## Scoring a File After Building a Model

May 10th, 2013 by: DSA

Two difficult issues come up frequently in response modeling: how to take different offers into account when building the model from a promotion that included multiple offers, and how to go about scoring the promotion file for future use when you may want to use different offers at different times – including new offers that you haven’t tested or even thought of yet.

To begin addressing these questions, let’s assume we’re building a direct mail ZIP-code-level customer-acquisition model for a continuity media club (this means the only things we know about the persons to whom we are mailing are the census data associated with their ZIP code) and that the mailing on which the model will be based went to 100,000 persons.

As noted, the total mailing pulled 3 percent, with the \$2.95 offer pulling 5 percent, the \$4.95 offer pulling 3 percent and the \$6.95 offer pulling 1 percent. The first decision we have to make is whether to build three models, one for each offer, or one model that in some way captures the effect of the different offers.

The statisticians will tell you that in general it’s better to build one model with a larger sample size than to build three individual models each based on a smaller sample size. What’s more, in this case the sample size for the \$6.95 offer is so small that a separate model couldn’t be built even if we wanted to build one. So, the model that we finally come up with will have to “capture” the effect of offer. We’ll show you how that is done later.

Let’s further assume that we decide to start off the modeling process by asking our statistician to bear with us and select only variables we personally think are important. Then we decide to start with the variable AVERAGE HOUSEHOLD INCOME IN THE ZIP OODE (AVGHHINC). And we run our model and discover that the variable is “significant” and that the model looks like this:

Response= a- b(AVGHHINC)

You read this equation as follows: a person’s expected response is equal to some number “a” minus some other number “b” times the average income that exists within the ZIP code in which the person resides. This indicates that response to this promotion is inversely proportional to income. Said another way, the higher the average income in the ZIP code into which we are mailing, the less likely there will be response to our continuity offer. A not uncommon finding.

Let’s try another variable – percentage of married women in the ZIP code (%MAR). We run this simple model and find the following:

Response= a-b(%MAR)

In this new model we have different values for a and b, but again the negative sign indicates that as the percentage of married women in the ZIP code increases, the response rate to our continuity promotion goes down. In our model it now looks pretty clear that we are doing better, at least in terms of response, in downscale markets where income is low and the percentage of unmarried women is high.

Next, let’s assume that we have exhausted our look at variables one at a time and that these two demographic variables are the only ones found to be significantly related to response. Now, we ask our statistician (who by now is going crazy because this isn’t really the way you would proceed) to build a multivariate model attempting to include both the income and marital-status variables. With the result being:

Response = a -bl(AVGHHINC) – b2(%MAR)

In this two-variable model, the values of the “a” and the “b’s” would again be different from our earlier models, but the negative signs associated with the b’s would continue to indicate that as income goes up and as the percentage of married women in the ZIP code goes up, response goes down.

Up to this point, we have said nothing about offer, but our traditional response analysis tells us that response is very significantly affected by offer. To build offer into the model, we have to create something called a dummy variable. Dummy variables can take on only two conditions: a yes represented by a 1 if the condition exists, and a no represented by a 0 if the condition does not exist.

In this case, any one person could have received any one of three offers. In situations such as these, where there are more than two categories, the rule is that you create one less dummy variable than the number of possible categories.

Let’s call these dummy variables D1 and D2. In the model-building process, each person will now have two variables added to their record: D1 and D2. One more step is required. We have to decide on a convention or rule for assigning the 1 ‘s and O’s.

But we can be arbitrary. Let’s decide that a person will receive a 1 on the dummy variable Dl if they received the \$2.95 offer. This means that anyone who received the \$2.95 offer will receive a 1 on D1 and will automatically receive a 0 on the dummy variable D2.

Let’s also agree that anyone who received the \$4.95 offer will receive a 1 on the dummy variable D2 and will therefore receive a 0 on the Dummy Variable Dl.

Persons who received the \$6.95 offer will receive a 0 on both dummy variables DI and D2.

Now we ask our statistician to the run the model again and we get the following:

Response = a – bl(AVGHHINC) – b2(%MAR) +b3Dl – b4D2

What do we have? Since both D1 and D2 appear in the final equation, we can assume that the modeling process found them to be “statistically significant” – as we knew they would be by looking at the response analysis.  Additionally, if we were using real numbers instead of letters, we would also see that the value of b3 (the coefficient of Dl) would be larger than the value of b4. We know this because the lift from the \$2.95 offer is larger than the lift from the \$4.95 offer.

We also know that this new model, which takes offer into account, will “fit” the data better because we know that response is affected by offer and that a model that failed to take offer into account, when offer was important, would have to do relatively poorly.

So, what about scoring the file? While the modeling procedure has quantified the effect of offer, what needs to be done now is to score everyone on the database under the assumption that one or more, or perhaps none of the offers, will be mailed to this entire file.

Let’s suppose we decide that we might want to promote some segment of the file either the \$2.95 offer or the \$4.95 offer and we agree that we will not promote the \$6.95 offer because of its low response rate. If both scores were computed for everyone on the file, given the additive nature of this particular model, and the files were sorted and ranked twice in descending order of each score, the rankings under each scoring system would be identical.

The estimates of each person’s probability of response to each of the offers should be fairly accurate, assuming a logistic regression model were used to create the final scoring equation, and could be stored on the database for future use when promoting these offers.

However, if all we want to do is rank the file in descending order of response, without regard to the absolute accuracy of the expected response rate, it would be sufficient to score the file leaving out the effect of offer (let D1, D2, and even a=)). Sort the file and create 10, or 20 or 100 segments and score only the segment number on the database.

This segment number can be used in the future when deciding how deeply you want to promote into the file. Unfortunately, this decision requires still another analysis. You begin this analysis by questioning what the response rate be would if you promoted the entire file and, given the power of the model(s) built (i.e., how much better or worse than average the response rate for each segment is), how deeply into the file you can reasonably expect to promote… assuming, of course, no difference in expected backend performance.

But that’s a subject for another blog article.

## Measuring And Modeling Attrition

April 3rd, 2013 by: DSA

Customer attrition is a universal concern of database marketers. However, depending on the kind of marketer you are,  the nature of the measurement and modeling problem will differ.

For example, continuity marketers – a category that includes ”of the month” products, collectibles and cable television - have a common problem in that their customers may leave at any time.

Mobile phone, subscription/magazine companies and credit card companies have a different problem in that their customers are usually prepared to stay for at least the term of their subscription, in the case of magazines, or for the first year , in the case of credit cards.

In other words, there is a specific renewal date or fee-billing date, and there is an understanding that the customer will have the option to continue or discontinue the relationship at that date.

Catalog companies are in an altogether different situation, one in which there is no obligation on the part of either the catalog company or the customer to ever do business with each other again (though the first-time catalog shopper will not be surprised to receive additional catalogs in the mail).

Let’s focus first on the continuity situation.  It’s not unusual to hear continuity marketers discuss attrition in terms of the percentage of the entire customer base that leaves each month. Usually this number is in the range of 2 percent to 5 percent per month, and it’s a meaningless, misleading and harmful number – especially when it’s defined as the dependent variable in an attrition model.

The problem is that in a continuity environment, controllable and predictable attrition does not occur at equal  rates across the customer base;  most of it takes place in the first three to six months.  The exact length of time depends on the unique business and the billing practices, but in general it takes a month or two for customers to make a decision about a product or service, and then another couple of months to do something about it and have that action appear in the database.

Attrition in this early period can be as high as 10 percent to 20 percent. After a customer has been with you six months (certainly after a year). attrition becomes so low that it is next to impossible to predict or to act upon. Unfortunately, we can’t stop our customers from moving, losing their jobs or dying.

That’s why retention efforts – anti-attrition efforts – should be directed at new arrivals during-their first four to six months of the company/customer relationship. The next question would seem to be whether these high-percentage attriters can be modeled. In other words. can their attrition be predicted so that retention efforts can he directed more toward those customers likely to leave during this first-six-month period?

Technically. the answer in most cases is yes. Of course,  it depends on the data available. If demographic data is a good predictor of performance in your market, then that type of data is certainly available.  Other data sources such as the information collected on a coupon or application can be particularly useful. And, very early performance information – the amount of time taken to pay for shipment or first-month billing is generally predictive.

However, the real question is not whether early attrition can he predicted, but whether or not it makes good marketing and economic sense to limit a retention program to any segment of your new customers.

A new-customer retention program will undoubtedly involve making a special effort to welcome your new customers, make sure they understand your product or service, reinforce their purchase decision, and in general make them comfortable with your business.

It’s arguable whether this process should be limited exclusively to those new customers expected to leave your business for one reason or another.

In any event, I want to re-emphasize that you should stop thinking about attrition as an average, as if it happened with equal frequency across the entire customer file. Attrition doesn’t work that way, and before you try to model it you need to focus in on your new customers, find out the exact points of weakness for your business and then develop marketing programs to persuade your customers to stay with you. Predictive modeling may or may not be a part of this process.

As for credit card attrition (I’ll leave magazines and catalogs for another day), modeling involves two issues: when to model and how to avoid “no-brainers”.

Most bank-card companies have found that they can predict “cancels” fairly well anywhere from three to six months prior to the fee-billing date. Modeling closer to the fee-billing date produces great models, but there is no time left to implement a targeted retention program.

Clearly, those customers who haven’t used the card at all or haven’t used the card within the last 90 days, will have a high predicted-cancel rate. If you let this population into the model, they’ll undoubtedly fall into the top deciles.  That’s OK. but you probably don’t want to keep these cardholders because they are unprofitable.

The retention effort should therefore be focused on members who are predicted to be in the second, third and maybe even the fourth deciles. Many of these “relatively” high-potential cancels will still be profitable if your retention program can convince them to stay.

## How To Develop A Customer Contact & Retention Strategy

March 6th, 2013 by: DSA

Last year we wrote an article on ‘Measuring and Modeling Attrition”. In that article we argued that if your objective was to cost-effectively reduce attrition from say 10% per year to 7.5%, you had to begin by focusing your retention efforts on those customer segments where the expected attrition rate was well in excess of the average, and that you began this process by building a retention model which would identify these segments.

But, we also argued, especially with regard to retention of new customers, that:

“…the real question is not whether early attrition can be predicted, but whether or not it makes good marketing and economic sense to limit a retention program to any segment of your new customers.

A new customer retention program will undoubtedly involve making a special effort to welcome your new customers, make sure they understand your product or service, reinforce their purchase decision, and in general to make them comfortable with your business.

It’s arguable whether this process should be limited to only those new customers expected to leave your business for one reason or another.”

In this article we would like to suggest a strategy for dealing with new customers that employs a combination of relatively new database marketing techniques, market research, and some old fashioned direct marketing testing practices.

Many if not most database marketing companies, be they clubs, continuity programs, catalogs, or financial service providers, treat all new customers the same way for at least the first three to six months of their relationship, some treat all new customers the same way for the entire first year.

After six months, or certainly after one year, enough is known about the customer to alter future contact strategies.

If this practice applies to your company, i.e., if in your company all new customers are treated the same way for the first six months to one year, then ask yourself this question:

“How much are you spending on customer marketing per month, and are you sure that this amount is the optimal amount? Should it be 25 % or 50% higher. Should it be 25% or 50% lower?”

Unless you’ve tested alternative spending strategies you won’t be able to answer this question.

But, before running off and testing alternatives, ask yourself other two questions. First, do you think all of your new customers have the same potential, i.e., the same lifetime value? Second, do you think that all of your new customers have the same needs or requirements.

If the answer to either or both of the above questions is ‘No”, which, of course, is the correct answer, then your course of action should be clear.

You need to learn how to estimate each new customer’s potential and also understand something about each customer’s special needs or requirements.

And both tasks are reasonably simple to accomplish.

Customer potential can be modeled based on the information you have about the customer at the time of acquisition. This information can be taken from applications or coupons, or can be predicted using models based on census or household data. If the first contact is a transaction, for example, a sale or an investment, then information about the transaction (the size of the investment, type of product purchased, etc.) can also be used in the initial prediction.

Customer needs and requirements can be learned be asking new customers to complete questionnaires contained in welcoming packages, or from information requested on the application, enrollment form or coupon. Segmentation based on this information can be relatively simple: for example, married couples with children versus married couples without children versus singles; or the segmentation can be based on mathematically defined clusters.

In some situations the initial segmenation may not take place immediately because not enough information exists to make a meaningful judgemnt. In these situations you still want to test alternative initial strategies, and then more your customers into their proper segments as soon as their early behavior, or their response to your questionnaires, indicate that there is a basis for segmentation.

Once customers can be segmented in terms of their potential and/or special requirements you’ll find that it is a relatively easy task to decide on two or perhaps three alternative strategies to be tested against each segment. Focus group research can help you define the alternative strategies but only testing will tell you which strategy is most cost effective for each segment.

Obviously, your early predictions will not be 100% accurate but your highest potential customers will generally distinguish themselves from your average and low potential customers. Over time, as more behavior information becomes known, customers will move from segment to segment and your contact strategy with regard to each customer should change acccordingly.

The process of testing alternative marketing strategies is time consuming but straightforward. The test may take six months to a year or longer to evaluate, but the results are worth the effort. For example, the next time somebody asks you if you are spending too much or too little on customer marketing, you will be able to answer them and support your answer.

## Modeling, Managing And Marketing To Unique Customer Groups

February 5th, 2013 by: DSA

To a large extent, the role of predictive modeling deals with the issue of selecting customers or prospects for a single promotion, or a series of promotions. The underlying premise being that some promotion universe is about to become the target of a single offer, or campaign, however all available names will not be promoted for reasons having to do with either: (a) budget, or (b) profitability.

In the first instance the budget is not sufficient to promote all of the names available, so a mechanism is needed to identify the most profitable segments within the promotion universe. In the second instance the budget may be sufficient, but not all names can be promoted profitably. That is, if everyone were promoted, the response rate would fall below breakeven; so a process or a model is needed to identify segments of the promotion universe whose expected response rate is above breakeven, or some agreed upon rate of return.

This scenario is useful for teaching purposes, because it is simple and our attention can be focused on the statistical issues involved in building reliable models. But the real world is more complicated. Not all customers and/or prospects should necessarily be placed into the same promotion universe against which a single selection model will be applied.

For example, we all know of companies that produce only one monthly or one quarterly offer, catalog, etc.  All customers selected for promotion receive the same promotion. And, in some cases this same very same promotion vehicle, especially if it is a catalog, is sent to prospects as well as to customers. (In this context prospects are defined as unconverted leads or inquirers, i.e., names on the database that have not yet made their first purchase, as opposed to rented names or compiled names.)

By now, everyone has figured out some strategy for deciding who gets promoted and how often. This contact strategy may be based on common sense, on simple or complex RFM models, or on predictive statistical models. For example, a company may choose to promote unconverted leads twice a year, and “model” the customer base so that one third of the file receives 12 catalogs a year, another third of the file receives 8 catalogs a year and the bottom third of the file receives only four catalogs a year.

While this approach to managing contact strategy is a step in the right direction, the problem is that the only marketing element being tampered with is frequency of contact. Anyone and everyone who is promoted is shown the same thing. And, that’s the problem we want to address in this section.

The hypothesis to be tested and hopefully proved is that not all leads and not all customers should receive the same promotion. To throw all customers and/or prospects into the same pool, develop a frequency strategy for individual groups, and then promote them all the same promotion is easy to do, and efficient to do from a cost perspective, but not necessarily the right thing to do.

One way to approach this problem is to first divide the potential promotion universe into customer and or prospect groups to whom you would intuitively do different things.

Some of the more obvious groups are:

1. One time inquiries or leads from cold catalog promotions or inquiries, or print, or broadcast ads;
2. Leads that have responded more than once to your promotions and/or advertising but have never purchased;
3. Customers who have made only one purchase;
4. Customers who have made multiple purchases but from only one product line;
5. Customers that have purchased from multiple product lines.

To test the hypothesis that profitability can be increased by treating different customer segments differently you need to establish specific objectives, and then develop different promotion strategies to test against your current control strategy. For example:

1.  For one-time inquirers you might be looking for more information before attempting to close your first sale. Depending upon the characteristics of your product or service, different strategies for converting leads into buyers may be required. In the financial services area we have seen situations where it is more cost effective to use a two-step questionnaire and conversion strategy to reactivate dormant leads than it is to attempt to sell the leads directly.

2.  For multiple inquiries or inquires that have responded to your request for more information, your objective must be to close that first sale. A very targeted low risk (to the customer) offer may be the way to achieve this objective.

3.  For one time purchasers, you need to develop the buying habit, to turn a one time sale into an on-going relationship. One time buyers are a major problem (or opportunity) for many database marketing companies. All too often one time buyers do not think of themselves as your customers –you just happened to be there at the right time with right offer for a particular product that the customer needed right away.

4.  For customers that have purchased in only one category, you need to get them into the “other” sections of your store. This phenomenon is common to all businesses but especially in men’s and women’s clothing.  “I like their shirts, but I buy my ties elsewhere” is the kind of comment frequently heard in focus groups. Why not just send a tie to your best shirt customers, or at least a coupon with a substantial discount for trying a tie?

5.  For customers that have purchased from multiple categories your objective should be to understand the most common product affinities, and then develop customized promotions that take advantage of these affinities. The concept is simple: if most people who buy product X also buy products A, B and C…  find people who have purchased A. B & C but not X and offer them product X.

In situations where the number of customers and/or prospects within each target group is large it often makes sense to organize marketing responsibility around each of the target groups. For example, one person may be assigned the responsibility for converting one-time buyers into repeat purchasers. Another person assigned responsibility for reactivating dormant leads. Both difficult but critical tasks.

Assigning specific management responsibility to individual customer or prospect segments insures that each segment will receive the attention it deserves. This approach to organizing around the customer prevents management from focusing its attention on easier and often more interesting things to do, such as generating more new leads from advertising.

Companies that fail to make this commitment may find that they are building a large database of unconverted leads, or that their one-time buyers never make the commitment to become steady repeat customers.

## Database Marketing: Where Are We After 25+ Years

January 9th, 2013 by: DSA

Of course, no one really knows when database marketing actually began to distinguish itself from traditional direct marketing, but somewhere around 1985 is probably not too far wrong. Some ‘database historians” trace it back to the start of the American Airlines Frequent Flyer Program which began in 1981 and their Gold and Platinum programs which we think began in 1985. (The best the folks at American Airlines could remember is that they began issuing upgrade stickers in August of 1985 — how’s that for marketing trivia.)

Another metric, albeit a rather personal one, relies on the history of the DMA Seminar Programs. In 1983 Rick Courtheoux began teaching the database course at the DMA, and we began teaching the Statistics and Modeling course in 1987.  Listening to participants describe their reasons for being in the course provides an interesting insight into the acceptance of database marketing over the years.

Back in 1987, very few participants came from companies that were actually doing modeling, but they had all heard about and were interested in learning more. Today, nearly every participant, primarily users rather than builders of models, comes from a company that is actively using models, either internally or with an outside vendor, and they want to know how to evaluate the models presented to them. To meet the demand for even more information, in 1993 we introduced the Advanced Modeling Course, just for statisticians and modelers, and that would have been impossible just a few years earlier.

By the same token, in 1987 very few companies had marketing databases, if you define a marketing database as a consolidation of one or more business files with promotion files that are accessible to end-users without programmer intervention. Today, while not everyone has one, probably more than 80% do, the rest are thinking about it, and are pretty sure they want one. Interestingly, among those that do have databases the frustration level is still very high relative to data quality and end user software tools.

So, where are we really after 25+ years?

Just about everyone has bought into the basic precepts behind database marketing and those basic precepts include recognition of the fact that not all customers have the same needs, nor have they the same potential, and therefore they should not all be treated the same way. And, given the power of our database tools, our ability to segment customers and predict and track behavior, plus our improved fulfillment and production systems, that have changed the economics of marketing to small groups or even to individual customers, there is no reason to do so.

Nevertheless, not everyone is there yet.

Interestingly enough the barriers to good database marketing today seem to have as much to do with organizational issues and believe it or not, traditional direct marketing skills, as they do with computer technology and predictive modeling.

It’s one thing to know that not all of your customers should be treated the same, it’s another thing to know how they should be treated. Consequently, we find most companies struggling with the question of how to arrive at an optimal contact strategy for meaningful segments of the database.

Ok, they say, these are my best customers, measured in terms of either actual or predicted performance and segmented in terms of attitudes and behavior, and maybe even by demographics.

Now what? How much should I spend on them, and what should I say to them — different from what I say and do to my average customers, or even my worst customers. And the simple, and perhaps the only answer is to test alternative strategies, which gets us back to old fashioned direct marketing testing which isn’t bad, but does seem to be becoming a lost art as we collectively tend to focus on the more esoteric aspects of database marketing.

Then there are the organizational issues. Just about everyone concurs that any customer base can be broken into meaningful segments — using either mathematical models or common sense. Common sense segmentation based on demographics and summary measures of purchase behavior generally suffice, mathematical models are required when the segmentation is based in large part on detailed product level purchase data.

But regardless of how you segment, any type of segmentation leads to segments which lead to segment managers which leads to inevitable conflicts between established product managers who have budgets to spend and goals to meet and the newly formed segment managers who mayor may not goals and frequently do not have budgets.

The conflict is inevitable, but it is not necessarily bad. In fact, the conflict can be healthy because attempts to resolve it require a corporate commitment to an environment in which both product managers and segment managers understand their roles and responsibilities and both have their own budgets, goals and objectives.

In this new organizational environment (actually similar to the relationship between manufacturers and retailers) product managers would do well to think of themselves as manufacturers with their own sales objectives and their own national and regional advertising budgets (which influence both consumers and retail decision makers); and think of the segment managers as retailers (also with their own promotion budgets and sales objectives) who need to be sold on accepting their product, and who have the power to decide which products to display and how to merchandise them, i.e., in direct marketing terms, which products to promote and how to promote them.

Think about it, you’ll see that it works. In this framework the segment manager owns the customer and controls direct contact with the customer. It’s the responsibility of product manager to supply the segment manager with products that will sell in the segment managers’ segment. If the segment grows both in terms of sales and profits, it’s because the segment manager has access to the right products for the segment and has promoted those products correctly.

When this happens the segment manager wins, the product manager wins, the segment grows, and the company grows. It’s just that simple. Well maybe it’s really not all that simple, so we’ll have more to say about segment management and contact strategy in future articles.

## What Is This Thing Called “EDA”?

December 10th, 2012 by: DSA

Regardless of whether you are building a traditional statistical model (multiple regression, logistic regression, etc.) or planning to use a neural net solution, it’s very important to subject your data to a thorough analysis before the actual modeling begins.  EDA or Exploratory Data Analysis encompasses all of the work the analyst does Before the start of the final modeling process.  Having said that we need to define what we mean by “the final modeling process”.

The final modeling process is the calibration of the regression equation or the neural net solution.  It happens after the analyst has decided which variables he or she will consider to be eligible for inclusion in the model, i.e., it happens after the EDA is completed.

This definition implies that not all of the variables available to the analyst are necessarily candidates for inclusion in the final model.  Some may be eliminated, some may need be “transformed” in some way, and some new variables may be created from the original set of variables.

This definition also suggests that the very last thing you want to do is drop all available variables into either a regression procedure or a neural net solution.  Trust us, that would be a very bad idea.

Now let’s get more specific.

There are four basic steps in the EDA or Pre-Modeling process.

1. The analysis of all available independent or predictor variables without regard to the dependent variable — the variable being modeled.

This step provides a basic understanding of the data, which in turn is the groundwork for segmenting and modelling the universe.

The analysis of the set of potential independent variables involves graphical displays.  Variables are studied one-at-a-time using Stem- and Leaf, and Box-and- Whisker plots. These graphics provide a host of information:

• simple data checks for erroneous input
• posting of minimum and maximum values
• gaps or missing values
• clumps or concentration of values
• depiction of the shape of the data.

Included in this set of potential variables are new variables that the analyst creates that are intuitive combinations of existing variables.   This process frequently results in more powerful or more predictive variables.   For example, the simple step of dividing Sales- To- Date by the Number- Of-Times Mailed or by Days-On-File results in more meaningful and therefore more predictive variables.

2.  The analysis of the relationship between each independent variabJe and the dependent variable

For categorical variables cross tabs (or univariate CHAID analyses — see our discussion of CHAID below) are used to depict the relationship between category levels and measures of the dependent variable.

For continuous independent variables a correlation analysis or a correlation table provides an index number (the correlation coefficient) which measures the degree of the linear relationship between each independent variable and the dependent variable.

Obviously many significant relationships are not linear.  Some relationships may be better described by a curve, or by two or three relatively straight lines, with each having different slopes and intercepts.

Scatterplots (graphs with the independent variables plotted on the x-axis and the dependent variable plotted on the y-axis) could in theory show this relationship, but scatterplots involving huge datasets are next to impossible to read and interpret.  Therefore, a technique referred to “smooth-plots” is employed for this purpose.

Basically, a “smooth plot” involves:

• sorting an independent variable (income, age, prior sales, etc.) in descending or ascending order
• slicing up the resulting file into ten or twenty equal pieces
• calculating the average value of each slice of the independent variable and the average value of the dependent variable corresponding to each slice.

The results are then plotted on a graph, and if the plot is not linear its underlying pattern is easily observed.   The next step is for the analyst to try different transformations (change the value of the independent value to its log its square root, etc.) to arrive at a more linear relationship .

The importance of linearity has theoretical underpinnings, but the intuitive reasoning is that straight-line relationships are easy to ‘see’ and therefore easy to analyze and put into segmentation and predictive models.

3.  The Analysis of sets of independent variables with or without regard to the dependent variables.

In many database marketing situations there is a strong correlation among the independent variables, this is particularly true when demographic data, especially census data, make up the set of available independent variables.   And, in most if not all modeling situations highly correlated independent variables are to be avoided.

In other situations there is simply too much data at the detail level (sku data for example) to include in a modeling exercise.  Fortunately, there is a technique called Principal Component Analysis which can be used in these situations.

Briefly, Principal Component Analysis creates new variables or components from the many original variables such that the components retain the most of information of the original variables.   The components – actually weighted sums of the original variables – can now be interpreted to reveal the relationship among the original variables.   These new variables can be used later on in a regression model or in a neural net solution.

4.  Segmentation of the universe with respect to the dependent variable (response or some measure of performance such as sales, payments, retention, etc.).

This step provides separate, distinct and non-overlapping groups of prospects/customers which are homogeneous in terms of response and/or performance.

The segmentation provides:

• Additional insight into the potential predictive power of individual variables.
• The not necessarily intuitive identification of new variables that are combinations of existing variables, better known as interaction variables, which may have potential predictive power.
• An expected maximum spread of predictive power for a given model (i.e., a benchmark to evaluate a regression model or neural net solution).

Segmenting the universe by the pre-selected criterion or dependent variable is performed by the decision tree analysis called CHAID.

Briefly, CHAID is a statistical technique that partitions or segments a population (prospect/customer universe) into separate, distinct and non-overlapping groups, such that the means of the dependent variable for all groups have the greatest variance or spread, i.e. the group means are as different from each other as possible.

The partition is based on the relationship between dependent variable and predictor or independent variables (whose selection is in part guided by the results of the earlier analyses).  The partitioning is done sequentially. CHAID finds the best predictor to make the initial division of the population, resulting in say k subgroups.

The k subgroups are in turn treated to k separate CHAID analyses.  For each subgroup, CHAID starts from ‘scratch’ finds the best predictor to maximize the spread of the mean of dependent variable for that subgroup. This splitting of the groups and subgroups continues until pre-set statistical stopping rules are met.

## How To Tell If You Have A World Class Marketing Database & Decision Support System

November 5th, 2012 by: DSA

From time to time we’re asked for our definition of a ‘World Class Database”.  This question frequently comes from the large accounting and consulting firms who have pitched a Fortune 100 firm on the need to have one, and now the client now demands to know what it is before they invest in it.  Is it bigger than a data warehouse?, does it come with built in data mining facilities?, is it managed by a self-learning neural net?, etc., and of course is it relational, or at least relational-like?

And, as all of us who have had any real experience with databases know, there’s no single answer to this question.  Everybody’s definition will be a little different and every vendor’s definition may well have something to do with the capabilities of their particular system.  With this caution in mind, we offer the following definition.  Actually, it’s not a definition at all but rather a listing of the capabilities that we think characterize a ‘World Class” Database.  So, if your database can do all or most of the functions listed below then you may truly have a ‘World Class” database system.  If it can’t, then it may be time to upgrade your system.

Here then is our list.  The functions or capabilities listed are ”kind of” in order of importance, but that’s hard to do after you get past the first ten or so.  Our comments on our own list are in italics, where appropriate.

1. Provide users with access to the database via a ”user friendly” graphical interface. Admittedly, user friendly is in the eye of the user.

2. Select names for promotion based on any user defined criteria:  ad-hoc queries, scoring models, pre-determined schedules, unexpected events. This is a base line capability.

3. Split names selected for testing and output coded files for mail, email, telemarketing, etc. production.

4. Prioritize the name selection procedure so that names selected for two or more promotions within the same promotion period are allocated to the most profitable use.  This assumes that someone in the organization has figured out how to do this.

5. Support suppression logic and controls so that no customer is contacted more frequently than the corporate standard permits.  This assumes a corporate standard exists.

6. Keep track of responses and perform response analysis.

7. Profile any customer or prospect group according to any other variable or set of variables in the database.

8. Maintain survey data (business or consumer) as well as overlay demographic and lifestyle data for profiling and modeling purposes. Survey data is tricky, especially when the surveys keep changing.

9. Calculate the to-date lifetime value of a customer and the expected incremental lifetime value of that same customer.  The shorter the incremental period, the greater the accuracy of the forecast.  Long term predictions of the lifetime value of individual customers are tricky at best.  Especially, if your strategy is designed to change customer lifetime value.

10. Calculate the to-date and expected lifetime value of all customers acquired from any defined source and calculate the return on promotion expenditures.  This is relatively easy and is the basis for all Return-on-Promotion calculations.  If you can t do this you probably aren’t ready for a database.

11. Produce sample files suitable for statistical analysis using more advanced modeling tools such as SAS, SPSS, CHAID, KXEN, R, etc.

12. Facilitate the testing of alternative contact strategies designed to maximize the return on marketing dollars.  Always be careful of words like facilitate and support.  They often mean the database won’t go out of its way to prevent the analysis from happening.  Testing alternative strategies is no less important in today’s database environment than it was ten or twenty years ago. In fact it’s probably more important to insure that our segmentation schemes really work better than the ‘old fashioned” control strategy.

13. Keep track of customers according to their segment, original source code, or test cell.  Critical, but very hard to do in many industries such as financial services where the original source is unknown and re-investments are hard to track.

14. Keep track of Customer Loyalty Programs, issue rewards and facilitate the evaluation ofthe program. There’s that word facilitate again.

15. Export data to other computer platforms or PC systems including other databases, word-processing and spreadsheet programs.

16. Maintain promotion history data, both mail and phone and record phone disposition results.  This is important.  A marketing database that does not keep track of past promotions, especially telemarketing promotions, is not a marketing database, no less a ‘World Class” database.

17. Provide timely responses for decision support by answering the most complicated queries in seconds or minutes, as opposed to hours.  Timely is important.  If you’re waiting weeks for answers from IT you might think that you would be happy with turn-around times measured in days or hours.  You won’t be for long.

18. Allow new files to be added to the database and accept changes in the structure of existing files without extensive re-design and/ or re-programming.

19. Perform updates by either replenishing files or updating only changed records.  Marketers always focus on access and response time, but update times are equally important, if not more important, especially if the update means that the database is down.  Thus, the next point.

20. Perform updates, at night or on the weekend, while continuing to provide users with access to the database during normal business hours. When this is not possible two versions of the database may be required. A current version available to users and a version always in the process of being updated.

21. Score database records using statistical models that have been developed outside the database.

22. Rank database records by model scores and assign records to percentile groupings based on scores.

23. Add fields to existing tables or new tables to database as new data elements and data relationships are identified.

24. Provide flexible updating at the field, table or full database level.

So, those are the functions and capabilities we associate with a ‘World Class” database.  Can you think of others, or do you think some of the capabilities are costly and unnecessary.  Let’s hear from you.

## When You Model Response, What Are You Modeling?

October 8th, 2012 by: DSA

One of the simplest predictive models to build is a response model based on only one previous promotion — provided each response represents equal profit potential.  Unfortunately, this is hardly ever the case.  In most database  marketing situations the initial response to a promotion is only the beginning of the story, and you need to consider the ”rest of the story”.

For example, many database  marketing situations involve responses that are really inquires or requests for more information.  The initial response, the inquiry, generates a stream of follow-up customer contacts.   These communications result in a conversion rate which in turn results in a net order rate.  For example, a 5% response rate coupled with a 20% conversion rate yields a 1 % response rate.

For the modeler (and for the marketer) the options are to:

1. Build one model to predict response, and one model to predict conversion, and then multiply each person’s expected response rate times their expected conversion rate, to arrive at their expected order rate.

2. Build just one model based on the net order rate.  In the example referred to above we would model the 1 % net order rate.  To model a response rate, or any other outcome, means to find segments of the promoted population whose performance (response or sales, etc.) is significantly above or below average.

What’s the correct decision?   There isn’t one.   But the correct approach is to build models of all of the possibilities and see which strategy works best for your particular situation.  While this seems like a great deal of extra work, it really isn’t.  By the time the data set is checked for accuracy, and by the time the variables which predict response and conversion have been identified and corrected for such things as non- linearities, interactions, multi-collinearity, missing values, etc., the amount of incremental work required to produce all possible models is minimal, two to three days at most, if that.

In some cases you will find that the net order rate is so low and the number of orders so small that it’s not practical to model net orders directly.  In other cases an examination of the variables that are correlated positively with response will be found to be correlated negatively with conversions.  For example, low income may be correlated positively with response but negatively with conversion, or net orders.  In this situation the process can produce insights into how your business works, and therefore how it can be improved.

Another useful result of modeling both response and conversion separately, is that the conversion model can often be used to manage or direct your fulfillment effort.  Prospects with a high probability of conversion may be sent more fulfillment efforts than prospects with a low probability of conversion and vise versa.  For example, many companies send the same number of conversions to all inquirers; these companies could dramatically increase their bottom line results by redirecting their conversion efforts.   Not only could prospects with high probabilities of conversion receive more efforts, but some of these efforts might be by phone, a medium that could be unprofitable if applied to all prospects.

Another situation that calls for going beyond simple response models, are new customer acquisition mailings which result in customers with wide variations in lifetime value (the net present value of the contribution to overhead, advertising and profits.)   Clubs and Continuity Programs are the classic examples of businesses that use ‘bribe offers” such as 4 books for \$1.00 or 12 DVDs for \$0.99.  These offers result in the club or program making an investment in each new customer, and that investment is expected to be recovered over the life of the customer.

A club or a continuity program that modeled only response could quickly put themselves out of business, because their models would tend to attract customers who were most interested in the bribe and least interested in the offerings of the club or program.   Therefore, these kinds of companies must build models that predict both response and performance.  Even then a fair amount of skill and judgement must be used to apply these models correctly.

Because customers with high probabilities of response will also usually have relatively low lifetime value estimates the indiscriminate use of both models could result in all prospects having the same expected net value per name mailed, making the models useless.  In these situations the models have to be applied on a list by list basis.  In some cases only the response model will be used — when response is the problem and performance is above average.  In other cases, lists with high response rate, the reverse situation might be true and only the performance model will be used.

The bottom line — think twice before building a simple gross response model.

## Making Segmentation and Predictive Modeling Work Together

September 10th, 2012 by: DSA

To make your Customer Contact Strategy as effective as possible you may need to use a combination of both segmentation and predictive modeling.  Here’s a quick review of how they differ and how they can be made to work together to maximize your return on your marketing dollars.

Segmentation
Segmentation involves assigning people to groups such that the people in anyone group are similar to each other and different from the people in any other group. Geographic areas, as well as people, also can be segmented. Claritas’ Prism product and their many competitors are examples of geographic segmentation.

Frequently, segmentations are based on lifestyle as well as demographic characteristics. Thus, a segment might consist of men between the ages of 39 and 54, with incomes above \$50,000, who live in urban areas, employed in white collar jobs, and who like to play golf, travel and drink white wine. This segment is obviously different from a segment that consists of women under 35, employed in low wage factory occupations, with a completely different set of leisure activities. And, and this is the point of segmentation, you would certainly want to speak to each segment differently, and tailor your offer to meet the needs of each segment.

This kind of segmentation can be done using simple common sense sorting rules, or by using complicated mathematical techniques such as factor and/or cluster analysis.

Who’s Responsible For A Segmentation Model
Additionally, it should be noted that Segmentation, as opposed to predictive modeling, to be discussed shortly, is much more arbitrary than predictive modeling. Forex:ample, it is up to the modeler (preferable a statistician working with a marketing person) to decide how many segments to include in the final segmentation solution. Going back to our upscale male example, that segment itself probably could be divided into two or three segments. Whats more, different mathematical clustering schemes can produce different results, so a great deal of care and attention needs to be used when building segmentation schemes based on lots of data and mathematical models.

The final segmentation scheme you decide upon should have as much to do with how you plan to market to each segment as it does with the mathematics of the solution. In other words, marketing must play a major role in the selection of the final segmentation. It makes no sense to have an eight segment model, if your production facilities can only handle four segments, or if you can’t think of eight different marketing strategies.

Segmentation is not limited to demographic or lifestyle data. Segmentation also can be based on product purchase data and/ or survey results. For example, it’s common practice to segment files based on purchase data, either retail or mail order. Of course, the most frequent segmentations are based on survey results that probe respondents for their attitudes (feelings) and behavior concerning some product or service.

The key thing to remember is that segmentation says nothing about how the members of a segment will behave: respond to a mailing; convert from a lead to an order; stay with a program; make payments, etc. It does not directly nor reliably tell you what products you should offer the segment, or if your products will meet the needs of the segment, even if your segmentation is based on a needs analysis, though a needs based segmentation will of course give you more insights regarding products or services than a segmentation based on demographic and/ or lifestyle data.

Predictive Modeling
When the issue is behavior or more accurately predicting behavior, the technique used is predictive modeling, usually some form or regression analysis, coupled with old fashioned direct marketing testing.

Predictive modeling allows you to find portions of a population that will do better or worse than average. Frequently these “portions of the population” are also called ”segments” and that may be where some confusion arises.

For example, if a promotion pulls 3% response, predictive modeling will usually allow you to identify 10% to 30% of the mailed population that will respond much better than 3% (say around 8%) and another 10% to 30% that will respond at a much lower rate, say 1%.

Predictive modeling can be used therefore in any situation where there is some measure of average performance. In B-to-B for example, if 25% of your leads convert to orders, then predictive modeling will allow you to finds ”segments” or portions of the population that will convert rates significantly above or below 25 % .

Finally, a customer/prospect database can be first segmented according to demographic, lifestyle, attitude or product purchase data and then predictive models can be built against each segment.

So, we might find that our upscale male segment responds on average at 5%, but through predictive modeling we can find segments that do better or worse than 5 % . Similarly, our downscale women’s segment might respond at 10% and through predictive modeling we’ll find segments that do better or worse than 10%.

In summary, Segmentation puts people into relatively homogeneous groups and this helps us in the ”tone and manner” of our presentations. Predictive modeling helps us decide how frequently we should contact any potential population or segment of the population.

A note of caution. Even though segmentation helps you understand who you are talking to, and therefore should help you develop a more targeted creative execution, it’s not always the case that a more targeted execution based on a segmentation, or, better still, a selection based on something as clear cut as gender, will produce a result that is more effective than a promotion that is less targeted, more general, more global in its appeal. Only testing will tell you that, and testing is no less important in today’s sophisticated database environment than it was twenty years ago, before we were building both segmentation and predictive models.

## 40 Things Database Marketers Should Never Do, Say Or Even Think About (Part 4)

August 7th, 2012 by: DSA

Recently, DSA was asked to come up with 10 great database marketing ideas.  We couldn’t think of ten great ideas, database related or otherwise, but it was easy to think of 40 or so things that database marketers shouldn’t do. Here the third installment of our  Top 40 list of things not to do. (Not in order of importance.)

31)  Never assume that block group or household level models will be more cost effective than zip code models.

The models will be stronger for sure, but not necessarily strong enough to cost justify the added costs measured in both dollars and extra lead time.

32)  Never model the obvious.

You don t need a model to tell you that targeted promotion customers buy more than print customers, or that middle aged customers buy more life insurance than those in their 20′s.

33)  Never trust anyone who tells you that their solution is proprietary and they can’t go into the details.

At least that’s our opinion.

34)  Never be product driven if you can be segment driven.

Focus on the needs of the customer segment and the right product will get to the right person, hopefully at the right time.

35)  Never model overlay data until you’ve modeled all of your internal data.

Model them both together and you won’t know if you could have done as well, or almost as well, without the purchased data.

36)  Never use a variable in a model that doesn’t improve your validation results, even if it’s declared statistically significant.

The fewer the variables in a final solution, the more likely the chance of the model holding up over time.

37)  Never use the results of small tests (check the confidence intervals) to roll-out large quantities.

Unless you want to lose your job.

38)  Never use dummy variables to code census variables or to classify purchase data when there are a large number of purchase options.

We’ll explain this in another article devoted just to this topic. In the meantime just don’t do it.

39)  Never assume the results of a model (the spread from top decile to bottom decile) will hold up exactly on a roll- out.

No one know why for sure, but it happens a lot, so plan for it.

40)  Never give a person with responsibility for marketing current products responsibility for developing new products.

Its too easy to work on current products and new products may not receive the attention they require.

Well, in the immortal words of Forrest Gump…  “That’s all we got to say about that”.  Please use our Top 40 list to avoid some of the pitfalls that we have seen over the years, and even come with some of your own items that we didn’t list.