Early Warning Delinquency Prediction

To develop and maintain predictive solutions that scores all live accounts (meeting certain criteria*) in the portfolio every month on their propensity to get disconnected or bankrupt in the next six months time. This information is then used by the collections function for optimal collection efforts.

*Accounts that are in collections (missed payment of at least one bill) and have due amount greater than certain amount suggested by the client.

Background and Challenges

Our client, a leading telecom company of the US, was experiencing an unusual growth in their recoveries portfolio. The number of customer getting disconnected or bankrupt monthly was growing every month. The client wanted to preemptively identify customers (based on predictive scores) who were most likely to discontinue their service within next six month at the time of coming into collections for the first time. This insight could then be used to segment the customers and use differential strategies for collections as well retention based on the model scores.

Our Approach

Data warehouses that the team got access to primarily consisted of- month on month billing data, payment data and product related information. Based on client inputs, the ‘target’ customers (customers discontinuing their services in next six month) were tagged in the data. It was observed that different customer segments were exhibiting different set of payment behaviors, meaning there were clusters with distinct behavioral trends. Clustering algorithms were run to identify distinct sub-populations in the portfolio. All sub-populations needed separate predictive models.

For each sub-population, various statistical techniques like uni-variate/bi-variate analysis, selection by regression etc. were used to identify important characteristics that govern a customer’s payment behavior. The target being dichotomous in nature, logistic regression was identified as the best suited modeling approach.
End output of these models under development was a score for each of the customers. Higher the score, the greater is the likelihood of a customer getting disconnected or a service termination in next six months time.

Project Process Map

Phase 1: Predictive Characteristic Analysis
An analysis showing the predictive power of various Strata extract attributes and derived attributes available at the observation point. It includes rank-orders account predictors for the “bad” definition (objective function).

Phase 2: Sub-Population Analysis
Assess the improvement to overall predictive strength of developing models for other sub-populationsAssess the improvement to overall predictive strength of developing models for other Verizon sub-populations

•  Includes creation of prototype models for selected sub-populations
•  Includes performance charts and K-S statistics of prototype models

Phase 3: Final Behavior Model Specifications
Final documentation of model(s) showing predictive strength and specifications for implementation
Includes performance charts and K-S statistics of final model

• Documentation of final model attributes and point allocation or coefficients
•  Ready for input into database engine

Phase 4: Model Maintenance and Fine tuning
We study the rank ordering charts and model results every month and in discussions with the client, make necessary changes to various models for different sub populations on a regular basis. The team dedicatedly works with the client management on a full time basis and carry out various analyses to support them use the models most efficiently.

Model Results

Benefits

Using our approach our client now has a scientific approach for the collection strategies to optimize their collection efforts. This tool is being used in live collection system with an automated alert system to raise indications of customers in terms of high, medium or low risk level at time of their first default itself.