The word “Force multiplier,” in military usage, refers to an enabler or a combination of enablers, which make a given force more effective than that same force would be without it. Similarly, in a business context, the predictive power of customer behavior modeling is stronger with a combination of segmentation, scoring models, unstructured text mining and social network analysis. When used in combination, these techniques deliver greater value than when used in isolation. Using post-paid churn prediction modeling in the Telecom Industry as an example; this article explores the theme of combining multiple techniques to deliver a complete predictive model.
In the Telecom Industry, there is a constant risk of customers turning to a different carrier. In order to ensure that they minimize attrition, most operators have a recurring scoring model which scores their customers based on their likelihood to switch carriers. For example, the following indicates some of the important predictor variables which strongly influence the likelihood that an existing post-paid customer will churn.
As illustrated below, there are 4 levels at which this scoring model can be created:
Stage-1: Churn Scoring Using the Scoring Model
At Stage-1 of the maturity model for scoring customers, an organization can build a generic scoring model using logistic regression as an example. Once the model is statistically significant, customers can be scored and a targeted campaign can be run for the customers with the highest propensity to switch carriers, with the goal of preventing this from happening.
Stage-2: Segment-Based Scoring Model
At Stage-2 of the maturity model for scoring customers, an organization can ideally run a behavioral segmentation model to see if the influencers of churn are predicated on the behavior they exhibit. For example, customers are segmented based on their activity (call behavior ratio of night time calls vs. day time calls, SMS vs. Internet usage), payment behavior, and risk dimensions, and then a customized scoring model is applied for each behavioral segment.
Stage-3: Segment + Text Mining Based Scoring Model
At Stage-3 of the maturity model for scoring customers, organizations can ideally leverage unstructured information about a customer. To give an example, there are inbound calls received at call centers asking about features of a product, complaints about a billing error, etc. Also, there are outbound calls from the call centers to customers urging them to pay on time or use a complimentary product. These rich conversation transcripts are often not mined for insights. In this case, a text mining process when fed with these call center texts can synthesize keywords and themes. These can then be overlaid on the behavioral segments to enrich the understanding of the customer’s behavior. Also, a set of key churn watch-words can be deployed. For example, if the word “billing error” appears in more than 6 inbound conversations in the last 3 months, it can be a significant predictor for churn models.
Stage-4: Segment + Text Mining + Social Network Analysis Based Scoring Model
At Stage-4 of the maturity model for scoring customers, organizations can construct social network graphs for every customer if the data exists. For example, by looking at call patterns in the Telecom Industry, one can reverse engineer the social network of the post-paid subscriber and flag his network influence. The social network influence of a subscriber can be one more behavioral dimension that can be considered, while building the scoring model.
As demonstrated, by slowly stacking multiple techniques on top of one another, organizations in the telecom industry can incrementally enrich the churn scoring model. This same approach can be applied to analytical models in other industries including consumer packaged goods, travel and transportation, insurance and banking.
To conclude, here are a few key points: