Analytics Tool Predicts Customer BehaviorPosted 2012-08-03 Email Print
WEBINAR: On-demand webcast
Next-Generation Applications Require the Power and Performance of Next-Generation Workstations REGISTER >
Seminole Gaming needs to be able to “see into the future” when it comes to their customers’ behavior, so the company implemented analytics software. Could such ‘fortune-telling’ help your business?
By Ralph Thomas
Seminole Gaming operates seven casinos in Florida on behalf of the Seminole Tribe, and two of these facilities are Hard Rock-branded hotel casinos. The company has more than 11,000 slot machines, 300 table games, dozens of restaurants and nearly 10,000 employees.
In the past, we developed a proprietary analytics tools, including an enterprise data warehouse and tools forbusiness intelligence, campaign management and sales force management. We leveraged these tools in many ways but saw a significant benefit by executing test-and-control methodology on thousands of customer segments.
However, even with these successes, we were missing something. Our segmentation strategy relied solely on prior behavior, largely using traditional RFS (recency, frequency, spend)-driven segmentations. We needed to be able to “see into the future” when it came to our customers’ behavior, so we decided to purchase SAS analytics software in 2010.
On the first day of training at Seminole Gaming, we built a practice model using SAS’ Enterprise Miner, taking real customer data and running it through a simple decision tree. The goal was to gain experience using the software, but our practice problem was to understand what behavior was predictive of a customer returning to one of our properties during the course of a month. That this was a practice model did not deter us because the model was built on real data, and the outcome of the model looked reasonable.
At the time the model was built, the analytics department was part of the marketing department, so taking action on the results was not difficult. A new program was put in place to take action based on the two metrics uncovered. These programs were tested via traditional test-and-control methodology to ensure that the marketing treatments drove incremental gains above and beyond the cost of the program.
While one of the metrics proved fruitless (the incremental revenue was canceled out by the incremental cost), the second metric was shown to drive significant incremental profit each month: well over $1 million a year. So, essentially, the cost of the SAS program was paid for during the very first training session on the software
Typically, aside from an accidental discovery like the one described above, the low-hanging fruit for any company that engages in a large volume of direct mail is finding ways to eliminate the cost of sending mail to customers who are unlikely to respond to an offer.
But what of email and social media? While the newer electronic channels are very important marketing tools, in some businesses (in particular the casino business) direct mail is a much more effective channel.
Seminole Gaming maintains a large direct-mail budget and was able to easily build a response model for its direct mail relating to concerts at Hard Rock Live, located on the grounds of Seminole Hard Rock Hotel & Casino in Hollywood, Fla. The response model was built using Rapid Predictive Modeler, a new tool that is part of Enterprise Miner 6.2.
We identified the 35 percent of our customers most likely to respond to one of these offers. For the 65 percent least likely to respond, we consolidated mailers—advertising multiple concerts, instead of a single concert, in each mailer. While the single-mailer-per-concert approach proved most effective for customers likely to respond, advertising multiple concerts in a single mailer saved mail costs and increased response. This response model was a second success, generating another incremental gain of more than $1 million in profit annually.
Additional response models have been built, and are continuing to be built, to ensure that the direct-mail program is as efficient as possible.
After we’ve identified people to whom we should not be mailing, our next strategy was to identify those to whom we should be mailing, but currently aren’t. Like response models, this is an important strategy for any company that relies heavily on direct mail.
In the gaming industry, we have a wealth of data about our customers’ purchases. A significant portion of our player base uses the casino rewards card: In local markets, 70 or 80 percent of all business is often tracked by such rewards cards. For each of these players, we know their play choices and outcomes for every machine or table on every day.
Traditional casino direct-mail programs rely heavily on metrics such as average daily actual (a measure of how much money a player loses on a given day), average daily theoretical (a measure of how much a player would have lost if he or she had no more or no less luck than expected, per day), average daily worth (a calculation that combines the two metrics above) and points earned (a measure of how much total play a player has given the casino). Since it’s unprofitable to mail every customer for every marketing campaign, casinos typically draw limits on how deeply they will mail.
Let’s look at an example. Imagine there’s a campaign that offers $5 in free play to any customer whose average daily worth is greater than $50. In effect, the marketer has determined that offering $5 to any customer whose value is more than $50 drives enough incremental trips to cover the expense of the offer to the customers who were going to come anyway.
We will assume testing has been done to prove that $5 is the optimal offer for a $50 customer. The question becomes, can we find “hidden gems” in the customers who are below $50 in value? The methodology to do this is fairly straightforward. Go back in time and look at all customers whose value prior to time period X is less than $50 but who came back after that time period and were worth more than $50. This is a basic change-of-behavior model that the Rapid Predictive Modeler can handle easily.
We’ve begun testing such models, and our very first effort proved we can find hidden gems and give those customers the right offer to drive enough incremental trips to ensure that the campaign is profitable. Each month we build more of these models to accommodate additional campaigns at multiple properties. This simple concept is proving to drive strong results.
The models above focus on using SAS software to find models that will return profitable results quickly. However, Seminole Gaming hasn’t stopped there. We also spend time doing R&D, investing energy in projects that may not see a direct return immediately.
One such project was to build association models to perform market basket analysis on the slot machines that our customers play. As a simple example, suppose our casino had only three slot machine games: A, B and C, and two customers: Bob and Rita. Suppose Bob and Rita each spend $100 per visit, but while Rita plays only game C, Bob splits his money between games A and B. In this case, we would say that games A and B are associated with each other, whereas game C isn’t associated with any other games.
If you add millions of customers and thousands of slot machines to the mix, the model gets much more difficult, but fortunately our new software includes easy-to-use association models to cut through the complex data.
When Seminole Gaming embarked on this project, we didn’t know what we’d find. It turns out we found very surprising associations on the slot floor. These associations cannot be revealed here, but we can say that the methodology behind slot decisions has been fundamentally altered thanks to the findings from this model.
In the future, Seminole Gaming hopes to continue along the two paths described above: creating models that will have an immediate, positive impact on the business and using the SAS software to investigate the data in an R&D environment.
The next step after developing association models is to cluster the data. This is not easy. If you multiply millions of customers by thousands of slot machines, you get a matrix with trillions of cells.
The goal is to tackle this matrix and uncover hidden groupings of customers and slot machines. Once uncovered, the sky is the limit. Imagine a whole new way to segment your customers based on how they select their products—or a whole new way to analyze, or even locate, your slot machines based on these choices. It’s very exciting work that we’re looking forward to tackling.
Deploying the models described above has had a significant positive impact on the Seminole Gaming business. First, there are intangible benefits: The knowledge received from association models is constantly influencing our slot product decisions.
In addition, since the casino operators are seeing benefits from the models already deployed, they are now thinking in terms of predicting behavior, rather than simply using prior behavior. They are brainstorming new and better ways to get future value out of our SAS software.
In terms of tangible benefits, the models built have been tested in direct-marketing campaigns, and have been statistically proven to provide an annual benefit exceeding $5 million. Overall, we are on pace to exceed $30 million in realized benefit over the first five years using this software.
Ralph Thomas is vice president of strategic analytics and database marketing at Seminole Gaming, of Hollywood, Fla.