Big-data Analytics for Casinos


In today’s highly competitive gaming market, where casinos have increased their range of offerings to include retail, golf, and theater - patrons are looking for more than just table games and slot machines. With such a wide range of available entertainment options, the most valuable patrons are no longer just those that spend the most on the casino floor, but those who spread their spending across all the property offerings. In order to have holistic view of patrons behavior requires a complete picture of all patrons’ spending behaviors and a careful analysis of their preferences. Therefore, it is no longer enough for a casino to know who the high rollers are. They must go deeper into the patron base and identify those whose activities go beyond the casino floor, and the only way that they can do that successfully is through analytics. Some of the typical analytical questions in the gaming industry are:

a.) How much is a patron worth?

b.) How much can we expect a patron to lose in the future?

c.) Who are the most valuable patrons?

d.) What patrons come together?

e.) What patrons are most likely to abuse an offer?

f.) What patrons are most and least likely to respond to an offer?

g.) What offers perform the best?


This question is of paramount importance in the casino industry. To predicting a patron’s future behavior, is not easy and requires a number of variables, many of which are not readily available to casinos, and needs to be acquired from third-party data providers. These variables include income, ethnicity, and reasons for a trip (Convention vs. vacation). Once the worth of a patron is determined, patrons can then be segmented into groups based on other behaviors and effective marketing campaigns can be developed around those behaviors.

On a question of patron's worth there are two components – the financial sources of worth (i.e., gaming or hotel) and the unit of time to which it refers (daily, weekly, monthly, etc.). In addition we can refer to historical worth, which is already known, or future worth, which is unknown. Worth can also be broken down into various sources (i.e., slots, tables, poker, race and sports) depending on the business issue being addressed. Definition of worth depends on how knowing this worth will be exploited. For instance, daily gaming worth would be most useful for building a campaign with a daily free play offer. For example, two separate patrons, each with an annual worth of $10,000, will have a completely different day worth if one patron comes one day a year and the other comes one day a week. Alternatively, annual worth might be more useful than daily worth to identify the patrons to target for an exclusive event. In other words, aggregate measures of worth are ideal for identifying the most valuable patrons to the business overall, whereas individual measures are usually suited more toward identifying valuable patrons for specific offers and campaigns.

Most revenue sources are fairly straightforward – room revenue is how much the patron paid for a room, restaurant revenue is how much they paid for food and drinks, etc. Gaming revenue, however, is a little more complicated because the probability is involved. There are two important measures used to assess a patron’s gaming worth - actual and theoretical loss. Actual loss is 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 type of games played. Theoretical loss tends to be more heavily relied upon for predictive analysis and is a much stronger predictor of future behavior, as actual loss is usually used to measure campaign performance and profitability.

Below are the general formulas used to calculate theoretical loss for tables and slots:

a.) Table Theoretical Loss = Average Bet x Time Played x Speed of Game x House Advantage

b.) Slot Theoretical Loss = Coin in x Hold Percentage


In order for business to be able to use data mining to estimate predicted worth in the future, they need to know simple metrics based on historical behavior, such as Average Daily Theoretical Loss or Average Trip Theoretical Loss which will produce fairly accurate predictions of future worth. For example, a model built to predict future gaming trip worth might be generated based on historical information about theoretical win, actual win, credit line, and time on the device, nights stayed, and average bet. Models can also be built by using such categorical variables as predictors, as gender, ethnicity, age range, or other demographic variables.


In addition to patron's future worth modeling, there are other analytical methods to determine a patron’s value to the business. One way to identify the best patrons is to try to separate the skilled gamblers from the unskilled. That can be achieved by calculating the percentage of trips where the player actually lost money. For instance, did a player with four trips lose money on all four of those trips? Although this might just be an indicator that the patron will play until he is out of money or time, it also is a fairly simple approach to identify the patrons that do not come away as winners very often. This is an instance where the actual loss might be a good predictor of worth, as we would rather have these patrons in the casino. This is highly useful for games requiring skill, such as table games and video poker, but might not be very convenient for slot players where there is little skill involved and constant losses are more likely to be attributed to lack of luck.

Although slot machines are not really skill-based, we can still differentiate between patrons by looking at the strategies and behaviors of slot players. One quick and easy way to separate slot patrons is to compare how much play they have on participation machines relative to own machines. 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 advantageous to the casino.

A slightly more complex metric for slot players is to look at their average bet relative to the maximum bet on the games they play. Usually, the maximum bet needs to be played in order to be eligible for jackpots and progressives. Given two patrons of similar theoretical worth, the one that plays closer to the maximum allowed bet is more likely to hit a jackpot than the one who doesn’t. Usually 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. This metric could be informative on its own, or could be used as either a predictor in a model for future worth or a decision tree predicting whether a patron will respond. These are just a few examples of how data mining can provide beneficial information to differentiate between players that might otherwise seem very similar.


Another important consideration in the discussion of patron worth is household worth. This relates to the combined worth of multiple patrons that tend to make their trips together. This can be hard to identify, as these patrons might stay in one room or separate rooms, or one patron might only come when accompanied by another patron. Additionally, the other patron might make trips without the first patron. Although identifying household worth can be tricky, it can pay huge dividends by helping to account for revenue that looks like two separate individuals but can be combined into one “household”. Many patron management systems contain the functionality to link accounts so that patrons that come together (i.e., married couples) can be easily identified. Unfortunately for the casino analyst, patrons might not be allowed to have linked accounts because of system limitations or because of business policies based on tax and gaming regulations.

Nevertheless, data mining can be used to identify groups of patrons that come together without linked accounts. First, we need to identify patrons that make their trips at the same time as one another. Second, we can use a combination of various Travel/Restaurant/retail charges (whether they have charges from the same outlet on the same day).

The above method is a great way to identify patrons with trips over the same period of time that have some commonality about their behavior that we can use to be fairly certain the patrons are together. For some of the more subjective measures (i.e., room, floor, city, time/type of play) it’s a good idea to be more conservative about how many overlapping trips the patrons have. For instance, two patrons with one overlapping trip and rooms next to each other may or may not be in the same “household” group. However, those same patrons with 5 overlapping trips, each with rooms next to one another, are much more likely to be in the same household grouping. In this manner, household grouping can identify a group of four patrons that are of “middle of the pack” worth individually, but come together and stay in the same room every time and thus are worth more as a group. Now, we can adjust our marketing efforts and send a letter based on their combined worth and the knowledge that we’re not really marketing to four unique individuals, rather to a group of related patrons.


Some factors that are likely predictors of abuse are age (younger patrons are more likely to abuse), gender, and history of abuse. Additionally, survey data (e.g., from follow-up surveys after a patron’s visit) that are linked to individual patrons can be used to identify other predictors. If a patron thought he was treated unfairly and had a bad experience in the past, they might take an offer for a free room as revenge for that bad experience. When they come in they take all the perks available from the offer then walks across the street to gamble and spend money so they can later get a better offer from across the street.

By identifying the patrons at risk of abusing offers, the business can determine how to market to those risky patrons. For instance, someone might reach out to the patron to try and rectify the situation if they had a bad experience. Instead of sending them to the general offer for a free room, they would be sent with an offer that requires them to play to a certain level or they will have to pay for their room. However, it might be better to not even give them the option to get a complimentary room. In this case, it is helpful to know what offer is the best type of offer to send a patron and whether it’s even worth the money to send them offer at all.


In addition to predicting the future worth of patrons, it is important to understand which marketing campaigns are the most effective for driving response, revenue, and profit. In general, certain offers are preferable to others, and specifically certain offers will be better for certain patrons. Common components of marketing involve offers for rooms, restaurants, retail, and gaming (i.e., free play). While knowing the probable future worth of a patron is critical for determining the reinvestment level for which a patron is eligible, patrons’ behaviors and interests can be used to identify the offer(s) that will be most appealing to each patron and generate the most profitable response.

Offers that include free rooms and gaming free play are historically the strongest drivers of response. However, a complimentary room is not the ideal offer for every patron. On one hand, some patrons won’t be eligible for a free room because their predicted gaming worth is too low to warrant a free room. On the other hand, not every patron that’s eligible for a free room has to be offered a free room to respond. Some may even be willing to pay for a discounted room or even a full price room. By analyzing the likelihood that a patron will respond to a certain offer or offers, casino analysts 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. Some of the more common statistical approaches are logistic regression, decision trees, and neural networks. Essentially, these statistical methods use historical data to find the factors that are related to patron's response. These methods have historically been used in direct marketing analysis to identify the best types of offers and the most likely responders. In order to build accurate and predictive response models, historical data about response is required. The likelihood of response might be a broad measure of response that refers to the likelihood of a patron will respond to any offer, or it might be specific to the likelihood of response to a specific type of offer.

There are at least three main uses of response modeling that can improve marketing results:

  1. Identify the likelihood of patrons to respond to the offer
  2. Identify the offer(s) to which patrons are most likely to respond
  3. Predict when a patron is likely to return

A response model can lead to lower mailing costs by identifying patrons that are very unlikely to respond to a particular offer. In the previous example, the business can identify the likelihood of response from all eligible patrons. After that, they can identify the most valuable patrons that are more likely to respond. This allows the business to estimate the expected response from the most valuable patrons and eliminate mailings to the patrons that are of lower worth and/or are unlikely to respond. If the initial list of patrons does not account for all the tickets bought, then the lower worth/high likelihood to respond patrons can be contacted. Rather than sending out the blanket offer to 100,000 patrons as in the year before, only 50,000 patrons are mailed and the offer is redeemed by higher worthy patrons. Occasionally, response likelihood models will lead to easy decisions, such as cutting out low worthy patrons with a low likelihood to respond. However, more complex situations might arise since response models are never perfect. No matter how good a model and historical data are, there is always a chance that a patron identified as unlikely to respond will respond. This can lead to a segment of patrons which is identified as being the least likely to respond to the offer in the example above. However, this is just a little segment of patrons with a high predicted worth, and if just one of the patrons in this segment responds the entire cost of the mailing will be covered.


Casinos marketing department may have wide array of offers for which they need likelihood scores in order to identify the best offer that the patron is most likely to respond to. One specific offer can be to go to some casino event or show, offer for a room, spa offer, etc. Analytics can help to select and optimize relevant offers and predict expected return on these offers, bearing in mind that monetary return and likelihood of response are often negatively correlated.

 If you send one targeted offer to a patron chances of him responding is greater but profit is lesser than if you send bulk offer, and vice versa. And so, response modeling can help identify which offer has the best chance of driving a response. If this is done for a segment as a whole, we can maximize response and profit by ensuring that the most expensive offer (in terms of reinvestment) goes to the patrons that might be less likely to respond. It might be tempting to offer the guest the option to choose which offer they want (especially if the offers are of equal value), but there is evidence to suggest that narrowing down a guest’s choices is actually better and more profitable. In this case, it’s a great idea to test whether a targeted offer based on a likelihood model performs better than an offer that gives the patron a choice between the available offers. Additionally, survey or preference data might be available that informs the business of what type of offers patrons are interested in receiving. Patron preferences can also be used in response models to determine whether a stated preference is actually a good indicator that a patron will respond to an offer targeted towards those interests.


Apart from knowing if the patron is the most likely to respond it would be highly beneficial to know when patron will return, and one way to have some idea of that is through building a predictive model. Key to such model is historical data of his previous visits to see if there are any time, or seasonal patterns. Even if such data doesn't exist for specific customer – knowing his peer-group return pattern could be beneficial, but such proxy information is likely to result in a weaker model and prediction. Historical data can help to identify segments of patrons that are expected to make trips weekly, monthly, quarterly, annually, bi-annually, and so forth. Marketing can integrate information from predicted worth, optimal offers, and time to subsequent trip to maximize campaign success in a number of ways. The business can save money by adjusting the frequency of offers for patrons that are not likely to come back for longer periods of time. Instead of sending the patron monthly offers, they can send quarterly offers with longer valid windows that allow more time to book. For instance - if the patron only comes annually around his/her birthday, we might only send an offer annually around the patron’s birthday. Conversely, campaigns might be created with the goal of increasing the frequency of visits from higher worth patrons. Casino marketing should have the goal of generating trips sooner than expected and converting patrons into more frequent visitors. Additionally, time to next trip analysis can be used to identify when it has been too long and the business is at risk of losing the patron. In this case it might be useful to send an offer using “last chance” “we miss you” message. The offer might also need to be slightly better than what the guest has received in the past. By knowing when a patron is likely to return, we can adjust marketing strategies appropriately in order to save money on mailing costs, retain guests and increase loyalty.

One of the key benefits of using analytics in the casino industry is in allowing casinos to do micro-segmentation of their base. For example, groups that never touch the casino floor might be just as valuable if they utilize your other revenue-generating outlets during off-peak periods, or bring heavy gamers with them. There can be weekday and once-a-month segments, spa guests from a drive-in market that prefer to eat in specific restaurant, etc. This micro-segmentation allows casinos to be much more laser-focused in your marketing and service efforts, and your patrons will feel that you really understand them. Further, new patrons exhibiting similar behavior to your valuable segments can be identified and nurtured, so casinos do not have to build a patron history before they are able to do any analysis. Behavioral indicators signal an opportunity for the casino to take action to encourage or discourage expected behaviors. For example, churn models predict when an individual patron is at risk for leaving, so they can take appropriate action before they defects. And once patron value is known, present and projected, his affinities and behaviors as well his demographic characteristics – casino can use analytic to figure what is the next best marketing action that should either drive his behavior, maintain and increase his value, or to stop him churning to competitor.

No Comments Yet.

Leave a comment