Building a Smarter Response Model

You get an offer from a company such as MCI or Citibank, and several days later you get another offer from the same company, with the same or different terms.

This has probably happened to most readers. Looking at this problem from the marketer’s standpoint, most large marketing companies carry on multiple marketing campaigns offering different products and channels to every appropriate household in America. They don’t really want to send multiple offers to the same customers. This is wasteful, expensive and often confusing and irritating to customers who get the same offer with different options.

To avoid these problems, big finance and telecommunications companies recruit staffs of statisticians who build models for each mailing campaign. The idea is to improve response and reduce mailing costs. Running models before each mailing, however, is labor intensive and expensive. It can take weeks to develop good response models.

MarketSwitch (, in Dulles, VA, has found an innovative solution to this problem. It has developed three software products that major companies such as CapitalOne, MCI Worldcom, AOL and SkyAlland have adopted. Here is how they work.

The three MarketSwitch products are a response-propensity model for single offers, a cross-selling optimizer for outbound campaigns and a real-time offer optimizer for inbound Web and call centers.

The response-propensity model. This does what most response models do: It aims to increase the rate of return for a single offer by identifying prospects who are most likely to respond, based on available demographic and behavioral data.

The MarketSwitch model, using neural networks, logistic regressions and radial basis functions, is similar to other products on the market. Where it differs, however, is that it comes packaged in neat, easy-to-use software that can be run by marketers with no training in statistics.

The package predicts the return on investment of each campaign by calculating the net present value of respondents. It can accept what-if scenarios, including a limited marketing budget or differing goals such as maximizing profit, on the one hand, and acquiring the most customers, on the other.

Where MarketSwitch really takes off, and why it is attracting major clients, is in building on the response-propensity model to create its cross-selling optimizer and real-time offer optimizer.

The cross-selling optimizer. This product begins with a problem that confronts large U.S. marketers: making multiple offers across multiple channels to the same 100 million U.S. households. The company may have 10 or 20 product offers that are potentially available to many millions of prospects and customers. Which is the best offer to make to each person? Each customer may have a unique propensity to respond to a particular offer or channel (mail, e-mail, telemarketing, etc.). There are essentially three ways to decide which offer to make to each person:

• Eligibility (whether he is eligible for the product).

• Response (how likely he is to snap at it).

• Profitability (how much money we will make if he buys it).

In addition, the solution varies with the size of the marketing budget and the overall goals of the marketing program – profits or customer acquisition.

Using traditional models, there are two solutions.

In the first solution, you rank all customers by propensity to respond (or profitability) and market to the top few deciles – the ones that show a positive return on investment.

The second solution ranks the offers to find the one group of offers that provides the highest return (either profit or number of customers). You would make this group of offers to a large percentage of the database.

There is a third solution when using the MarketSwitch cross-selling optimizer. You optimize the offer for each customer, including your business constraints. This results in a complex pattern of customers and offers. Some people will get several offers. Many will get none. The software calculates and recalculates the offer to each prospect and customer until the net present value of the entire customer base is maximized.

The problem solved by this software is bigger than can be solved by normal statistical modeling. With 20 possible products and channels and 100 million customers, there are more than 2 trillion possible solutions that must be tested to determine which one maximizes profits. The cross-selling optimizer can solve the problem in about a day. The result is an improved response rate for a complex campaign of 10 percent to 30 percent. The savings for a major marketer such as MCI or CapitalOne can be more than $1 million on a single campaign. No other software product on the market is capable of solving such a complex problem in a reasonable period of time.

The real-time optimizer. The profit-maximizing problem faced by a Web site or a call center is even more complex. It requires immediate answers. With cross-sell planning, after all, you have several days to figure out the marketing solution that maximizes profits. How can you solve the same problem when a customer visits you on the Web or makes an inbound call? On the Web we can accumulate a great deal of relevant information about a prospect or customer. We may know:

• Previous purchases.

• Demographics, including age, income, home value and family composition.

• What he did when he visited the Web site last week.

When he comes back today looking at the site again, we have to decide what to show him. The content is totally under our control. We can show him sporting outfits or baby products, or something else. But we have to move fast. We have to decide:

• What categories of products will appeal to him?

• What products within his categories should we feature?

• What offer will get him to respond best?

• What will be most profitable for us?

This modeling problem has to be solved within a couple seconds to be of any use. This is what the real-time optimizer does for a call center agent or site. The software has to be powerful enough to determine the most optimal offer strategy in real time, yet simple enough for such strategies to be created quickly, by marketing personnel with little training, in an ad hoc fashion.

There are three different Web users who need real-time solutions:

• The Web publisher or portal.

• The corporate Web site.

• The ad network.

A portal is trying to improve its profits by offering banner ads that maximize click throughs while balancing commitments. For example, you may have promised a partner that you will show his banner on your site for at least 10,000 impressions per day. At the same time, 30 other advertisers want space on your Web page. You know a lot about many of the visitors to your site. You want to use what you know to offer each visitor banners that are likely to be most appealing, while at the same time meeting your commitments to your advertisers. When a visitor shows up on your site, you have a fraction of a second to optimize your real-time offer solution and get it up on the page for the visitor to see. This is where the MarketSwitch software proves its worth.

The corporate Web site has a similar set of problems. When a visitor arrives, you have to use what you know about the visitor to show him what he wants to see, and what is most profitable for you to sell to him. Every time a visitor visits, he may:

• Disappear in a few seconds.

• Linger over the site, checking out various options.

• Register his name and address.

• Buy something.

After each visit, you can use the corporate Web site ad optimizer to rebuild your model and strategy automatically based on actual results. Your solutions get better by the hour.

Ad agencies place banner ads on third-party Web sites and ad networks. On Madison Avenue, advertising agencies can only guess what happens when a prospect looks at a client’s print or TV ads. On the Web, the agencies know exactly what happens, and they know it within an hour after the ad’s placement. For that reason, Web advertisers expect instant results. They expect their agency to be on top of what is happening every hour of every day, fine-tuning the ads and their placements to maximize results and minimize costs. The agency can:

• Change the wording on an ad on a rotating basis.

• Change the wording on an ad to include personal data about the visitor.

• Drop the ad from a site entirely.

• Heavy up on placements on successful sites.

• Coordinate the optimizing strategy for this ad with the strategy for dozens of other ads that they have to place on the same sites at the same time.

The ad network optimizer measures the success of banner ads at the domain level and builds optimization strategies based on actual results and economic calculations. The software is able to readjust the optimization strategies hourly to maximize profitability. The resulting strategies are passed on electronically to the ad server, which automatically displays the chosen ad on the chosen Web page.

Does all of this sound futuristic? It does to me. Yet the software to do all of these things exists and has been installed in several major client call centers and Web sites. The software is expensive. Prices begin at $450,000 and go up to about $1.5 million for the largest users.

Why would anyone pay such high fees for modeling software? Because it works and saves many multiples of its cost in the first year. All of MarketSwitch’s current clients tested the MarketSwitch product against their own statistical team’s products before buying. They made a purchase only after they could see that this new product really did what MarketSwitch claimed for it.

• Arthur Middleton Hughes is vice president for strategic planning at MS Database Marketing (, a database marketing and e-commerce firm in Los Angeles. Reach him at [email protected]

Related Posts