Modern marketers often frame their role as developing customer-centric marketing strategies to maximize some metric of importance, such as response rates, ROI, or cost per acquisition (CPA). Predictive models are analytical tools that provide us a formal way for maximizing (or minimizing) those metrics. This post is Part III in a series exploring how predictive models can be used in new customer acquisition. You can read Part I here, and Part II here.
The first step in building a predictive model might actually be the most important. Working to understand the goals and context around past marketing efforts can help inform the new model build and possibly uncover potential problem areas. Asking the right questions helps put together an analysis plan.
For example, some helpful questions to ask are:
What acquisition marketing has been done in the past?
Predictive models are built on historical customer data, and they are particularly good at finding which variables are important for differentiating people in that data. This can be both good and bad. If marketing in the past was targeted at a particular group – perhaps from a purchased list of people – the new model will learn simply about those people and be less applicable to a broader population. We call this problem selection bias. If that set of people was working well in the past and we are trying to replicate that success in a larger population, we may be fine. However, if we are looking to target a potentially different audience, we will have to change strategies. For example, we might recommend planning test campaigns to a broader audience so we can build a data set better suited for training a model.
What is the Lifetime Value of a new customer?
Having a rough idea of the value of a customer is a good benchmark for planning the profitability of a campaign. A common question for every acquisition marketer is: “How many people should I target?” We can answer this question by estimating response rates from the model and calculating both cost and expected revenue to estimate how much we should target and remain profitable. The marketer can also make more informed decisions about what methods of targeting are available. Without knowing customer value, all of this is little more than a guess.
What do we already know about current customers?
This is where our clients’ domain knowledge is invaluable; they are, after all, the experts on their own customers. Marketing databases have thousands of possible features that can be used for modeling and analysis. Any initial understanding can help us determine where to focus or if we need to source new attributes that might be missing. A common example is when we learn distance to store locations (or competitor locations) is an important factor. We know an important preparation step in that case is to bring in store locations to calculate a person’s distance to the nearest store.
Now that we understand the problem and have an analysis plan, we can begin creating a data set for building the model.