Recently there’s been plenty of focus on predictive analytics. We were recently privileged to create a Take 10 webcast on this subject for MarketingProfs . Why all the interest? Companies want to be able to apply a variety of statistical techniques from modeling, machine learning, data mining and game theory. This way they can uncover relationships and patterns in order to predict behavior and events, such as attrition, propensity to purchase, incremental lift to maximize impact and optimize marketing mix and spend.
These models assign scores or ranks to each customer based on probabilities in order to predict a single behavior, such as which customers are most likely to buy a specific product(propensity to purchase modeling), which customers are most likely to be influenced by a specific promotion (response modeling), or to calculate customer lifetime value.
As with any model development, you will need to perform the usual data cleansing, transformation, initial and ongoing validation and refinement. These steps will help you begin creating a propensity model.
First, you need a suitable modeling sample. This requires enough records (thousands) that are recent enough to be relevant so that you can simulate various scenarios and perform the appropriate analyses. Odds are you will be using a variety of internal data sources, such as transaction, contact, weblog, text, and campaign data as well as appending external data to improve the quality of your model. The more instances of what it is you are trying to predict the more robust a model you can create.
Second, once you have your sample, check it carefully for biases.
Third, establish your criteria and ranks based on weighted attributes and build the model.
Fourth, similar to testing a new pharmaceutical, test your model with both a treatment and a control group. It will be essential to have clean control groups so that comparisons are truly actionable. This allows you to find the buyers, responders, etc. in both groups and also ascertain what kind of people did not perform the desired behavior in the control group but did so in the treatment group. These are the customers whose behavior was impacted only because of the treatment. You can now build a propensity model.
In parting, once you create and implement the model it will be important to communicate the results and the value generated as a result of the model.