End to End Customer Journey Solutions Utilizing Differentiated Data Assets
This blog delves deeper into a presentation delivered with Equifax at the Telecommunications Risk Management Association (TRMA) Spring conference in Tucson, AZ.
Having served in a consulting capacity to the telecommunications, pay tv, and cable industry for the last eight years, I’ve been fortunate to see a great deal of cross-functional areas within these businesses, including supply chain, engineering, tech ops, finance, accounting, and credit and collections. Inevitably, each of these areas is looking for an edge that will differentiate their function from its peers and identify ways to improve their performance in a meaningful way. Any change made in these various areas will impact the customer’s journey, which is ultimately how we arrived at the title of our presentation: End to End Customer Journey Solutions Utilizing Differentiated Data Assets.
With the explosion of focus on customer experience in this industry, the mapping of the customer journey is a common occurrence. For example: how does each choice an organization makes, starting from the buy-flow through to when and how to contact them once they’ve entered into a state of delinquency impact the customer’s journey?
Recession Looming
How does a looming recession change how we as an organization will handle the customer journey? Nobel-winning economist Paul Krugman was quoted in February at the World Government Summit as saying that “there was a good chance that the U.S. will have a recession late this year.”
According to a report released by the New York Fed earlier this year, $584 billion in new auto loans and leases appeared on credit reports in 2018 — the highest level seen in 19 years. With the high volume of prime originations, there has been a shift in the quality of the outstanding pool of auto loans, and as of the 4th quarter of 2018, 30% of the $1.27 trillion in outstanding debt was originated to borrowers with a credit score over 760.
Particularly interesting from the report was a note that flow into serious delinquency (90+ days delinquent) in the 4th quarter of 2018 crept up to 2.4%, substantially above the low of 1.5% in 2012.
Industry Trends
Piggybacking off the trends seen on a more macro level, we begin to dive deeper into how these trends bare out in the telecommunications industry as we see similar growth in the late stage aging buckets.
We also see a significant impact from potentially fraudulent customers identified as “Never Pay” customers and the overall increase in balances at the time customers are written off. We conclude that the balance increases are primarily attributed to the overall increase in cost of service, but are also influenced by the increasing value of unreturned equipment and treatment strategies which don’t modify the behavior of the consumer allowing them to accumulate large delinquent balances.
TACKLING THE PROBLEM
As we previously stated, ultimately each of these departments is looking for their proverbial leg up, and the credit and collections teams are no different. They spend millions of dollars annually attempting to mitigate risk during customer acquisition through risk modeling and fraud prevention solutions, but it seems that all too often the next steps in the customer journey are overlooked.
Our premise, and what we try to enact in our customers, is the idea of predicting consumer behavior early and often in the customer lifecycle. This approach avoids any potential biases that a consumer might create and truly relies on the predictive nature of the model.
Our models are typically focused in two areas:
- Propensity to Pay — Identifies a customer’s likelihood of paying. This prediction will evolve over time.
- Never Pay — Identifies the probability that this customer will never pay us.
***Quick caveat: A never pay, for the purposes of this discussion, is focused on customers who never made a payment following the activation of service. With that said, if a customer made a deposit to obtain service but made no further payments, they are identified as a Never Pay.
The interesting thing about these two models is they dissect a unique set of subscribers, and while a customer may have a low propensity to pay score, it doesn’t necessarily mean that their never pay score won’t be high. Obviously, in cases where both of those statements are true, we encourage our customers to act accordingly during the customer journey mapping process.
Mechanics
As with any machine learning model, there are three core steps in the process: Train, Test, Operationalize. Then repeat.
Other critical parts of the training process include the munging and feature engineering aspects of model development. It is crucial during these steps that the business, in this case, the credit and collections team, work closely with the data science team building the model. Close collaboration will ensure that the right attributes and features are considered for the model. It has been my experience that you and your team undoubtedly understand the data better than the data warehousing and data scientists who are working with the data.
Differentiated Data
There are five key differentiated data sets we considered when putting together the presentation, but this list is by no means intended to be exhaustive. If you believe there are other relevant pieces of information outside of those listed here, we encourage you to work with your data warehousing teams to effectively ingest and stage those datasets for use in the model.
Biller Demographics
It probably comes as no surprise, but as a provider, you have a myriad of information about your consumer population and their behaviors within the historical data in your system. Some of this may be challenging to absorb or consume due to the high transactional volume nature of the system or the fact that your consumer population is split between multiple systems, but with some diligence, you can effectively capture the demographic of your consumer footprint.
These demographics will be biller sourced values aggregated at a zip code level. Generally speaking, these provide valuable insight into how your customers have performed historically, and when used in prediction, will determine the historical persona a particular individual most resembles.
An example of metrics included in these demographics might be the NSF percentage which would detail the insufficient funds transactions as a percentage of total payments by Zip. Additionally, you might track values like the write-off percentage, the average length of service, and the charge back percentage.
Biller Attributes
While the biller demographics are focused at an aggregate level, the biller attributes will be specific attributes about the newly onboarded consumer that we’ll want to capture and monitor over their customer journey. Though they may change over time, these attributes will give us and the model insight into behavioral changes within a consumer that may suggest an issue looming.
You may find that these attributes are not readily stored at a subscriber level in your billing system, and it may be necessary to complete the data munging process mentioned earlier to capture these values.
An example of a biller attribute might be an e-Bill Indicator, which would suggest the customer is set up to receive electronic statements. This attribute alone might not be a strong predictor, but that this couple with an invalid email address on their customer record would identify a common scenario for delinquency in the cable and pay-tv space.
Account Linking Demographic
Account linking demographics utilize the billing information to identify relationships in your existing customer population. These do not necessarily match on common values like Social Security Number, but they would be matches on phone numbers or email address.
Once a relationship is identified it is essential to understand and record similar demographic data captured in the billing demographics, but relatively only to these linked accounts. This data would include things like the number of inactive accounts, write-off amounts, and the number of unreturned equipment pieces.
System Disposition
This data set will be sourced from the AR management or Collections management solution. Agents provide this data either in internal (early stage) collections or by third-party agencies (late stage) write-off collections. These dispositions capture the sentiment of the customer and allow us to characterize the interactions.
OPERATIONALIZE
To operationalize these types of models and effectively alter your current strategy, there should be several considerations about the AR management or Collections management system you are utilizing.
The system should have the capability to complete granular segmentation of your account population using any underlying attribute or system generated attributes (e.g., the scores of the models defined in this post).
Additionally, your system should be capable of establishing event triggers that will consume or obtain new data as it becomes available. The system should also capture dispositions or outcomes from both internal agents and outsourced vendors. Lastly, the system should be capable of dynamic rescoring and re-prioritization of accounts.
Want to learn more about the collections industry? Comment below a few topics you’d be interested in reading next!
