Today's direct marketing companies continue to grow and expand, and they are moving from traditional recency, frequency, monetary value segmentation to more sophisticated segmentation using model scores.
In today's marketplace, where relationship marketing is key to success, customer file modeling is a cost-effective, powerful segmentation tool. However, there are challenges associated with implementing new and improved segmentation based on statistical modeling techniques.
RFM: traditional segmentation criteria. Typically, marketers use RFM data from the customer file to segment their files into categories. It is often estimated that RFM segmentation provides 75 percent to 85 percent of segmentation capabilities. As a marketer refines the RFM segmentation, he may add a product select on older segments to lift the performance of those cells. For example, 49-plus-month multibuyers with shoe purchases may be more likely to purchase from a footwear offer than 49-plus-month multibuyers with no shoe purchases. Similarly, a gender select on top of RFM can help boost a cell's performance. As the number of RFM cells increases and the matrix expands, increased lift in performance is gained, but the cells become unmanageable. This is one indicator that a customer file model may be appropriate.
Customer file modeling. The objective of a customer file model is to increase profits and improve name flow. Additional benefits include reducing the marketer's time and maximizing the information on the marketing database. Additional data, such as product purchase activity, promotion history, returns and demographics, can be incorporated into models with seamless implementation.
Model development packages and methods vary. Common statistical techniques include logistic regression, ordinary least squares regression, CHAID and neural networks, yet most statistical methods result in a scoring formula designed to predict sales, response, profits or returns. The marketer relies on the predictive score to forecast contribution levels by segment and determine mailing depth. Given that the technique is appropriate for the data, modelers generally agree that most modeling techniques will result in similar performance as long as the model developer has a good understanding of the data and its application.
Challenges and solutions. Challenges that a marketer faces when transitioning from the use of RFM segmentation to model scores include dealing with a “black box” syndrome, forecasting model segment performance and understanding the effects on name flow. Once the model is built, testing, benchmarking and post-analysis are tools that will help raise the comfort level. Common techniques are:
· Use validation results to determine the strength of the model. Mailing performance from a specific mailing or season is generally used to develop a model. The validation of the model should confirm that the model variables are significant on an entirely different mailing sample. The validation should illustrate the model's segmentation power and stability.
· Post-analyze mailings not included in the development of the model. By selecting one or more mailings to post-analyze the model segmentation against, the marketer will understand whether the same model that was developed on a particular mailing or season will apply across other mailings or time frames. The model may or may not be applicable to all mailing drops across all seasons, depending on the difference in the offer and seasonal influences.
· Set up an in-the-mail test to measure cost versus benefit of using the model. This will help the mailer understand the gains needed to offset the costs involved with model development and application.
Testing. Here is one method of testing the model. When evaluating names selected using RFM versus names selected using model segmentation, the following distribution will occur: Each technique will select names that are in common with the other technique. This is generally in the 70 percent to 75 percent range. The model will identify names selected by RFM that the model predicts are unprofitable. The model also will identify names not selected by RFM that the model predicts are profitable. Each of these two groups must be tested to understand the value and efficiency of the customer model.
Benchmarking. Another recommendation is to establish a benchmark to understand how the model is performing. Here is how a benchmark works:
· Identify a segment of names that is predictable and stable – for example, 0- to 24-month multibuyers with a monetary value of $50-plus.
· According to history, develop sales-per-mailing levels of how the benchmark performs in various seasons.
· From the development mailing, identify how each model segment performs compared with the benchmark.
· When using the model, always select the benchmark as a segment to be mailed in addition to the model segments.
· Forecast how the benchmark and all model segments selected will perform in the next mailing using the indices created in the third point.
· After four to eight weeks, read results and identify whether the model is performing against the benchmark in the same relationship as when developed. If not, determine whether the model is still performing at an acceptable level.
Results. A customer file model often achieves significant improvements over traditional RFM segmentation. For the typical mailer, annualized benefits have been as high as $250,000 to $1 million.
In summary, a model can significantly improve the profits of each customer mailing and can improve name flow by reactivating names not normally mailed using RFM. It also can simplify a marketer's efforts.
