In the first post about predictive analtyics we learned about the essential building block of predictive analytics: the predictor. This is a value calculated for each entity (say, a customer) who’s actions or behaviors are to be predicted – for instance the recency, in months, since a customer’s last purchase.

Prediction power is enhanced if you use more than one predictor at a time. In doing so you are creating a model. Models are the heart of predictive analytics. In this post I’ll discuss how you can find the “best” predictive model. I put “best” in quotes because from a practical standpoint, unless you assume unlimited time and resources you may be best off finding a model that improves your results (e.g., reduction in customer churn) over previous experience. Today there is available very powerful modeling software and well-trained and talented statisticians, but the number of variables to consider in any predictive model (across demographics, transactions, behaviors) can be extremely large making determination of the “best” model cost prohibitive.

Fortunately, taking an incremental, continuous improvement approach can yield solid results for most any business and the promise that results will improve over time. A common tool is to develop a yield curve. For example, plotting the results of a predictive model for churn with amount of churn on the Y axis and percentage of customers contacted in a retention campaign on the X axis will show a curve the decreases to a point — i.e., up to a certain percentage of a universe of customers contacted, attrition rates will fall — but will bottom out and then move upward. Meaning that not all customers will respond to a retention campaign and you are best off contacting only those predicted to respond well. After that point, you are best leaving the balance of the universe of customers alone – either because they are not likely to churn anyway or because the predictive models say campaigns to retain them will be unsuccessful (and possibly other methods are needed – along with models that might predict how these approaches can be equally tuned to expend effort on just those predicted to be successful).

Now, although the model does not work perfectly, the socring and ranking of customers according to their likelihood to be retained provides clear guidance on how to invest in retention programs to yield the best results. It will prevent campaigns to retain customers that are too aggressive (trying to retain those that are not likely to respond positively, or wasting effort on those that are likely to stay).

There is a great deal more to predictive analytics than I’ve covered in the past two posts. But I hope one message is clear: you can gain practical improvements in marketing results or other customer touch points through the use of analytics that don’t need to be complex (at least to start) nor perfect. Commitment, willingness to experiment and continuous improvement are what’s really required.

Thanks for reading and I’ll look forward to comments.