How to Incorporate RFM Segmentation With Predictive Models
With predictive models all the rage in direct marketing analytics these days, recency-frequency-monetary value segmentation may seem outdated. But the truth of the matter is, RFM still has a place in modern database marketing when used side by side with predictive models. After all, the best predictor for future behavior is past behavior, and that’s exactly what RFM offers.
RFM should be used in concert with predictive models to optimize database marketing plans. Generally, predictive models do a superior job of predicting sales and have better ROI than segmentations based on RFM. But RFM is still useful to help validate predictive models before you contact customers.
For example, when your database firm brings you attractive statistics, you should verify those statistics before adjusting your marketing plan. Review the sample sizes, statistical and sampling techniques, treatment of outliers, treatment of incomplete information (blanks, nulls, garbage), transformation of variables, etc. Then, you should ask for a cross-tab comparing its score versus RFM data.
The cross-tab should contain not only population counts, but also ratios such as life-to-date (LTD) dollars/buyers, average order size, average days since last purchase and LTD orders/buyer. Look for a correlation between the best scores and the best ratios. If there are attractive ratios in the basement of your predictive model, you may be missing out on opportunities with good customers.
While RFM provides a broader view of segments than predictive models, it also has some advantages over them. For instance, you can build it yourself and do so quickly; it’s portable across industries; can quantify seasonality issues to a degree; applies to all customers in your database; and can be used for reactivation, cross-sell, renewal and acquisition campaigns.
Certainly, RFM is not a replacement for inferential statistics from predictive models. Rather, it should add value in its unique ability to amplify management understanding, ensure data quality and substitute for predictive models when models are not practical.