Preparing Effective Predictive Modeling
If you are someone who enjoys cooking, then you are likely to have several automated tools that make your time spent in the kitchen more enjoyable. But the quality and taste of your gourmet creation are not dependent solely on what type of food processor you use.
Several key factors affect the final outcome. First, you need to plan your meal. Next, you need to decide on and gather the ingredients. The quality of those ingredients is crucial to the success of your culinary masterpiece. Finally, you must perform the taste test. While it is probably the most enjoyable part of the project, it is still critical to evaluating whether your efforts are truly a success.
So why talk about cooking in a marketing newspaper? Because the parallels are uncanny. The next time you attend a marketing conference, notice how the latest automated modeling tools are promoted in the exhibition hall. You will come in contact with vendors touting "ease of use" or "no analytic skills necessary." Don't be fooled. While these tools are excellent for performing many important and time-saving functions, if not used properly the results can be disastrous.
Those who have been working in this field for many years know the pitfalls inherent in these claims. They know that the success of any data modeling project requires not only a good understanding of the methods, but also solid knowledge of the data, market and overall business objectives. In fact, in relationship to the entire process, the model processing is only a small piece.
The entire model building process is the focus of my new book, "Data Mining Cookbook: Data Models for Marketing, Risk, and Customer Relationship Management."
I describe each step with case studies. I begin the discussion with the first basic step, planning the menu. This involves determining the goal or clearly defining the objective from a business perspective. For example, to make clear what you want to predict, you might consider asking these questions, which can be addressed through the use of profiling, segmentation and/or target modeling:
• Do you want to attract new customers? Targeted response modeling on new customer acquisition campaigns will bring in more customers for the same marketing cost.
• Do you want those new customers to be profitable? Lifetime value modeling will identify prospects with a high likelihood of being profitable customers in the long term.
• Do you want to avoid high-risk customers? Risk or approval models will identify customers or prospects who have a high likelihood of creating a loss for the company. In financial services, a typical loss comes from nonpayment on a loan. Insurance losses result from claims filed by the insured.
• Do you want to understand the characteristics of your current customers? This involves segmenting the customer base through profile analysis. It is a valuable exercise for many reasons. It allows you to see the characteristics of your most profitable customers. Once the segments are defined, you can match those characteristics to members of outside lists and build targeting models to attract more profitable customers. Another benefit of segmenting the most profitable and least profitable customers is to offer varying levels of customer service.
• Do you want to make your unprofitable customers more profitable? Cross-sell and upsell targeting models can be used to increase profits from current customers.
• Do you want to retain your profitable customers? Retention or churn models identify customers with a high likelihood of lowering or ceasing their current level of activity. By identifying these customers before they leave, you can take action to retain them. It is often less expensive to retain them than it is to win them back.
• Do you want to win back your lost customers? Win-back models are built to target former customers. They can target response or lifetime value depending on the objective.
• Do you want to improve customer satisfaction? In today's competitive market, customer satisfaction is key to success. Combining market research with customer profiling is an effective method of measuring customer satisfaction.
• Do you want to increase sales? Increased sales can be accomplished in several ways. A new customer acquisition model will grow the customer base, leading to increased sales. Cross-sell and upsell models can also be used to increase sales.
• Do you want to reduce expenses? Better targeting through the use of models for new customer acquisition and customer relationship management will reduce expenses by improving the efficiency of your marketing efforts.
Once you have established your objective, the next step is gathering the ingredients. The best ingredients for a powerful model are clean, actionable, accessible, relevant data. Depending on the type of model, the data may be purchased, created inhouse or a combination of both. I cannot overemphasize the importance of these first two steps. The level of their accuracy and relevancy directly influences the entire modeling process.
Now it's time to reach for the tools. And there are many from which to choose. Neural nets, classification and regression trees, and genetic algorithms all promise great results with their powerful methods. But I prefer good, old logistic regression. If you take the time to clearly define your objective, obtain quality data and carefully transform the predictive variables, your model will be not only powerful but also very robust.
Once the model is complete, validation is essential. While there are numerous statistics available for evaluation, I prefer to look at the effect of the model on the bottom line. This involves performing decile analysis that feed into gains tables and gains charts. These common sense validation techniques provide a wealth of actionable information.
We're all busy people. Speed is becoming the No. 1 competitive advantage. Modeling tools, like food processors, can help us reach our goals more quickly. Just be sure to consider the entire process so you can ensure perfect tasting models every time.
• C. Olivia Parr Rud is executive vice president at Data Square, West Chester, PA, a database marketing consulting firm offering business intelligence solutions.