Using Analytic Profiles To Improve Customer Retention
In my last blog, Best Practices for Analytic Profiles, I talked about the importance of creating individual profiles for key entities (e.g., customers, devices, machines, cars, wind turbines) as part of the data science process. Now, I’d like to walk through an example of how we can apply these profiles to improve customer retention by identifying and scoring “unusual” customer activities and behaviors (vis-à-vis their individual profiles).
For this customer retention example, we compare a customer’s current activities (purchases, claims, returns, social posts) and current state (location, day of week, time of day, weather at location, holidays, seasonality, events at location, traffic at location, nearby coffee shops, nearby restaurants) to their profiles in order to determine the next best offer or flag anomalies worthy of investigation and/or action.
The analysis process for the “Improve Customer Retention” business initiative is laid out below in Figure 1:
- Step 1: Establish a hypothesis that you want to test. In our customer retention example, our test hypothesis is that “a customer issue that goes unresolved for 5 days or more has a significant impact on customer attrition.”
- Step 2: Identify and quantify the most important metrics or scores (combinations of multiple metrics) to help predict a certain business outcome. In our example, the metrics and scores that we’re going to use to test our customer attrition hypothesis includes Customer Tenure (in months), Customer Satisfaction Score, Average Monthly Purchases, and Customer Loyalty Score. Notice that the metrics don’t all have the same weight (or confidence level). Some metrics and scores are more important than others in predicting performance given the test hypothesis.
- Step 3: Employ the predictive metrics to build detailed profiles for each individual customer with respect to the customer retention outcome test hypothesis.
- Step 4: Compare an individual’s recent activities and current state with their profile in order to flag or score unusual behaviors and actions that may be indicative of a customer retention problem. In our customer retention example, we might want to create a “Customer Attrition” score that quantifies the likelihood that particular customer is going to leave, and then create specific recommendations as to what actions or offers can be made to retain that customer.
- Step 5: Continue to seek out new data sources and new metrics (lead indicators) that may be better predictors of attrition. This is also the part of the process to try to improve the confidence levels of the metrics and scores using sensitivity analysis and simulations like the Monte Carlo experiments.
- Step 6: Integrate the analytic insights, scores, and recommendations that came out of Step 4 into the key operational systems (likely CRM, marketing, service and call center for the customer attrition business initiative) in order to ensure that the insight uncovered by the analysis are actionable by front-line employees.
Profiles: The Organization’s Key Intellectual Property
Here is what we’ve learned by employing this analytic approach built around detailed profiles:
- In a Big Data world, we can now build individual profiles at an unprecedented level of detail. For a personalized marketing initiative, we built 120 million profiles (individual customer by major product categories). For a product quality initiative for a high-tech manufacturer, we’ll be building 332 million profiles (contract manufacturer by product ID by test stage) against which we’ll be monitoring, flagging, and scoring potential supplier quality, line stop and on-time shipment problems.
- As we move from use case to use case, we are constantly enhancing the profiles of our key entities. With each use case, we will uncover new metrics and scores that can be added to the customer or machine profiles to improve the predictability for new use cases. We’re finding that some of these new metrics can be used to improve the predictive accuracy and confidence levels of previous use cases (see Figure 2).
These customer, partner, product profiles end up becoming key intellectual property that evolves from use case to use case, and the predictive power of these profiles will grow as these profiles are deployed across additional use cases.
Finally, I want to send out a special thanks to Team Godzilla (Srini, Wei, and Shriya) for all their hard work on making this approach work for our current client Vision Workshop engagement!
Here they are setting up a data science lab in a local Starbucks over the Memorial Day weekend. Nothing like a little Starbucks mixed with some data science to nurture some creative thinking!