Win-Win Marketing

Give your sales force more receptivetarget markets while focusing onlong-term profitability

You work for a large company with many products and customers. The CEO calls you into her office. “You IT people have a lot of customer and company data,” she says. “What real insight can you deliver to our sales force to help it present the products a customer wants and promote our long-term bottom line?”

A sale is a meeting of minds and ideally is a win-win situation, advancing the goals of both the company and the customer. This lab piece will report on a “profit-focused” approach to cross-sell modeling that considered product profitability, cast different sizes of nets in grouping prospects, and matched customer characteristics to favored products. This method showed a three-fold increase of projected returns over a simpler customer identification model.

To deliver to the company’s needs to maximize long-term profits, you need an analysis of product profitability streams and an understanding of who might want a product. To deliver to the customers’ needs (and thus the company’s too), you need to know what kinds of products their current situations or conditions merit. For the sales person (real or virtual), you need to be able to suggest product mixes that best fit the customer’s needs and conditions and offer the most return.

This project was based in part on work done for a number of banks and investment companies. Thus the profit analyses described were for situations where customers tend to hold products for a long time generating an even incremental revenue stream. Other product patterns could easily be overlaid.

Expectation Value and More

The other day I had this conversation with my friend Joe.

Joe said: I’m going to flip this dollar coin. What would you pay for the right to pocket it if it comes down heads?

Forty cents, I replied.

Joe: What? How did you come up with that?!

Half the time it will come down heads. A 50-percent chance of winning a dollar is worth 50 cents to me. That’s the expectation value, the probability times the payout.

Joe: So, what’s that 10-cent difference for?

Five cents for overhead and five for my profit margin.

Joe, unable to resist a gamble, took my four dimes and flipped the dollar coin. That’s my kind of game.

There is an essential difference between a businessperson and a gambler when it comes to dealing with uncertainty — something both have to do. The gambler gains some personal satisfaction even if over the long run he loses. The businessperson, however, is in for the long haul and wants to leave room for a profit margin on top of all risks and costs.

As with my valuation of the coin gamble, there are two main components to analyzing the potential impact of marketing additional products: probability and value of a customer response. In addition, you should make sure you don’t miss any potentially profitable customers.

To calculate long-term value, you need to estimate a customer’s likely term of stay — in terms of, let’s say, years — and what each year means for the bottom line.

Longevity: Mean Holding Time

One day, while working for a large bank, I was notified that a major pension fund, with which we had never done business, had deposited 10 billion dollars. After three months, the fund wanted a different approach. Two months later it jumped across the street. This tale makes a stark point: Customer loyalty and therefore the longevity of accounts is at least as important as the size of accounts.

In a study, I found one of the best indicators of a customer’s likely stay is the number of accounts the customer holds. Here, for simplicity, I used the formulation that on average best fit all the customer sets I examined:

Mean longevity in years = a multiplied by the square root of n;

n = number of years;

a = a constant that needs to be estimated in each context.

A person with four accounts, on average, will stick around twice as long as a person with one.

Profitability Per Year

A detailed analysis of a company’s earnings and each product’s margin is important to have, and an OLAP environment is particularly suited to yielding such an analysis. Again for simplicity, I used another good indicator developed from experience: The sum of absolute value of all accounts was a good proxy for the customers’ profitability. The average is a good indicator of the amount of money likely to be put into new products. To make the analysis more precise, I added differential weighting factors by calculating the profits on broad classes of products.

Now that you have methods for estimating the long-term value of new product accounts, the next step is to determine the comparative probabilities of customers accepting offers.

Data Mining for Interest

One approach to this calculation is to ascertain groups of customers who should be interested in a product and then determine their interest level. Once you define the group, the interest level (probability) comes from a simple frequency calculation: The number of people in the group who have already bought the product, divided by the total number of people in the group.

You can criticize data-mining methods for their pattern-recognition origins; rather than aiming to optimize general solutions, they are set to obtain the best classifications or patterns. The problem becomes more apparent when you look at the specifics: If most of the people classified to want a product already have it, the list of prospective candidates remains small. If you look for new candidates among those individuals most like the ones who already have a product, you may not find the most profitable prospects. However, it is possible that some candidates who are just a bit less similar, and who are excluded from the pattern, could be very reasonable and profitable prospects. (See Figure 1.)


FIGURE 1 A pattern-recognition method puts in a constraint (vertical line), to separate categories of customers. This breakdown leaves out nearby potentially profitable customers.

Furthermore, considering more prospects is better than less. But looser models will not give good estimates of customers’ interest levels. Here, I used a nested set of regions, each representing a more tightly defined explication of who has the product. Each individual obtained the rating of the best region he or she landed in. (See Figure 2.)


FIGURE 2 Loosening constraints brings more prospects into consideration.

Decision trees have an advantage here. They are designed to generate a hierarchically ranked set of relevant variables. The first variable split does the first best job of dividing those who have the product from those who do not, and so on. With an ordering of variables’ classification value you know which variables to drop or relax to increase the number of customers described with the smallest loss of regional homogeneity. By dropping the least important variable first, you can gain the greatest increase in model generality for the lowest cost in forecast power.

This study took three differing tree models set for three different levels of generality or inclusiveness. Customers fitting the tightest test criteria got the certainty rating of this most homogeneous group. Customers identified as a prospect in the second tightest group but not the first received this group’s certainty score. To be more conservative, we used two test samples for each model and used the lesser probability rating to assign a probability to each region.

We then had all the parts: a comparative probability measure and a measure of long-term profitability as constructed from multiplying annual profitability by mean longevity. Each customer gets a profit-gain projection for adding any of the products. A ranking of top product recommendations and profit levels for any customer can be part of the information delivered to the sales end.

To test this “profit focused” approach it was compared to a control model. The control used a single decision-tree analysis for each product and the same customer attributes as the more sophisticated analysis. In this lab test the estimated total of added income from the more sophisticated marketing method was nearly three times more than the control. Actual marketing campaigns taking advantage of differential product profit values have been measured to beat those without by as much as 50 percent.

Furthermore, the chance a customer might accept an offer is also a function of how it is proposed. For example, if family size is one key predictor of the need for personal loans, and an increase in family size is the reason a person is being reclassified as to loan interest, this is something the sales person really ought to know. So this kind of analysis can further add directly to the impact of the sales process.

Much of the work described here was done on a Data General Corp. Clariion storage array, using 400MHz Xeon Dell/Intel four- and two-way servers, WhiteLight Systems Inc. OLAP, Belmont Research Inc.’s TableTrans, and SPSS Inc.’s Clementine data-mining software. For more detail, see www.dsslab.com

Barry Grushkin is a researcher at the DSS Lab (www.dsslab.com) in Cambridge, Mass. You can reach him at bgrushkin@dimsys.com.

Much of the work described here was done on a Data General Corp. Clariion storage array, using 400MHz Xeon Dell/Intel four- and two-way servers, WhiteLight Systems Inc. OLAP, Belmont Research Inc.’s TableTrans, and SPSS Inc.’s Clementine data-mining software. For more detail, see www.dsslab.com