Mar 28, 2016
Qualex Consulting Services, Inc4300 Biscayne Bay, Suite 203Miami, FL 33137.
www.qlx.com
Copyright © 2012 Clive J. Pearson. All rights reserved.
No part of this book may be reproduced, stored in a retrieval system, or transmitted by any means without the written permission of the author, excepting brief quotes used in reviews.
First published by Qualex Consulting Services on 4/15/12.
First Printing: April 15, 2012
Printed in the United States of America.
Cataloging data may be obtained from the Library of Congress
ISBN 978-0-557-73420-7
CHAPTER TWO
PREDICTIVE ANALYTICS
“In business, as in baseball, the question isn't whether or not you'll jump into analytics. The question is when. Do you want to ride the analytics horse to profitability...or follow it with a shovel?”
~ Rob Neyer, ESPN
Overview
Predictive analytics refers to a variety of statistical techniques that
analyze current and historical facts to make predictions about future
events. Using such techniques as predictive modeling, machine
learning, data mining and game theory, predictive analytics can build
models that exploit patterns found in historical and transactional data,
patterns that can identify business risks and potential opportunities.
Predictive analytics is not a new technology, it is a decades
old, proven technology that encompasses such disciplines as statistics
and data mining. A forward-looking technology that uses past events to
predict future activity, predictive analytics arose out of the management
information systems and standard reporting world of the 1970s. The
drill down technology of the 1980s led to the data warehousing and
OLAP cubes systems of the 1990s, which allowed for more complex ad
hoc data querying. The analytics and Business Intelligence solutions of
the 2000s evolved into the more complex world of predictive modeling
and optimization of today. These systems can do more than report, they
37
can actually help predict future business activity. The Predictive
Analytics Pocket Guide4 (2009) defines predictive analytics as such:
Unlike Business Intelligence applications, which
merely present summaries of historical data, predictive
analytics focuses instead on the prediction of future
outcome of events not yet observed in the data. These
predictions are made by creating a model from the
observed data using statistical techniques. These
models can range in complexity from simple linear
equations to powerful techniques such as neural
networks5.
Predictive analytics can be broken down into three different
types of models:
1. Predictive: these analyze past performance to predict the
likelihood that an individual customer will exhibit a specific
behavior in the future.
2. Descriptive: these identify different relationships between
customers to group or segment them for marketing or other
purposes.
3. Decision: these predict outcomes of complex decisions,
relationships, products and/or processes.
Predictive analytics extracts information from data sets and
uses it to anticipate future trends and behavior patterns based on
statistics and data mining (Ramakrishnan and Madure, 2008). The most
4Available at www.predictivesource.com.
5A predictive analytics technology that can learn the relationship between inputs and output through training.
38
important element of predictive analytics is the predictor, “a variable
that can be measured for an individual or other entity to foresee future
behavior” (Ramakrishnan and Madure, 2008). The real trick is to find
the predictive model best suited for the outcome one is trying to study
(Ramakrishnan and Madure, 2008) and this is no easy feat.
“Predictive analytics also encompasses models that seek out
subtle data patterns to answer questions about customer performance,
such as churn prediction, fraud detection and propensity to buy
additional products and services” (Ramakrishnan and Madure, 2008).
Predictive analytics solutions include SAS's suite of analytics products,
IBM's SPSS, EMC's Greenplum and the Revolution's R open source
product. Whichever solution is used, predictive analytics can enhance
customer acquisition and retention, identify cross-sell and up-sell
opportunities, identify customer lifetime value, spot fraud detection,
determine the life cycle of a slot machine and help direct and improve
marketing campaigns. Predictive analytics can even “perform
calculations during live transactions to guide a decision”
(Ramakrishnan and Madure, 2008), but without data mining predictive
analytics would be useless.
Data Mining: An In-House Goldmine
Data mining – the process whereby hidden patterns within data sets are
discovered – is a component of predictive analytics that entails an
analysis of data to identify trends and patterns of relationships among
data sets (Ramakrishnan and Madure, 2008). To put is simply, data
mining helps transform raw data into usable information. In their article
Neural Networks in Data Mining, Singh and Chauhan (2009) state that
data mining is the:
39
business of answering questions that you've not
asked yet. Data mining reaches deep into databases.
Data mining tasks can be classified into two
categories: Descriptive and predictive data mining.
Descriptive data mining provides information to
understand what is happening inside the data without
a predetermined idea. Predictive data mining allows
the user to submit records with unknown field values,
and the system will guess the unknown values based
on previous patterns discovered in the database.
By employing automated predictive analytics to sift through a
casino operator’s customer database, data mining can discover hidden
opportunities and connections that might otherwise be missed. Many
casino operators have terabytes and terabytes of data – everything from
customer player card information to information about a customer’s
room preference – and sifting through this information to discover
meaningful connections would be an impossible task without data
mining.
Data mining and predictive analytics aim to identify valid,
novel, potentially useful and understandable correlations and patterns in
datasets (Chung & Gray, 1999) by combing through copious amounts of
data to sniff out patterns and relationships that are too subtle or complex
for humans to detect (Kreuze, 2001). Data must be gathered from
disparate sources and then seamlessly integrated into a data warehouse
that can then cleanse it and make it ready for consumption. Trends that
surface from the data mining process can help in monetization, as well
as in future advertising and marketing campaigns.
For casinos, data mining can cull through data from such
disparate sources and departments as sales and marketing, thereby
allowing users to measure patron behavior on more than a hundred
40
different attributes, which is a far cry from the three or four different
attributes that statistical modeling used to offer.
Unlike traditional statistical analysis, which relies heavily on
hypothesis testing, data mining tries to identify relationships and
interdependencies that affect a marketing-related opportunity or
problem (Thelen, et al., 2004). While traditional multiple regression
methods can only use a limited number of complexity levels, neural
networks and decision trees can easily handle up to 200 predictor
variables (Thelen, et al., 2004), allowing them to do much more
complicated computations.
Normally, with statistical modeling, an analyst poses a simple
question such as: “Are higher-income people prone to be more loyal to
a casino player card than those with lower income levels?” The
hypothesis would elicit two responses, either “yes” or “no.” Data
mining, however, can reveal factors that contribute to casino loyalty;
factors that the analyst might never have thought to test for. According
to Thelen, et al. (2004), the data mining process is as follows:
1. Identify the business opportunity
2. Cleanse the data
3. Transform the data into meaningful information
4. Confirm the model
5. Tweak and perfect the model
Since data mining systems are inherently reliant on so many
departments, they can be difficult and complicated to implement.
Marketing managers, corporate strategists, statisticians and IT directors
are all required to add their input. Casino operators should keep in
mind, however, that data mining will only be successful if their casino
patrons are willing to provide information on themselves (Thelen, et al.,
2004). Although player cards provide a wealth of information, if the
41
patron doesn’t trust the casino with information beyond what is gleaned
from the player cards, the casino will only have a incomplete view of
that individual patron (Thelen, et al., 2004), which severely limits the
predictive analytics capabilities.
Data Mining Techniques
Regression models: Regression analysis is the process of predicting
the continuous dependent variable from a number of independent
variables. It attempts to find a function which models the data with the
least error. Regression analysis can be used on data which is either
continuous or dichotomous, but cannot be used to determine a causal
relationship. Regression analysis focuses on establishing a
mathematical equation as a model to represent the interactions between
the different variables under consideration. Regression models are
particularly effective to find patron worth because the model can be used
to score historical data to predict an unknown outcome (Sutton, 2011).
Multiple regression models “utilize a variety of predictors and the
relationships between those predictors to predict future worth” Sutton
(2011) states. As an example, Sutton (2011) explains that “a model built
to predict future gaming trip worth might be generated based on historical
information about theoretical win, actual win, credit line, time on device,
nights stayed, and average bet.”
Linear regression models: These analyze the relationship between the
response or dependent variable and a set of independent or predictor
variables. This relationship is expressed as an equation that predicts the
response variable as a linear function of the parameters. These
parameters are adjusted so that a measure of fit is optimized. Much of
the effort in model fitting is focused on minimizing the size of the
42
residual, as well as ensuring that it is randomly distributed with respect
to the model predictions. An important assumption of regression
analysis is linearity, which defines a straight line relationship between
Independent variables and dependent variables. For example, in Figure
1, we could make the assessment that an increase in ad spend also
increases sales and, using the straight line, we could predict how much
the sales would be affected.
For the casino and hospitality industry, regression models can
be used to predict a patron's future worth (Sutton, 2011). Multiple
regression models “utilize a variety of predictors and the relationships
between those predictors to predict future worth” Sutton (2011) states.
As an example, Sutton (2011) explains that “a model built to predict
future gaming trip worth might be generated based on historical
information about theoretical win, actual win, credit line, time on
device, nights stayed, and average bet.”
43
Figure 1: Linear Regression chart
Neural networks: Artificial Neural Networks (ANN) or often just
called “Neural Networks” are non-linear statistical data modeling tools
that are used when the exact nature of a relationship between input and
output is unknown. In their article Neural Networks in Data Mining,
Singh and Chauhan (2009) claim that a neural network is:
a mathematical model or computational model based
on biological neural neworks, in other words, is an
emulation of biological neural system. It consists of
an interconnected group of artificial neurons and
processes information using a connectionist
approach to computation. In most cases an ANN is
an adaptive system that changes its structure based
on external or internal information that flows
through the network during the learning phase.
They can be used to find patterns in data. A key feature of neural
networks is that they learn the relationship between inputs and output
through training.
There are three types of training in neural networks;
reinforcement learning, supervised and unsupervised training, with
supervised being the most common one. Neural Networks (see Figure
2) are data processing systems whose structure and functioning are
inspired by biological neural networks. Their fundamental
characteristics include parallel processing, distributed memory and
adaptability to their surroundings.
44
For casino and hospitality marketing purposes, neural
networks can be used to classify a consumer's spending pattern, analyze
a new product, identify a patron's characteristics as well as forecast
sales (Singh and Chauhan, 2009). The advantages of neural networks
include high accuracy, high noise tolerance and ease of use as they can
be updated with fresh data, which makes them useful for dynamic
environments (Singh and Chauhan, 2009).
Logistic regression - This method transforms information about a
binary dependent variable into an unbounded continuous variable and
estimates a regular multivariate model. Logistic regression (see Figure
3) is a generalized linear model. It is used mainly to predict binary
variables (with values such as yes/no or 0/1). Thus, logistic regression
techniques may be used to classify a new observation whose group is
unknown, in one of the groups, based on the values of the predictor
variables.
45
Figure 2: A Neural network
A/B testing: Also known as split testing or bucket testing, A/B testing
is a method of marketing testing by which a baseline control sample is
compared to a variety of single-variable test samples in order to
improve response rates.
A classic direct mail tactic, this method has recently been
adopted within the interactive space to test tactics such as banner ads,
emails and landing pages. For casino marketers, A/B testing is the most
effective way to identify the best available marketing offer (Sutton,
2011). It can test “two different offers against one another in order to
identify the offer that drives the highest response and the most
revenue/profit” (Sutton 2011).
Decision trees: Used to identify the strategy that is most likely to reach
a goal. It is a decision support tool that uses a graph or model of
decisions and their possible consequences, including chance event
outcomes, resource costs, and utility. Decision trees are sequential
partitions of a set of data that maximize the differences of a dependent
variable (response or output variable). They offer a concise way of
46
Figure 3: Logistic Regression chart
defining groups that are consistent in their attributes, but which vary in
terms of the dependent variable.
The construction of a decision tree is based on the principle of
“divide and conquer”: through a supervised learning algorithm,
successive divisions of the multivariable space are carried out in order
to maximize the distance between groups in each division (that is, carry
out partitions that discriminate). The division process finalizes when all
of the entries of a branch have the same value in the output variable,
giving rise to the complete model. The further down the input variables
are in the tree, the less important they are in the output classification
(and the less generalization they allow, due to the decrease in the
number of inputs in the descending branches). Figure 4 shows a
decision tree for responses to a marketing campaign using age and zip
code as the variables.
Figure 4: Decision Tree
47
For the casino and hospitality industry, decision trees can be
used “to identify patron characteristics that can predict the likelihood of
a patron (or segment of patrons) to abuse an offer” (Sutton, 2011).
Time series model: A time series is an ordered sequence of values of a
variable at uniformly spaced time intervals. According to the
Engineering Statistics Handbook, time series models can be used to:
• Obtain an understanding of the underlying forces and structure
that produced the observed data;
• Fit a model and proceed to forecasting, monitoring or even
feedback and feedforward control.6
A time series model (see Figure 5) can be used to predict or
forecast the future behavior of a variable. These models account for the
fact that data points taken over time may have an internal structure
(such as autocorrelation, trend or seasonal variation) that should be
accounted for. For the casino and hospitality industry, a Time Series
Analysis can be used to forecast sales, project yields and workloads as
well as analyze budgets.
Figure 5: Time Series Model
6http://www.itl.nist.gov/div898/handbook/pmc/section4/pmc41.htm .
48
Nearest Neighbor Method: Initially introduced by J. G. Skellam, the
Nearest Neighbor Method is a technique that is based on the concept of
similarity where “the expected and observed mean value of the nearest
neighbor distances is used to determine if a data set is clustered”
(Skellam, 1952). This method constructs a classification system
without making assumptions concerning the shape of the function that
relates the dependent variable with the independent variables. The aim
is to identify in a dynamic way observations in the training data that are
similar to a new observations that we want to classify. This method
does not impose a priori any assumptions about the distribution from
which the modeling sample is drawn. It involves a training set with
both positive and negative values.
Discriminant Analysis: Discriminant or discriminant function analysis
is a method used to determine which weightings of quantitative
variables or predictors best discriminate between two or more than two
groups of cases and do so better than chance (Cramer, 2003). It is a
method used in statistics, pattern recognition and machine learning to
find a linear combination of features which characterizes or separates
two or more classes of objects or events.
Because of its ability to classify individuals or experimental
units into two or more uniquely defined populations, discriminate
analysis can be used for market segmentation and the prediction of
group membership. The discriminant score can be the basis on which a
prediction about group membership is made. For example, the
discriminant weights of each predictive variable (age, sex, income, etc)
indicate the relative importance of each variable. For example, if age
has a low discriminant weight then it is less important than the other
variables. For a casino and hospitality marketing department, use of
discriminant analysis can help predict why a patron frequents one
casino over another. Discriminant analysis is specifically useful in
49
product research, perception/image research, advertising research and
direct marketing.
Figure 6 shows a generic Discriminant Analysis Model.
Figure 6: Discriminant Analysis Model
Survival or duration analysis: A branch of statistics that deals with
death in biological organisms and failure in mechanical systems (see
Figure 7). It involves the modeling of time to event data; in this
context, death or failure is considered an “event” in the survival
analysis literature – traditionally only a single event occurs, after which
the organism or mechanism is dead or broken. Survival Analysis is the
study of lifetimes and their distributions. It usually involves one or
more of the following objectives:
• To explore the behavior of the distribution of a lifetime.
• To model the distribution of a lifetime.
• To test for differences between the distributions of two or
more lifetimes.
50
• To model the impact of one or more explanatory variables
on a lifetime distribution.
By applying survival analysis to revenue management models,
casino operators can gain a truer picture of their table games revenue
(Peister, 2007).
Figure 7: Survival Analysis Model
There are several other data mining techniques that can be
used but the ones listed above are the most commonly used ones in the
industry and much of what you will need to glean from your data can
be discovered by using them. Once the data has been mined, a business
intelligence solution can tell you what's going on in your data while a
predictive analytics program can actually analyze current and historical
trends to make predictions about future events.
51
Predictive Analytics: Actionable Intelligence
Customer analytics have evolved from simply reporting patron
behavior to segmenting customers based on profitability, to predicting
that profitability, to improving those predictions (because of the
inclusion of new data), to actually manipulating customer behavior
with target-specific promotional offers and marketing campaigns.
Predictive analytics can graph a customer’s value over time as
well as anticipate that customer’s behavior. From this analysis, a casino
operator can tailor highly specific, laser-focused marketing campaigns
to each customer in the casino’s patron database. By consolidating the
various patron touchpoint systems throughout the casino property, the
casino operator can create a full view of each patron.
Drawing on data from casino player cards, predictive models
can set budgets and calendars for the casino's gamblers, calculating
their predicted lifetime value in the process. If a gambler wagers less
than usual because they may have skipped a monthly visit, the casino
can intervene with a letter or phone call offering a free meal, a show
ticket or gaming comps. Without these customer analytics, casino
operators might not notice what could be a slight, almost imperceptible
change in customer behavior that portends problems. For example, if a
long-time customer decides to cash in all their player card points
perhaps it’s because they are dissatisfied with their last experience at
the casino. Predictive analytics can quickly spot these trends and alert
casino management to the issue so that they can approach the
individual to find out if there is a problem. This kind of personalized
attention can go a long way in appeasing disgruntled customers, which
might be the difference between retaining or losing them as a customer.
Predictive analytics can glean data from a variety of disparate
sources, including:
52
• Data integrated throughout the casino's gaming systems.
• Feedback information derived from post-visit surveys.
• Web data mining from customer’s individual online
behavior.
• Social media websites.
With predictive analytics, gaming organizations can easily
segment their customers and coordinate marketing campaigns to
effectively target each segment across each outbound channel. For
example, if a casino customer is scheduled to receive all of his or her
event promotions via e-mail, the predictive analytics solution will
automatically remove him from concurrent campaigns being run
through other channels. This ensures consistency and also improves
customer satisfaction, since the organization respects the customer’s
contact preference and doesn’t inundate him or her with multiple offers.
Moreover, a predictive analytics solution monitors channel capacity
and usage to eliminate overload, while distributing campaigns equally
across the various channels. If one channel is at risk of overload, the
solution automatically shifts the remainder of a campaign to a different
channel to ensure completion. This enables organizations to maximize
the capacity and value of each channel without resorting to time-
consuming manual monitoring.
Manipulating Customer Behavior
Successful marketing is about reaching a consumer with an interesting
offer when he or she is primed to accept that offer. Knowing what
might interest a patron is half the battle of making a sale and this is
where customer intelligence and predictive analytics come in.
Customer analytics has evolved from simply reporting customer
53
behavior to segmenting customers based on their profitability, to
predicting that profitability, to improving those predictions because of
the inclusion of new data, to actually manipulating customer behavior
with target-specific promotional offers and marketing campaigns.
Predictive analytics can deconstruct a casino’s massive data
warehouse, making the information held within these databases more
meaningful. It can extrapolate trends, invent and validate a hypothesis,
as well as predict future activity. Predictive analytics can be used in a
myriad of ways, but mostly for cross-selling/up-selling, campaign
management, customer acquisition, budgeting, forecasting and
attrition/churn/retention, amongst other things (see Figure 8).
Figure 8: Applications for Predictive Analytics
Casino operators can enhance their customer relationships by
cross-selling and up-selling items that the customer might actually be
54
interested in, rather than offering them products they are likely to
reject.
Predictive analytics can also enable call center personnel to act
on inbound calls by providing offers that are likely to be attractive to
certain caller profiles. Inversely, telemarketers can listen for such
trigger phrases as “Baccarat tournament” or “seafood buffet” or “hotel
room” to help casino marketers come up with the most enticing offer
for an individual patron. In addition, automated systems like kiosks and
customer service agents on the casino floor can use predictive analytics
to provide customers with appropriate offers during other interactions.
In their article “Knowing What to Sell, When, and to Whom,”
authors V. Kumar, R. Venkatesan, and W. Reinartz (2006) showed how,
by simply understanding and tweaking behavioral patterns, they could
increase the hit rate for offers and promotions to consumers, which then
had an immediate impact on revenue.
By applying statistical models based on the work of Nobel prize-
winning economist Daniel McFadden, researchers accurately predicted
not only a specific person’s purchasing habits, but also the specific time
of the purchase to an accuracy of 80% (Venkatesan and Reinartz,
2006). The potential to market to an individual when he or she is
primed to accept the advertising is advantageous for both parties
involved; marketers don’t waste time advertising to consumers when
they aren’t primed to accept the advertisements, but do market to
consumers when and where they might want to use the advertisements.
Predictive modeling is only useful if it is deployed and it
creates an action. Taking advantage of the more powerful, statistically
based segmentation methods, customers can be segmented not only by
dollar values but also on all known information, which can include
behavioral information gleaned from resort activities, as well as the
patron’s simple demographic information. This more detailed
55
segmentation allows for more targeted and customer-focused marketing
campaigns.
Models can be evaluated and reports generated on multiple
statistical measures, such as neural networks, decision trees, genetic
algorithms, the nearest neighbor method, rule induction, and lift and
gains charts.1 Once built, scores can be generated in a variety of ways
to facilitate quick and easy implementation. The projects themselves
can be re-used and shared to facilitate faster model development and
knowledge transfer.
In his paper Predictive Analytics, Wayne Eckerson (2007)
advises creating predictive models by using the following six steps:
1. Define the business objectives and desired outcomes for the
project and then translate them into predictive analytic
objectives and tasks.
2. Explore and analyze the source data to determine the most
appropriate data and model building approach and then
scope the effort.
3. Prepare the data by selecting, extracting, and transforming
the data, which will be the basis for the models.
4. Build the models, as well as test and validate them.
5. Deploy the models by applying them to the business
decisions and processes.
6. Manage and update the models accordingly.
By utilizing data from past campaigns and measures generated
by the predictive modeling process, casino operators can track actual
campaign responses versus expected campaign responses, which can
1 Cumulative gains and lift charts are visual aids for measuring model performance.
56
often prove wildly divergent. Additionally, casino operators can
generate upper and lower “control” limits that can be used to
automatically alert campaign managers when a campaign is over or
underperforming, letting them focus on campaigns that specifically
require attention.
One of the benefits of automating campaigns is that offers
based on either stated or inferred preferences of patrons can be
developed. Analysis can identify which customers may be more
responsive to a food/beverage offer, a room offer, and/or a free chip
offer. The result: more individualized offers are sent out to the casino's
patrons and, because these offers tap into a customer’s wants, desires,
needs and expectations, they are more likely to be used; more offers
used means more successful campaigns.
By understanding what type of patron is on its property, why
they are there, and what they like to do while they are there, a casino
operator can individualize its marketing campaigns so that they are
more effective, thereby increasing the casino property's ROI.
With predictive analytics, casino operators can even predict
which low-tier and mid-tier customers are likely to become the next
high rollers. In so doing, casinos can afford to be more generous in
their offers as they know that there is a high likelihood that these
customers will appreciate the personalized attention and therefore
become long term – and, hopefully, highly profitable – patrons.
ROI: Predictive Modeling in the Real World
It is hard to get an exact Return on Investment (ROI) figure with
predictive analytic solutions because many companies who have
implemented these solutions haven’t conducted formal ROI studies.
The very nature of an ROI study can be rather nebulous as it isn't
always easily quantifiable. However, according to Wayne Erickson
57
(2007), “companies with high-value analytic programs that have
calculated ROI invest on average $1.36 million and receive a payback
within 11.2 months. It should be noted, however, that this study was
based on responses from only 37 survey respondents.”
Although they are not in the casino industry, according to a
July 7, 2009 press release about their use of the SAS Analytics suite of
tools, the National Geographic Society saw “a return on investment of
200 to 300 percent, with the best-performing customer segments
realizing 50 percent overall campaign performance improvements”
(SAS, 2009).
Some other real world examples of how predictive analytics
have helped companies increase customer service as well as drive
profits straight to the bottom line include:
• To test a marketing campaign hypothesis, Harrah's chose
two similar groups of frequent slot players from Jackson,
Miss. Members of the control group were offered a typical
casino-marketing package worth $125 that included a free
room, two steak meals and $30 of free chips at the Tunica
casino. Members of the test group were offered $60 in
chips. The more modest offer generated far more gambling,
suggesting that Harrah's had been wasting money giving
customers free rooms. Thereafter, profits from the revamped
promotion nearly doubled to $60 per person per trip
(Binkley, 2000).
• When Pearl River Resort initiated a marketing campaign
using SAS PVO, they were surprised by the results, which
showed that not all high-value guests were the same, some
actually weren’t very big gamblers but their spend in other
parts of the resort more than made up for their lack of
58
gambling. This information was put to good use during
times when the casino was hosting poker or blackjack
tournaments, a time when the casino property knew that the
tables would be crowded. Pearl River Resort was able to fill
the resort with people the company knew had little intention
of venturing onto the gaming floor7.
• By using predictive analytics Harrah’s, was able to identify
a small group of customers who accounted for about 30% of
their overall gamblers (Binkley, 2000). These customers
spent between $100 and $499 per trip but actually
accounted for about 80% of the casino’s revenue and nearly
100% of the casino’s profits (Binkley, 2000).
Predicting a Patron's Future Worth
Modern casino analytics and patron management systems have
provided the gaming industry with an enormous amount of highly
detailed data about when, where, how often and how much patrons are
playing at a casino (Sutton, 2011). This is information that can be used
to better segment its customers as well as predict future behavior, and
improve marketing outcomes (Sutton, 2011). “Patron analytics are
essential for maximizing revenue driven by mass market marketing
campaigns,” Sutton (2011) argues in his paper Patron Analytics in the
Casino and Gaming Industry: How the House Always Wins. Sutton
(2011) claims that, when it comes to casino patron analytics, casino
operators must seek answers to the following questions:
• How much is a patron worth, how much can we expect a
patron to lose in the future, and who are the most valuable
patrons?
7http://www.sas.com/news/feature/02jun05/pr.html (Accessed April 2, 2012)
59
• What patrons come together?
• What patrons are most likely to abuse an offer?
• What patrons are the most and least likely to respond to an
offer?
• Which offers perform the best?
The single most important thing patron analytics must
determine is the patron's worth to the casino property (Sutton, 2011).
Predicting a patron's future behavior is not an easy thing to do as it can
be affected by a number of variables, including total income,
expendable income, ethnicity, reasons for a trip (convention vs.
vacation), among many other things (Sutton, 2011). However,
“although that information is often available to append through third
parties, there is still a plentiful amount of information found with in-
house data and this data can be used to build models and metrics to
predict a patron's future worth” (Sutton, 2011). Once worth has been
determined, patrons can be “segmented into groups based on other
behaviors and effective marketing campaigns can be developed around
those behaviors” (Sutton, 2011).
Determining a patron's worth is imperative because it helps
reveal how valuable that patron is to the casino, and this is information
that can dictate how much should be reinvested in the patron in the
future (Sutton, 2011). Sutton (2011) argues that, “There are two main
components of worth – the financial sources of worth (i.e., gaming or
hotel) and the unit of time to which it refers (daily, weekly, monthly,
etc.). Additionally, worth can refer to historical worth, which is already
known, or future worth, which is unknown” (Sutton, 2011).
Most revenue sources are fairly straightforward – room
revenue refers to how much the patron paid for the room, restaurant
revenue, obviously, is how much he or she paid for food and/or drinks,
60
(Sutton, 2011). Gaming revenue, however, isn't as simple because
probability is involved and there are “two important measures used to
assess a patron's gaming worth – actual and theoretical loss” (Sutton,
2011). Actual loss refers to how much money the patron actually lost
(or won), “whereas theoretical loss usually refers to the amount of
money a patron is expected to lose based on the amount of money
wagered, the time spent playing, and the probability associated with the
type of games played” (Sutton, 2011). Whereas actual loss is generally
used to measure campaign performance and profitability, theoretical
loss relies more heavily upon predictive analytics and is a much
stronger predictor of future behavior (Sutton, 2011). For Sutton (2011),
the formulas used to calculate the theoretical loss for table games and
slots are as follows:
• Table Theoretical Loss = Average Bet x Time Played x
Speed of Game x House Advantage.
• Slot Theoretical Loss = Coin in x Hold Percentage.
Once patron worth has been defined, data mining and
modeling techniques can be used to estimate predicted worth in the
future (Sutton, 2011). “Simple metrics based on historical behavior,
such as Average Daily Theoretical Loss or Average Trip Theoretical
Loss, will produce fairly accurate predictions of future worth,” Sutton
(2011) notes. “However, advanced predictive models are able to predict
worth with more accuracy and power by accounting for both patterns in
behavior over time and relationships between predictive inputs that
exist within casino data,” Sutton (2011) argues.
There are many different techniques that can be used to
develop models to predict future worth, the most common of which are
regression models (Sutton, 2011). Multiple regression models “utilize a
61
variety of predictors and the relationships between those predictors to
predict future worth” Sutton (2011) states. As an example, Sutton
(2011) explains that, “a model built to predict future gaming trip worth
might be generated based on historical information about theoretical
win, actual win, credit line, time on device, nights stayed, and average
bet.”
Regression models could also be “built using such categorical
variables predictors as gender, ethnicity, age range, or demographic
variables” (Sutton, 2011). “Regression models are particularly
effective,” Sutton (2011) concludes “because the model can be used to
score historical data to predict an unknown outcome, which is worth in
this case, with a certain degree of confidence.”
Identifying the Casino's Most Valuable Patron
One way to determine who the casino's best patrons are is to try to
separate the skilled gamblers from the unskilled gamblers (Sutton,
2011). Most casino databases don't have a very good measure of skill,
however, it is possible to look at whether a patron is usually a loser or a
winner (Sutton, 2011). “A quick and easy way to evaluate a player's
skill is by calculating the percentage of trips where the player actually
lost money,” Sutton (2011) explains. For example, as Sutton (2011)
states:
Did a player with five trips lose money on all five of
those trips? Although this might just be an indicator
that the patron will play until he is out of money or
time, it is also a fairly simple way to identify the
patrons that do not come away as winners very often.
This is an instance where actual loss might be a good
62
predictor of worth, as we would rather have these
patrons in the casino.
Although slot machines are not really skill based and loses
could be more attributed to luck, a differentiation can be made between
patrons by looking at behaviors and strategies of slot players (Sutton,
2011). “Since casinos have to pay a certain percentage of win or handle
to the slot manufacturer for participation games, patrons that primarily
play non-participation games are slightly more valuable to the casino”
Sutton (2011). Casino operators should keep in mind how much play a
slot patron has on participation machines compared to machines it
owns outright.
Another thing to look at for slot players is their average bet
relative to the maximum bet on the games they play (Sutton, 2011). In
most cases, a patron has to play the maximum bet in order to be eligible
for jackpots and progressives (Sutton, 2011). Given two patrons of
comparable theoretical worth, the one who plays closer to the
maximum bet allowed is more likely to hit a jackpot than the other
(Sutton, 2011). Sutton (2011) argues, rather counterintuitively, that “the
patron with the higher average bet would seem to be more valuable, but
since the lower bet patron is less likely to hit a jackpot, the lower bet
patron might be a lower risk.” Although this metric has usefulness on
its own, casino management could also “use it as either a predictor in a
model for future worth or a decision tree predicting whether a patron
will respond” to an offer (Sutton, 2011).
Identifying Patrons Who Come Together
Beyond a patron's worth is the combined worth of a household, which
“refers to the combined worth of multiple patrons that tend to make
63
their trips together” (Sutton, 2011). This can be difficult to identify
because patrons of the same household might stay in separate rooms,
take trips separately, or one patron might only come when accompanied
by another patron (Sutton, 2011). Although tricky, identifying
household worth can pay huge dividends by helping to “account for
revenue that looks like two separate individuals but can be combined
into one 'household'” (Sutton, 2011).
Although many patron management systems allow the linking
of accounts so that patrons who come together can be easily identified,
data mining can help to identify groups of patrons who come together
without linked accounts (Sutton, 2011). To do this, the casino must first
identify patrons who make their trips at the same time as each another.
Then a combination of various data points are studied to identify the
“households” (Sutton, 2011). For Sutton (2011), these include:
• Last name: this can identify relatives who come together.
• Address: this reveals roommates or patrons living together
with separate last names.
• Room or floor: patrons in a group tend to request rooms
that are near each other.
• City and State: this could reveal friends and/or relatives
who are from the same area.
• Time of day that games are played: this will reveal
players who are together on the casino floor, which is
obviously the behavior of friends or family members.
• Game type: this reveals patrons who are playing in the
same location on the casino floor, or at least close to one
another.
• Restaurant and retail charges: this reveals patrons who
have charges from the same outlet on the same day.
64
By grouping patrons into a “household worth”, a group of four
patrons who might be of modest worth individually, can “come together
and stay in the same room every time and thus are worth more as a
group” (Sutton, 2011). Armed with this grouping information, a casino
can adjust its marketing effort and send an offer that is based on the
patron's combined worth rather than on their individual worth (Sutton,
2011). It is a distinction that could be the difference between the offer's
acceptance or its rejection.
Minimizing Patron Abuse
Predictive models of worth should take into account the likelihood of a
guest playing on a future trip (Sutton, 2011). It is also advisable to
build a separate model that identifies patrons who are likely to use a
future offer but not play in the casino, thereby taking advantage of the
property (Sutton, 2011). “Since many offers in the casino industry tend
to be for complimentary rooms that are given to patrons upfront,
patrons that redeem offers and do not play have a considerable impact
on campaign success and profitability,” Sutton (2011) points out. For
marketing campaigns to be successful over the long term, it is
important to not only identify the patrons who are abusing the system,
but also to adjust the offers they receive (Sutton, 2011). According to
Sutton (2011) “Decision trees and logistic regression are common
statistical methods used to identify patron characteristics that predict
the likelihood of a patron (or segment of patrons) to abuse an offer.”
Factors such as age, gender and history of abuse are likely
predictors of abuse (Sutton, 2011). Survey data from post-visit follow-
up surveys can be used to identify these predictors (Sutton, 2011). “If a
patron had a bad experience in the past, they might take an offer for a
free room as revenge for that bad experience,” Sutton (2011) warns.
65
But by identifying those patrons at risk of abusing offers, the casino
property can decide how best to market to these individuals (Sutton,
2011). Instead of receiving the general offer of a free room, the patron
would be sent an offer that requires them to play up to a certain level or
they would be required to pay for their room” (Sutton, 2011).
Campaign Optimization
In addition to predicting the future worth of patrons, casino marketers
must know which marketing campaigns are the most effective at
driving up response rates, as well as which campaigns will increase
gambling revenues and property profits (Sutton, 2011). Understanding
a patron's probable future worth is critical in determining the eligible
reinvestment levels that make financial sense for the casino property
(Sutton, 2011).
“A patrons' behavior and interests can be used to identify the
offer(s) that will be most appealing to each patron and generate the
most profitable response,” argues Sutton (2011). Although offers of free
rooms and free gaming play are historically the strongest drivers of
response, the cost associated with them can be detrimental to the
property (Sutton, 2011). Sometimes it doesn't make good business
sense because “not every patron who is eligible for a free room has to
be offered a free room to respond – some might be willing to pay for a
discounted room or even a full price room” (Sutton, 2011). “By
analyzing the likelihood that a patron will respond to a certain offer or
offers, casino executives can optimize the offer that each patron is
given in order to maximize the amount of revenue and profit driven by
marketing campaigns as a whole,” concludes Sutton (2011).
A/B testing is the most effective way to identify the best
available offer (Sutton, 2011). A/B testing “involves testing two
66
different offers against one another in order to identify the offer that
drives the highest response and the most revenue/profit” (Sutton, 2011).
Logistic regression, decision trees, and discriminant analysis can also
be used to cull through a casino's historical data and uncover factors
that are related to whether a patron responds to an offer or not (Sutton,
2011). “These factors can then be used to assess the likelihood of
response based on the similarity of a patron profile to that of
responders,” explains Sutton (2011). Obviously, “to build accurate and
predictive response models, historical data about response is required”
(Sutton, 2011). “The likelihood of response might be a broad measure
of response that refers to the likelihood that a patron will respond to
any offer, or it might be specific to the likelihood of response to a
specific type of offer,” warns Sutton (2011). Effective response models
have a dual purpose; they can help identify which patrons are most
likely to respond to an offer, as well as reveal which offer patrons are
most likely to respond to (Sutton, 2011). According to Sutton (2011),
“There are at least three main uses of response modeling that can
improve marketing results:
• Identify the likelihood of patrons to respond to the offer.
• Identify the offer(s) to which patrons are most likely to
respond.
• Predict when a patron is likely to return.”
Determining which offers a patron is most likely to respond to
is only half the battle. It is also important to know exactly when that
patron is planning to make his or her trip as well (Sutton, 2011).
Although it's almost impossible to know exactly when a patron is likely
to return (without them making a reservation, of course), there are a
variety of predictive analytics methods that can help take out the guess
67
work (Sutton, 2011). Frequency analysis, regression, and survival
analysis can all be used to assess when a patron is likely return to a
casino (Sutton, 2011).
Knowing when a patron is likely to return can help identify
patrons who haven't been at the property in while and might be at risk
of leaving for good (Sutton, 2011). To identify these patrons, the casino
should discover the average or median time between a patron's trips
(Sutton, 2011). This could be segmented by geography, worth, or even
historical frequency, such as trips made weekly, monthly, quarterly,
annually, bi-annually, and so forth (Sutton, 2011). Patrons who “have
not made a trip within the decided amount of time for their segment are
subsequently flagged and dealt with appropriately” (Sutton, 2011).
Sutton (2011) believes a casino's marketing department should have
two primary goals; generate trips sooner than expected and convert
patrons into more frequent visitors.
A lesser goal would be identifying patrons who are at risk of
leaving the property for good (Sutton, 2011). In cases such as these,
sending the patron an offer using “last chance” or “we miss you”
language could help retain them (Sutton, 2011). The offer contained
within should probably be better than the patron has received in the
past to really catch their attention (Sutton, 2011). By knowing when a
patron is likely to return, a casino “can adjust marketing strategies
appropriately in order to save money on mail costs, retain guests and
increase loyalty” (Sutton, 2011). These are all important strategies that
should help drive profits straight to the casino's bottom line.
Conclusion
I started this chapter with a quote that asked the question, “Do you
want to ride the analytics horse to profitability...or follow it with a
68
shovel?” and I believe the return on investment of implementing a
predictive analytics solution in the casino business would be
substantial. Casino patrons the world over like the same thing – free
rooms, free gaming comps or free meals – all delivered within a
marketing campaign that taps into their own personal wants, desires,
needs and expectations. This can only be done with predictive
analytics.
Modern casino analytics and patron management systems
contain enormous amounts of highly detailed data about when, where,
how often and how much a casino patron is spending at a casino property
(Sutton, 2011). This information can be used to better segment customers,
to predict future behavior, and to improve marketing outcomes (Sutton,
2011). Sutton (2011) claims that, when it comes to casino patron analytics,
casino operators must seek answers to the following questions:
• How much is a patron worth, how much can we expect a
patron to lose in the future, and who are the casino's most
valuable patrons?
• What patrons come together?
• Which patrons are most likely to abuse an offer?
• What patrons are the most and least likely to respond to an
offer?
• Which offers perform the best?
Answers to all of these questions can be found by utilizing
predictive analytics, a variety of statistical techniques that analyzes
current and historical facts to make predictions about future events.
Using such techniques as predictive modeling, machine learning, data
mining and game theory8, predictive analytics can build models that
8The study of mathematical models of conflict and cooperation between intelligent rational decision-makers (Myerson, 1991).
69
exploit patterns found in historical and transactional data to identify
business risks and opportunities. Predictive analytics can be broken
down into three different types of models, predictive, descriptive and
decision. By combing through copious amounts of data to discover
patterns, trends and relationships that are too subtle or complex for
humans to detect, data mining, along with predictive analytics, can
identify potentially useful and understandable correlations and patterns
in a property's datasets.
For the casino industry, data mining can cull through copious
amount of data, data that is coming from such disparate sources and
departments as sales, credit, and marketing and then measure patron
behavior against more than a hundred different attributes. This is a far cry
from the three or four different attributes that statistical modeling used to
offer. The most common data mining techniques are linear regressions,
neural networks, logistics regressions, decision trees and A/B testing.
Predictive analytics can graph a customer’s value over time as
well as anticipate that customer’s behavior. From this analysis, a casino
operator can tailor highly specific, laser-focused marketing campaigns to
each customer in the casino’s patron database. Drawing on data from
casino player cards, predictive models can set budgets and calendars for
the casino's gamblers, calculating their predicted lifetime value in the
process. Armed with this information, casino properties can use predictive
analytics to:
• Create direct mailing campaigns.
• Create seasonal promotions.
• Plan the timing and placement of advertising campaigns.
• Create personalized advertisements.
• Define which market segments are growing most rapidly.
70
• Determine the number of rooms to reserve for wholesale
customers and business travelers.
Patron worth is the single most important thing patron analytics
must determine (Sutton, 2011). Once it has been determined, patrons can
be “segmented into groups based on other behaviors and effective
marketing campaigns can be developed around those behaviors” (Sutton,
2011). Determining a patron's worth is imperative because it helps reveal
how valuable that patron is to the casino, and this is information that can
dictate how much should be reinvested in the patron in the future (Sutton,
2011). Once patron worth has been defined, data mining and modeling
techniques can be used to estimate predicted worth into the future (Sutton,
2011).
Regression models are particularly effective to find patron worth
because the model can be used to score historical data to predict an
unknown outcome (Sutton, 2011). Multiple regression models “utilize a
variety of predictors and the relationships between those predictors to
predict future worth” (Sutton, 2011).
In addition to predicting the future worth of a patron, casino
marketers must know which marketing campaigns they are running are
the most effective at driving up response rates (Sutton, 2011).
Understanding a patron's probable future worth helps determine the
eligible reinvestment levels that make financial sense for the casino
property (Sutton, 2011).
A/B testing is one of the most effective way to identify the best
available offer to be made to a patron, but logistic regression, decision
trees, and discriminant analysis can also be used successfully (Sutton,
2011). Effective response models have a dual purpose; they can help
identify which patrons are most likely to respond to an offer, as well as
reveal which offer patrons are most likely to respond (Sutton, 2011).
71
Casino operators should keep in mind that data mining will only
be successful if their casino patrons are willing to provide information on
themselves. Privacy is a big issue and will always remain so in the mobile
age. Casino properties that can honor a patron's privacy demands will find
patron loyalty comes with voluminous amounts of priceless patron data.
This is data that can be used to create marketing campaigns that should
prove highly effective. By understanding what type of patron is on its
property, why they are there, and what they like to do while they are there,
casino operators can individualize their marketing campaigns so that these
campaigns are more effective, thereby increasing the casino property's
ROI.
72