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 II in a series offering a deep dive into how predictive models are developed and can be used in new customer acquisition. You can read Part I here.
Marketing strategies aimed at existing customers – such as retention, customer engagement, and cross-selling – are all benefited by data stored in the business’s data warehouse so a marketer or analyst can understand how a customer interacts with the company. Descriptive data such as the frequency and variety of orders, visits to the website, and interactions with the call center are all likely stored and accessible. Brought together, these attributes tell a story about individual customers.
Acquisition marketers, however, have less immediate access to data. By definition, acquisition marketing requires targeting people that are not current customers, and therefore do not yet have stored data to act on. This is where companies like Speedeon Data can assist.
At Speedeon Data, we build and source analytics data sets that can fill the gap for acquisition marketers. Our consumer prospect database, for example, contains over 200 million U.S. individuals and 110 million U.S. households. It has hundreds of columns representing attributes (often called features or variables) that describe people, households, and locations. These attributes are used to select people most likely to be interested in our client’s products or services.
Consumer prospects are just one example. For business-to-business marketing, we maintain a business prospect database that is structured similarly and serves the same purpose. In other instances, marketers may have a product or service that only makes sense when the consumer is in market for that product or service. For that scenario, we source and build a life event database that contains the dates of significant events such as when consumers move, get married, add a baby to their household, and more.
Taken together, these analytics data sets form the raw materials we can use for acquisition marketing.
Building a Predictive Model
Now that we have the raw materials, what should we do with them? If we have a fair understanding of what our current customers look like, it might be tempting to use the available features in the analytics data set to hand select prospects appropriate for targeting. We call this method of prospect selection a “business rules model.” Business rules can be an appropriate strategy when the audience is small and well-defined, but in most cases, it will run into problems.
One common problem we see is that even after defining the parameters for selection, the remaining audience might still be millions of people. If we need to reduce the size of the audience further, what should we use? Should we use another arbitrary variable? Should we just draw a random sample? There are not a lot of good options here. But perhaps the biggest problem with business rule selections is that we may not have a great understanding of what customers look like to begin with. Selecting an audience using our instinct will likely yield non-optimal results.
The better alternative is to build a predictive model. Instead of making educated guesses about the features that matter, we apply a machine learning algorithm that can “learn” the relationship between the data and outcome we want to predict. When we use the term model, we are talking about this learned relationship.
Another way to think of a model is as a function that maps input attributes to an outcome. We call these input attributes features, and the outcome a target. The features describing a person might be age, income, location, home value, etc., and the target a yes or no indicating if the event – such as responding to direct mail – occurred. The role of the machine learning algorithm is to find the optimal function that differentiates between the two outcomes.
A nice feature of most marketing models is that they provide a probabilistic interpretation. In other words, they do not simply predict yes or no, but instead produce a score between 0 and 1 representing the likelihood the event will occur. These scores are what make acquisition models useful; we can score every prospect (who has the required features) and use it as a prioritization for marketing.
How do we actually build these models? The details can differ depending on the type and requirements, but in general, the process is surprisingly consistent. Our modeling process typically has the following steps:
- Requirements gathering
- Collect and create a modeling data set
- Preprocess data
- Train a model
- Score the prospect universe
In Part III of this series, we will begin to explore the five parts of building a predictive model.