’th visit be wik . Note that the
inter-visit time before the first visit ( wi1 ) and after the last visit ( wi ( N i 1) ) are not fully observed as our
data is both left- and right-censored. We denote the observed inter-visit time before the first trip and after
the last one as wi*1 (= ti1 ) and wi*( Ni 1) (= T- tiNi ), respectively.
We use the competing risk framework (Klein and Moeschberger 2003) to model a player’s inter-
visit time and the choice of gambling region. We assume that for each player i and trip k, there are latent
inter-visit times associated with all J regions. The gamer chooses to visit the region that has the shortest
1
Our collaboration with the casino operator was at the corporate level. Thus, we had access to data at the corporate
level. However, the decision to give promotions to gamers is carried out at the individual casino level.
5
inter-visit time. Formally, for gamer i, region l and trip k, we assume that the latent inter-visit time ( silk )
is generated by an exponential distribution with ilk as the rate of visit. Let gamer i visit region zik =j on
an uncensored trip k ( k {2,..., N i } ). The likelihood for such an observation can be written as follows:
Here fexp denotes the exponential probability density function. Assuming that the visit to each
e
J
ijk wik ilk wik
lik ijk e = ijk exp wik ilk
.
[2]
l j l 1
Next, we focus on the first and last censored observations. Due to the memoryless property of the
exponential distribution, the likelihood of the first observation is also given by an expression similar to
J
li ( N i 1) exp wi*( N i 1) il ( N i 1) .
l 1
[3]
The likelihood of the entire data (i.e., for the Ni trips) available for player i can be written as:
Ni Ni J J
Li
ijk exp wik
k 1 ilk wi ( N i 1) il ( N i 1) ,
*
[4]
k 1 l 1 l 1
We link the rates for a trip to each region l to past win/loss using the following specification:
Here, il is the base rate of visit to region l and Rik 1 is a player- and trip-level covariate that captures
win/loss for player i from her most recent trip (i.e., trip k-1) to any region. Rik 1 takes value 1 if player i
wins money in her previous trip, and -1 otherwise. An inclusion of this variable allows a player’s win/loss
from past trips to influence the rate of visit for the current trip. The parameter is the sensitivity to this
past covariate – a positive coefficient indicates that past wins lead a player to have a shorter inter-visit
time while past losses lead her to have a longer inter-visit time.
6
2.2 Model of wagering behavior
As defined in the model of visit, player i visits region zik during her k’th trip. Let yik denote her
wager amount in that trip. This total amount is allocated between slots and table games. We denote the
amount wagered in slots as yikSLOT and the amount wagered in table games as yikTABLE . Further, the
As defined earlier, Rik 1 is a player- and trip-level covariate that captures win/loss for player i from her
most recent trip (i.e., trip k-1) to any region. The parameter is the sensitivity to this past covariate – a
positive coefficient indicates that past wins lead a player to wager more in the current trip while past
losses lead her to wager less. The parameters αi and z ik capture the gamer- and region-level effect
respectively on the total wager amount. Similarly, in Equation [8], which captures the proportion of wager
amount spent on slots, the term μi is a player-specific effect and z ik is a region-specific effect.
The term house advantage (or house edge) refers to the expected revenue by the casino per dollar
wagered by a player (Epstein 1997; Packel 2006). To illustrate, suppose the player wagers $y; then, the
E(x) = r * y, [9]
where E(x) denotes the revenue that the casino expects to make from the $y wager amount.
More specifically, consider a slot game called “21 Bell” in which three separate wheels, each
containing eight different symbols (7, Bar, Melon, Bell, Plum, Orange, Cherry and Lemon) are spun. If
7
the result of a $1 wager spin is any of the “winning combinations” (first column of Table 1), the player
We can calculate the associated house edge using the following expression:
The house advantage associated with the 21 Bell slot machine is 5.55%. This indicates that in the
long run, a player will tend to lose 5.55% of her total wagered amount.
The house edge may differ across slot machines. In some cases, different casinos may offer
differing house advantages even for the same slot machine. For the same 21 Bell slot machine, some
casinos may offer a different payoff table from that shown in Table 1. For instance, they may offer a
jackpot for the combination 777, which pays $500 (instead of $200 as shown in Table 1). In this case, the
In many table games, a gamer is required to make decisions during the course of the play. In such
games, the house advantage also depends on how good the player’s decisions are. For instance, consider
the game of blackjack, which is the most popular table game among players (Harrah’s Report 2006). In
blackjack, a player is dealt two cards in the beginning and then given the opportunity to select more. The
goal of a player is to beat the dealer by getting as close to a total of 21 without going over (referred to as
“busting”). After obtaining each card, the player is given the option of either “hitting” (i.e., getting an
additional card) or “standing” (i.e., not getting any additional card). As this game involves a player’s
decisions, the effective house advantage depends on whether the player hits/stands at the right situation. If
she makes each decision correctly, the house advantage is around 0.5% (Griffin 1999). However, a
deviation from the overall correct strategy results in a much larger house edge. In particular, if a “never
bust” strategy is followed (i.e., stand whenever a total of 12 or more is observed), the house advantage
rises to 3.9%. Thus, players with lower skill will incur a higher house advantage (as compared to players
8
with higher skill) and thereby give the casino higher expected revenues. This suggests that a
To summarize, the house advantage that a gamer faces depends on several factors. In slots, it
depends on the choice of game and the odds offered by a casino. In table games, in addition to these two
factors, the house edge also depends on player-specific skill. In the next subsection, we show how
To illustrate the nature of the net wins/losses data that the casino collects, consider the example
shown in Table 2.
Table 2 shows the within-trip data for a player (A) who visits Region 1. In this trip, he plays on
two different slot machines – one with a house advantage of 0.2 and another with a house advantage of
0.3 (column labeled as “Slot Machines/ House Advantage”). He wagers $100 in the former slot machine
and $200 in the latter machine. Thus, he bets a total of $300 over the entire trip. The column labeled as
“Expected Revenue” shows the revenue that the casino expects (theoretically) to make from this
particular combination of the wagered amounts and the house edges of the two chosen slot machines. The
casino then expects to make a total of $80 from this player’s betting in slots during the given trip. The
final column (“Actual Revenue”) is the actual revenue that the casino makes. Note that the actual revenue
may be either higher or lower than the expected revenue. In addition, the actual revenue may be positive
(the casino receives money) or negative (the player makes money). The net actual revenue that the casino
makes in this example from the player over the trip is $20 (= $40-$20). The last row labeled as
“Aggregated Information” is the information from this player summarized at the trip level — he wagers a
total of $300, the casino’s expected revenue is $80 and its total actual revenue is $20. This is part of
9
In our dataset (described in greater detail in a later section), we have aggregate trip-level
information for each player (e.g., $300, $80, $20) but do not have any intra-trip details (i.e., $100, 0.2,
$40; $200, 0.3, -$20). Thus, consistent with the information in the dataset, for a gamer i, during trip k and
chosen region zik, we employ an aggregate house advantage associated with slots ( rikSLOT ):
Given the above metric, the aggregate house advantage for player A in Table 2 is 0.27 (=
$80/$300). This aggregate house advantage can vary across players, trips and regions depending on the
We model the aggregate trip-level house advantage as a function of a person-level effect, which
captures differences in the choice of slot machines across players, and a region-level effect that captures
Let xikSLOT denote the actual net revenue to the casino from player i wagering in slots during the
k’th trip in region zik. Within our notation, xikSLOT > 0 indicates that the casino makes money, while
xikSLOT < 0 indicates that the casino loses money. Following Equation [9], the total trip-level expected
where yikSLOT is the amount wagered on slots and rikSLOT is the trip-level aggregate house advantage. Next,
10
given the skewness of the payoff in each slot wager, we approximate the total payoff as the expected
payoff (in Equation [13]) plus an error term which is skew-t distributed, as follows2:
xikSLOT yikSLOT rikSLOT X SLOT ikx SLOT , where ikxSLOT ~ Skew t. [14]
Table games involve active in-play decision making on the part of players. As described earlier,
the house advantage that each player receives may depend on her “skill” level. To model the skill level of
a player, we posit, consistent with industry practice, the existence of a prototypical or average (we use the
term interchangeably) gamer who has an average level of skill. Given the assortment of table games and
bets that a particular player selects to engage in during a trip, the aggregate house advantage for the
prototypical player can be calculated by combining the exact same assortment of games and bets with the
house edges that she would receive. We illustrate this notion with a hypothetical example.
Consider Table 3 that shows a particular assortment of wagered amounts in different table games
The column (“Prototypical Player House Advantage”) shows the house advantage that the
average player receives for this exact assortment of wagering amounts and chosen games. For the total
bet of $300, the “Expected Revenue” is calculated as before, i.e., Total Aggregate Expected Revenue is
$60 (=100*0.2 + 100*0.1 + 100*0.3). As discussed earlier, the casino keeps track of aggregate trip level
information for each player (e.g., $300, $60, $28) but not any intra-trip details (i.e., $100, 0.2, $40; $100,
0.1, $8; $100, 0.3, -$20). Thus, we denote the “Aggregate prototypical house advantage” in table games
based on the trip level information of player i (assortment of games and wagers) to region zik combined
with the house advantage received by the prototypical player as rikTABLE . This is calculated as follows:
2
We also modeled the total payoff using a normal distribution. Given the skewness in the payoffs, a skew t-
distribution (Jones and Faddy 2003), with fatter tails than the normal distribution and more skewed than the t-
distribution, gave a better fit.
11
Aggregate Trip-level Expected Revenue based on house edge received by an average player
rikTABLE
Aggregate Trip-level Wagered Amount in Tables for player i
[15]
Note that the aggregate prototypical house advantage for a player in table games is based on two
drivers: i) the selection of games and wagers of the player, and ii) the house edges received by an average
skilled player for that same assortment of games and wagers. For this example, rikTABLE is $60/$300 = 0.2.
This aggregate prototypical house advantage may vary across players, trips, and regions depending on the
types of table games offered in different regions and differing allocations to table games. We model
rikTABLE as follows:
In table games, players’ skill can alter the house advantage that they actually receive. We model
this aspect and specify the actual aggregate house advantage received by player i as follows:
Here, rikTABLE denotes the actual aggregated table house advantage received by player i in trip k,
and i represents his level of skill. We estimate this skill from our model. The parameter γi can range from
-∞ to +∞. If i is positive (negative) then the actual house edge for player i will be lower (higher) than
We use the Central Limit Theorem (see Appendix) to approximate the distribution of the revenue
that the casino wins/loses from gamer i in table games on a trip k ( xikTABLE ) by a normal distribution:
xikTABLE yikTABLE rikTABLE ikxTABLE , where ikxTABLE ~ N (0, yikTABLE X2 TABLE ). [18]
12
Equation [18] links the win/losses to actual aggregated table house advantage and, via Equation
[17], to the player-level skill. This completes the model for wagering decisions and how wager amount
Till now we focused on one player (player i) and showed the model specification. Next, we
incorporate heterogeneity across players. Let ln(λi) be { lni1 , lni 2 , ..., lniJ }, where il is the base
rate of player i’s visits to region l. We assume that all individual-level parameters in the model are drawn
from a common multivariate normal distribution - ln( λi ), i , i , iSLOT , iTABLE , i ~ N ( IND , IND ) . The use
of this multivariate prior distribution achieves two objectives. First, the gamer-specific base rate
parameters across regions are allowed to be correlated. Second, the parameters in the model of visits (visit
frequency) and wager/wins are linked. The estimation results will indicate to the strength of this linkage.
Similarly, the region-level parameters are drawn from a different multivariate normal
distribution, l , l , lSLOT , lTABLE ~ N ( REG , REG ) . For model identification, both vector REG and the
sum of each of the parameters across the J regions ( l , l , lSLOT and lTABLE ) are set to 0. We
l l l l
adopt a hierarchical Bayesian framework for simulation-based inference using MCMC methods. The
priors for the unknowns and the full conditionals are available from the authors upon request.
As our model contains past win/loss covariates, a simulation procedure is required to obtain
model-based predictions of various metrics of interest, such as wager amounts, and the associated
wins/losses. The prediction of total expected revenue, which accrues from a gamer, incorporates all the
intermediate predictions of how many times a gamer visits a region, how much she wagers and her
13
For a time period T*, the casinos’ expected revenue that player i will generate for region l
N i (T * )
Revil (T )
*
x
k 1
k
I ,
ik {l } [19]
where, N i (T *) denotes the number of times player i visits any of the J regions during the period T* , xil
is the total revenue generated from both table games and slots in region l and I{kl} is an indicator function
For each player and each posterior draw of her parameters from the MCMC chain, we generate a
total of 200 paths where a path for the player represents her visits to the various regions during the period
T* and the associated wagers/wins during each visit. Using the simulated wins/losses in each trip, we also
generate the indicator variable that captures the effect of past wins/losses on gamer behavior during the
next trip.
We generate 200 such paths for each gamer and each sample of her individual-specific
parameters. We then average over all 200 paths. Finally, we incorporate the uncertainty in the individual-
specific parameters by averaging the probabilities over all sampled values of the parameters from the
MCMC chain.
3. Data
Our dataset is provided by a major casino operator in the United States. It contains records of
1571 customers over a 28-month period from December 2004 to April 2007. During this time period,
these customers made a total of 8641 trips, where a trip is a visit to a casino region. For each player-trip
combination, we have information on when the trip takes place, the visited region, and for both slots and
table games separately, the total amount that the gamer wagers, her trip-level aggregate house advantage
and the net actual amount that she wins/loses. For a subset of 400 players, we also have demographics
data (age and gender) along with the above discussed transactional information.
14
Table 4 shows the overall data aggregated at the region level. For reasons of confidentiality, we
cannot name these regions but these are regions such as Las Vegas and Atlantic City. For each of the
seven regions, the table shows the number of observed visits, the total wagered amount, and the total
Table 5 shows the summary statistics of the data for the players. There is substantial variability in
terms of both visit frequencies and wagering behavior. The number of visits that a player makes varies
from 1 to 122 over the 28-month period, with an average across players of around 6 visits. In terms of
wagering behavior, the mean aggregate amount wagered in slots (tables) across all the trips per player is
around $37,800 ($1000). Further, not reported in the table, gamers tend to lose around 5.6% (17.6%) of
A closer investigation reveals that about 40% of the players in the data (635 players) make a
single visit during the data period. For the remaining players, the mean inter-visit time is 165 days
(around 5 months). More than 25% of these players visit more than one region. This percentage of
customers making cross-region trips is consistent with industry reports (see CIO magazine,
http://www.cio.com.au/index.php/id;609782439).
4. Empirical Results
We begin with a comparison of four model specifications based on their log-marginal likelihood
(LML). We compare our proposed model (Model 1) with three nested versions – Model 2 ignores past
covariates (i.e., κ = υ = 0) but includes the construct of latent skill, Model 3 ignores the skill level (i.e., i
= 0) but allows for past covariates and Model 4 excludes both past covariates and skill.
15
4.1 Model comparison
We use the MCMC draws from the simulation-based inference to calculate the LML for each
model. Higher LML denotes a better model. The LML for Model 1 is -62139.2, for Model 2 is -62259.0,
for Model 3 is -62614.0 and for Model 4 is -67095.0. Thus, Model 1 is best supported by the data based
on the criterion described in Kass and Raftery (1995). A comparison of the log-marginal likelihoods
shows that inclusion of both the construct of latent player skill and past covariates are important within
We further compare all four models on the percentage differences between the fitted values from
our model for the various metrics of interest aggregated over all players and regions and the observed
values. The fitted values are obtained by applying the simulation procedure described in Section 2.7 in-
sample. Table 6 shows these comparisons and indicates that our proposed model best fits the data. The
results also show that past covariates and latent skill contribute to the model in distinct ways. An
inclusion of past covariates helps in improving the fit of the wager amount, which in turn has an effect on
revenues from both slots and table games. An inclusion of the latent skill level improves the fit of the
revenues from table games. This latter result has face validity as player skill only affects his revenues
from table games and does not directly influence his revenues from slots.3
We begin by examining the player-level parameters ( IND ) for our proposed model.4 Table 7
shows the estimated parameters together with their 95% posterior interval. As expected, the parameter
related to amount of wagering ( ) is positive. The coefficient capturing the proportion of wager spent on
slots ( ) is positive as well and implies that, on average, players allocate more of their wagering amount
3
We also validated our proposed model by comparing the model-fitted wagers and revenues (for both slots and table
games) with their observed values. We found that our model provided a good fit to the data – the observed values
were within the 95% posterior intervals around the model-fitted values. Due to space constraints, we do not report
these results. They are available from the authors upon request.
4
Due to space constraints, we do not report the details on the region-level parameters for our proposed model. These
estimates as well as those for Models 2, 3 and 4 are available from the authors upon request.
16
to slots than to table games. We find that the average skill level ( ) is marginally negative. This suggests
that on average, gamers in our database incur a higher house advantage than the prototypical player. In
addition, after controlling for region-level effects, the average (across players) house advantage within
slot machines ( SLOT ) is around 0.10 (i.e., 10%) while the house advantage faced by the prototypical
player in table games ( TABLE ) is about 0.18 (i.e., 18%). These findings are also consistent with industry
reports on the average house advantage of around 20.5% (Vogel 2007). As the logarithm of player-level
rate parameters for the seven regions is normally distributed across players (see Section 2.6), we report its
population mean for all the seven regions. Finally, not reported in the table, we find that parameters
and , which capture the effect of past covariates, are positive and significant. The parameter is
estimated to be 0.22 with a 95% posterior interval of (0.19, 0.26) while the parameter is estimated to be
0.13 with a 95% posterior interval of (0.09, 0.17). These results suggest that past wins influence gamers
to both visit a casino faster (i.e., decrease their inter-visit time) and wager more.
Table 8 shows the estimated correlations among the player-level parameters. We find a positive
but low correlation (ρ = 0.25) between wager amount ( i ) and skill level ( i ).5 This indicates that an
observable player-level characteristic such as wagering amount is not a perfect indicator of player skill.
We also find a moderate level of correlation among several of the visit parameters ( i1 ,..., i 7 ) across
different regions. For instance, the correlation between i 3 and i 7 is 0.45 and the correlation between i 2
and i 5 is -0.56. These correlations have face validity as Region 3 and Region 7 are geographically close
to each other and hence a gamer might consider visiting both regions. Region 2 and Region 5 are far apart.
This indicates that there is dependence in gamers’ visit patterns across regions. Finally, in general, the
individual-level parameters associated with the model of wagering behavior and winning are weakly
correlated with the visit parameters. For instance, only the visit parameter to Region 1 ( i1 ) is moderately
5
Cohen (1988) suggests that a correlation of less than 0.3 is weak, between 0.3 to 0.5 is moderate and greater than
0.5 is strong.
17
correlated with the wager amount (ρ= 0.38) while the remaining show weak correlation. Similarly, the
strongest correlation between the visit parameters and skill level is only 0.28 (this is for visit parameter to
Region 6). This suggests that a player’s frequency of trips is also not a good metric for capturing player
skill.
These findings indicate that i) there is dependence in gamers’ patterns across regions and ii)
observable player characteristics such as frequency of trips and wagering amount are not perfect
5. Managerial implications
An important task for casinos is to target players based on their predicted revenue.6 For making
the predictions, we divide our data time period into two equal halves, and use only the first half of the
data (14 months) to calibrate our proposed model. We then create three holdout datasets of 5, 10 and 14
months respectively. We compare our model-based out-of-sample predictions with those from a
reasonable alternative model that uses observable player-level revenue within the calibration data and
assumes that a proportional level of revenue will be present in the holdout dataset. The proportionality
constant is dependent on the number of months in the holdout vis-à-vis the 14 months in the calibration.
For instance, for the holdout dataset of 5 months, the alternative model assumes that the level of revenue
for each player will be 0.357 (= 5/14) times the revenue in the calibration dataset.7 Similarly, for the
holdout dataset of 14 months (the same length as the calibration dataset), the alternative model assumes
that the level of revenue for each player will be the same as that in the calibration dataset.
6
In several conversations with the management of the casino we have been informed that they use predicted future
revenue of players for targeting purposes.
7
We also used a different methodology to make revenue predictions for the alternative model. For the holdout
dataset with 5 (10) months, we assumed that the level of revenue for each player was the same as present in the last
5 (10) months of the calibration dataset. The predictions from the alternative model based on this assumption were
worse than the reported predictions. These results are available from the authors upon request.
18
We perform three comparisons between the two model-based predictions based on i) predicted
ranking of players, ii) the root mean squared error (RMSE), and iii) Lorenz curves (Gastwirth 1972). For
the rank based comparison, we calculate the goodness-of-fit of the ranking system using Spearman’s rank
correlation coefficient, which ranges from +1 (perfect agreement to actual ranking) to -1 (perfect
disagreement to actual ranking). For the RMSE calculation, in the proposed model, we predict the total
revenue from each player across all the seven regions and then compare it with its observed value. (For
the alternative model, please see the above description of how we predict player-level revenues.) A
Lorenz curve is a graphical representation that allows a comparison of the percentage of one variable
against the percentage of another (Gastwirth 1972). Of particular interest to casino managers is the
percentage of total revenue that the casino collects from the top 10 %, 20 %, 50 % (and so on) of all
players. We determine player rank either by (i) our full model, or (ii) alternative measure such as past
cumulative revenue.
Table 9 shows the comparison between the two models based on Spearman’s correlation and
The ranking based on our model is better than the ranking based on the alternative model. For
instance, when the holdout dataset is 14 months, the ranking based on our model has a Spearman
correlation of 0.59 while those based on the alternative model has a lower Spearman correlation of 0.43.
We also use other observable statistics such as frequency of visits, wager amounts and other combinations
to develop player ranking. The ranking system based on our model is better (i.e., has a higher Spearman
correlation) than the ranking systems based on any of these observable statistics. In particular, for the
holdout dataset of 14 months, the ranking system based on visit frequency had a correlation of 0.36 and
the ranking based on frequency and wagered amount had a correlation of 0.46. The results of other
combinations are available upon request. All combinations indicate that predictions based on observable
19
We find another indication for the superiority of our full model when we compare the RMSE
across the two models. For each holdout dataset, we find that the RMSE from our model is smaller than
that from the alternative model. For instance, when the holdout dataset is of 14 months, the RMSE of our
model is 5556.23 while that of the alternative model is 6261.73. This difference indicates about 11%
We make a third comparison between the two models based on their Lorenz curves. Figures 3a-3c
show Lorenz curves that compare our model-based predictions with those from the alternative model for
each of the three holdout datasets. An “ideal” Lorenz curve would be one which starts out flat (the gamers
who are predicted to provide low revenue actually generate low level of revenue) and with an increase in
predicted revenue, the model starts to capture a greater percentage of the observed revenue. Thus, the
farther the Lorenz curve for a model is from the 45 degree line, the better the model is in targeting high
revenue players. The area between the 45-degree line and the Lorenz curve can be interpreted as the
As can be seen in the figures, for each of the three holdout datasets, the Lorenz curve for our
proposed model (the dotted line) lies below the Lorenz curve for the alternative model (the solid black
line). For the dataset with 5 (10) months holdout, the area between the 45-degree line and the Lorenz
curve of our model is 0.442 (0.432), which is larger than that of the benchmark model of 0.402 (0.398). In
the holdout dataset of 14 months, the area between the 45-degree line and the Lorenz curve of our model
is 0.427, which is larger than that of the benchmark model (0.397). These results indicate that our model
All three comparisons indicate that our model provides better individual-level revenue predictions
compared to the predictions based on the alternative model, highlighting the managerial importance of
20
5.2 Demographic characterization
As stated in the introduction, another important managerial goal is to understand how player-level
demographics relate to various behavioral aspects such as amount of wagering, skill level and others.
To study how demographics information is related to our model parameters, we re-estimate our
proposed model for a subset of 400 players (as described in Section 3.1) and include their demographics
(age, gender) in the heterogeneity specification across players.8 We assume that the player-specific
ln( λ ), , ,
i i i i
SLOT
, iTABLE , i ~ N (Z
i
IND
, IND ) where Zi contains player demographics. This is a typical
way of specifying heterogeneity across players with demographics (Rossi and Allenby 2003).
We estimate our model and find several interesting results, which are shown in Table 10. We find
that older players wager more and are inclined to allocate more of their wagers on slot machines vis-à-vis
table games. We also find that women have a tendency to wager more than men and, compared to men,
they allocate more of their wagers to slots versus table games. We also find some directional, though not
significant, evidence that men have higher level of skill than women. Finally, there is no evidence that age
or gender affects the rate of visit to a region (with the only exception being Region 3 where we find that
older players are more likely to visit).9 Having additional data with respect to demographics such as home
address, profession, level of education, relationship status might yield finer differences and it certainly
8
Due to confidentiality reasons, the casino operator did not feel comfortable sharing individual-level demographics
information for more than 400 players.
9
To examine further the usefulness of having the player-level age/gender demographics, we also estimated a model
without demographics on the data from the 400 players. We used the MCMC draws from the simulation-based
inference to calculate the LML for the two models. The LML for the model with demographics was -39610.60 while
for the model without demographics was -39622. In comparing the player-level revenue predictions from both
models (with/without demographics) with the observed player-level revenue, we found that the Spearman rank
correlation between the predicted player ranks based on the model with (without) demographics and actual player
ranking was 0.87 (0.85), the RMSE from the model with and without demographics, across all seven regions, were
very similar (around 12357), and the areas under the Lorenz curve for the model with (without) demographics were
0.359 (0.357).
21
6. Conclusions and Future Research
We developed a new model for gamers’ behavior in a casino setting and incorporated various
context-specific features. Our integrated framework contains several components: a model for wager
amount and net wins/losses in slots and table games, a model for the proportion that he wagers on slots
and table games respectively, a description of the odds offered by different regions, a specification of
player’s latent skill level and the model for visit and choice of region.
The results indicate that our modeling approach is quite useful. At the aggregate level, it provides
a very good fit to the data in terms of region-level wager and revenue amounts in slot machines and table
games as well as in terms of the total number of gamer visits. At the individual level, the integrated model
provides better evaluation of future player revenue vis-à-vis more traditional metrics.
Taking a boarder perspective, our research is related to other studies in entertainment industry.
While most past research in marketing that focuses on the entertainment industry addresses the motion
picture and home entertainment sectors (Eliashberg et al. 2006), some attempts worth noting in other
sectors include Broadway shows (Reddy et al. 1998) and theater performance (Putler and Lele 2003). Our
contribution in this paper is also to a non-motion picture sector -- the gaming sector.
With some modifications, our modeling approach can be applied to other marketing contexts. For
instance, in the context of salesforce management, a salesperson’s compensation is based on his/her sales
performance. It is unclear how much of the sales are due to product attributes (e.g., a product being good
sells by itself) and how much is due to his/her latent skill level in convincing buyers. An estimation of
such skill thus becomes vital for capturing true differences in performance across salespeople. Another
context is the area of hotel revenue management. Hotels are interested in predicting the amount of
revenue they would receive from each customer. However, customers differ in their usage of services
such as mini bar, business center, room internet service and others, which affect the overall revenue
received from them. Clearly, for accurate prediction of revenues it is important to capture such underlying
differences in service usage among customers. A similar problem occurs for other services such as car
22
rentals where customers differ in their usage of services such as in car navigation system, satellite radio,
The gaming industry offers modeling challenges as well as a large volume of rich data that allow
modelers to generate important managerial insights. The current analysis is a first step in that direction.
Our model can be further extended, however, by considering the following future research opportunities:
(i) Share of wallet issues: For targeting customers, most casinos are also interested in
how much of the overall revenues it is getting from a player’s gambling budget, i.e.,
what is its share of a player’s gambling wallet? There is some recent work (Du et al.
2007) which describes how a secondary dataset such as the one we used can be
augmented with primary data from customers collected using surveys to make
predictions about share of wallet. In the current context, a survey-based primary data
from players, in conjunction with our model, can be used for both revenue and share
of revenue predictions.
(ii) Promotional activities: Typically, casinos attract gamers by sending out a diverse
array of promotions (e.g., complimentary hotel stays/meals, cash backs) and giving
them compensations while they visit a casino. We did not have data on promotions
on the type of customers who are induced to visit (high skilled / low skilled) and their
wagering behavior.
possibility is to model skill as a dynamic construct, which may change over time
based on a player’s experience with a certain game. For example, a player may
become more skilled with blackjack as she becomes more experienced with the game.
(iv) Consumer demographics: In this study, we only examined the role of age and gender.
23
(v) Casino floor design: Casino industry managers are also interested in issues such as
how best to configure their games on the casino floor (Bayus and Gupta 1985), the
effect of environmental cues on gaming behavior (Griffiths and Parke 2003) and the
the per capita wagering on slot machines (Thalheimer and Ali 2003). By incorporating
such casino-level covariates into our model, we can provide some insights into casino
floor planning.
24
References
Abt, Vicki, James F. Smith and Eugene M. Christiansen (1985), The Business of Risk: Commercial
Gambling in Mainstream America, Lawrence: University Press of Kansas
American Gaming Association (2006), Industry Information, available at
http://www.americangaming.org/Industry/factsheets/statistics_detail.cfv?id=7, last accessed on
February 3, 2009.
Baker, Robert G.V. and David C. Marshall (2005), “Modeling Gambling Time and Economic
Assignments to Weekly Trip Behavior to Gambling Venues,” Journal of Geographical Systems, 7,
381-402.
Bayus, Barry L. and Shiv K. Gupta (1985), “Analyzing Floor Configurations for Casino Slot Machines,”
Omega, 13, 6, 561-567.
Becker, Gary S. and Kevin M. Murphy (1988), “A Theory of Rational Addiction,” Journal of Political
Economy, 96 (4), 675-700.
Cargill, Thomas F. and William R. Eadington (1978), “Nevada’s Gaming Revenues: Time Characteristics
and Forecasting,” Management Science, 24, 12, 1221-1230.
Clotfelter, Charles T. and Philip J. Cook (1993), “The Gambler’s Fallacy in Lottery Play,” Management
Science, 39 (12), 1521-1525.
Cohen, Jacob (1988), Statistical Power Analysis for the Behavioral Sciences, New Jersey: Lawrence
Erlbaum Ass.
Connolly, John and Brandon Rogoff (2008), “Keeping Track of U.S. Mall Visits,” Research Review, 15,
2, 5-9. Available at
http://www.icsc.org/srch/rsrch/researchquarterly/current/rr2008152/US%20Mall%20Visits.pdf,
last accessed February 9, 2009.
Cotte, June and Kathryn A. Latour (2009), “Blackjack in the Kitchen: Understanding Online versus
Casino Gambling,” Journal of Consumer Research, 35, 742-758.
Du, Rex Y., Wagner A. Kamakura and Carl F. Mela (2007), “Size and Share of Customer Wallet,”
Journal of Marketing, 71 (April), 94-113.
Eliashberg, Jehoshua, Anita Elberse and Mark A. A. M. Leenders (2006), “The Motion Picture Industry:
Critical Issues in Practice, Current Research and New Research Directions,” Marketing Science,
25, 6, 638-661.
Epstein, Richard A. (1997), The Theory of Gambling and Statistical Logic, Revised Edition, Elsevier
Science.
Gastwirth, Joseph L. (1972), “The Estimation of the Lorenz Curve and Gini Index,” The Review of
Economics and Statistics, 54, 306-316.
Griffin, Peter (1999), The Theory of Blackjack, 6th Edition, Huntington Press.
Griffiths, Mark D. and Jonathan Parke (2003), “The Environmental Psychology of Gambling,” In G.
Reith (Ed.), Gambling: Who Wins? Who Loses? 277-292, Prometheus, New York.
Gupta, Sunil (1988), “Impact of Sales Promotions on When, What and How much to Buy,” Journal of
Marketing Research, 25, 4, 342-355.
Gupta, Sunil (1991), “Stochastic Models of Interpurchase Time with Time-Dependent Covariates,”
Journal of Marketing Research, 28, 1, 1-15.
25
Guryan, Jonathan and Melissa S. Kearney (2008), “Gambling at Lucky Stores: Empirical Evidence from
State Lottery Sales,” American Economic Review, Forthcoming.
Harrah’s Report (2006), Profile of the American Gambler, available at http://www.harrahs.com/harrahs-
corporate/about-us-profile-of-gambler.html, last accessed on February 3, 2009..
Helsen, Kristiaan and David C. Schmittlein (1993), “Analyzing Duration Times in Marketing: Evidence
for the Effectiveness of Hazard Rate Models,” Marketing Science, 12 (4), 395-414.
Jones, M.C. and M.J. Faddy (2003), “A Skew Extension of the t-distribution, With Applications,” Journal
of the Royal Statistical Society, Series B, 65, 1, 159-174.
Kallick, M., D. Suits, T. Dielman and J. Hybels (1979), A Survey of American Gambling Attitudes and
Behavior, Research Report Series, Ann Arbor, MI: University of Michigan, Institute for Social
Research.
Kass, Robert E. and Adrian Raftery (1995), “Bayes Factors,” Journal of the American Statistical
Association, 90, 773-795.
Klein, John P. and Melvin L. Moeschberger (2003), Survival Analysis: Techniques for Censored and
Truncated Data, 2nd Edition, Springer, New York, NY.
Morrison, Donald G. and David C. Schmittlein (1981), “Predicting Future Random Events Based on Past
Performance,” Management Science, 27 (September), 1006-1023.
Morse, Edward A. and Ernest Goss (2007), Governing Fortune: Casino Gambling in America, University
of Michigan Press.
Motion Pictures Association of America (2009), Research and Statistics, available at
http://www.mpaa.org/researchStatistics.asp, last accessed on February 6, 2009.
NationMaster (2009), Cinema Attendance by Country, available at
http://www.nationmaster.com/graph/med_cin_att_percap-media-cinema-attendance-per-capita,
last accessed on February 6, 2009.
Nichols, Mark W. (1998), “The Impact of Deregulation on Casino Win in Atlantic City,” Review of
Industrial Organization, 13, 713-726.
Packel, Edward (2006), The Mathematics of Games and Gambling, 2nd Edition, The Mathematical
Association of American.
Putler, Daniel S. and Shilpa Lele (2003), “An Easily Implemented Framework for Forecasting Ticket
Sales to Performing Arts Events,” Marketing Letters, 14, 4, 307-320.
Reddy, Srinivas K., Vanitha Swaminathan and Carol M. Motley (1998), “Exploring the Determinants of
Broadway Show Success,” Journal of Marketing Research, 35, 370-383.
Rossi, Peter E. and Greg M. Allenby (2003), “Bayesian Statistics and Marketing,” Marketing Science, 22,
3, 304-328.
Stigler, George J. and Gary S. Becker (1977), “De Gustibus Non Est Disputandum,” American Economic
Review, 67 (2), 76-90.
Swami, Sanjeev, Jehoshua Eliashberg and Charles B. Weinberg (1999), “SilverScreener: A Modeling
Approach to Movie Screens Management,” Marketing Science, 18, 3, 352-372.
Thalheimer, Richard and Mukhtar M. Ali (2003), “The Demand for Casino Gambling,” Applied
Economics, 35, 907-918.
Vogel, Harold L. (2007), Entertainment Industry Economics: A Guide for Financial Analysis, 7th Edition,
Cambridge University Press, New York, NY, 353-419.
26
Figure 1: Overview of our modeling framework.
Timeline
wi*1 *
wiN i 1
0 T tiN i 1
ti0 ti1 ti2 tiN i 1 t iN i
27
Figure 3a: (5 months holdout) Lorenz curves for our model (dotted) and alternative model (solid black).
Figure 3b: (10 months holdout) Lorenz curves for our model (dotted) and alternative model (solid black).
Figure 3c: (14 months holdout) Lorenz curves for our model (dotted) and alternative model (solid black).
28
Combination Payoff (zi) (in $) Probability (pi) Payoff x Probability (pi x zi) (in $)
3 sevens 200 0.000125 0.025
3 bars 100 0.00075 0.075
3 melons 100 0.001 0.1
Melon, melon, bar 100 0.0005 0.05
3 bells 18 0.005 0.09
Bell, bell, bar 18 0.000625 0.01125
3 plums 14 0.007875 0.11025
Plum, plum, bar 14 0.002625 0.03675
3 oranges 10 0.0125 0.125
Orange, orange, bar 10 0.003125 0.03125
Cherry, cherry, any 5 0.03 0.15
Cherry, any, any 2 0.07 0.14
Expected payoff 0.9445
Aggregated 300 80 20
Information
Aggregated 300 60 28
Information
Table 3: Intra-trip data within table games for player i visiting region 1.
29
Region Number of Total Wager Total Revenue Wager Amount Average Revenue
Visits Amount ($, 000) ($, 000) Per Visit ($) Per Visit ($)
R1 745 17230.2 557.5 23127.8 748.3
R2 1841 11746 944.6 6380.2 513.1
R3 315 4437.6 193.3 14087.7 613.5
R4 438 1601.9 136.6 3657.2 311.9
R5 2831 13359.7 964.6 4719.1 340.7
R6 1888 8783.7 553.1 4652.4 293
R7 583 3919.4 216 6722.8 370.5
Table 4: Summary statistics of the dataset by each region.
Table 6: Model fit. For each model, the numbers denote the % difference between the fitted value and the
observed values.
30
95% Posterior
Parameter Mean Interval
Wager Amount (α) 6.10 (5.99, 6.19)
Proportion on Slots (μ) 4.40 (4.21, 4.59)
House Edge in Slots (τSLOT) 0.10 (0.09, 0.10)
TABLE
House Edge in Tables (τ ) 0.18 (0.16, 0.20)
Skill Level (γ) -0.13 (-0.24, -0.02)
Rate - region 1 -8.54 (-8.68, -8.38)
Rate - region 2 -9.09 (-9.24, -8.93)
Rate - region 3 -11.03 (-11.54, -10.62)
Rate - region 4 -10.94 (-11.40, -10.55)
Rate - region 5 -8.86 (-9.09, -8.66)
Rate - region 6 -8.69 (-8.82, -8.55)
Rate - region 7 -12.13 (-15.53, -11.39)
Parameter 1 2 3 4 5 6 7 8 9 10 11 12
1) Wager Amount (α) 1.00
2) Proportion on Slots (μ) 0.41 1.00
SLOT
3) House Edge in Slots (τ ) -0.04 0.14 1.00
TABLE
4) House Edge in Tables (τ ) 0.06 0.07 0.02 1.00
5) Skill Level (γ) 0.25 0.06 0.01 0.10 1.00
6) Visit Rate - Region 1 (λ1) 0.38 -0.02 -0.07 -0.06 -0.04 1.00
7) Visit Rate - Region 2 (λ2) -0.04 0.02 0.02 0.04 -0.06 -0.05 1.00
8) Visit Rate - Region 3 (λ3) 0.17 0.03 -0.06 -0.08 -0.20 0.30 -0.23 1.00
9) Visit Rate - Region 4 (λ4) 0.27 0.27 -0.02 -0.01 0.09 -0.18 -0.07 0.10 1.00
10) Visit Rate - Region 5 (λ5) 0.19 0.13 -0.03 0.04 -0.01 0.03 -0.56 -0.03 -0.03 1.00
11) Visit Rate - Region 6 (λ6) 0.21 0.16 0.07 0.01 0.28 -0.17 -0.46 -0.23 0.03 -0.13 1.00
12) Visit Rate - Region 7 (λ7) 0.16 0.13 0.03 0.03 0.02 0.16 -0.13 0.45 -0.41 -0.06 -0.02 1.00
31
Holdout - 5 months Holdout - 10 months Holdout - 14 months
Proposed Alternative Proposed Alternative Proposed Alternative
Comparison Metrics Model Model Model Model Model Model
Spearman Correlation 0.45 0.34 0.54 0.41 0.59 0.43
Root Mean
Squared Error (RMSE) 2622.8 2840.74 5171.61 5680.91 5556.23 6261.73
Table 9: Comparison of the model-based predictions from the proposed model and alternative model.
32
Appendix
Let X be the total net amount of loss by the gambler, Y be the total amount of wager. Let the expected net
value of a unit (dollar) wager be Zi = 1-r (i.e., r denotes the house edge), and let var(Z i ) 2 .
derived as follows:
Y
E ( X ) Y E ( Z i ) Y Y (1 r ) Yr
i 1
Y
Var ( X ) Var ( Z i ) Y 2
i 1
Since X is the sum of a series i.i.d. random variables, its density converges (in distribution) to a normal
33