Second Moment: Stats in Industry: Optimizing The Collection Strategy

home * about us * contact us * past features * columns * resource links * site map

Past Features

Columns

Resource Links

Delinquency rate is a key performance measure for any credit portfolio. Financing institutions devote a great deal of effort to collection activity to lower their delinquency numbers. The collection strategy traditionally employed by financing institutions is to assign collection efforts based on the number of days a borrower is delinquent, usually broken down into one of three categories, 1-30 days, 31-60 days, and >60 days. In other words, a borrower with an account that is 30 days delinquent is assumed to have a higher probability of making payment than a borrower whose account is 60 days delinquent. This type of collection strategy, however, ignores other key factors that may be even better indicators of risk.

In a paper titled “Optimization of Collection Efforts in Automobile Financing – A KDD Supported Environment” (published in the KDD-99, Proceedings of the Fifth ACM SIGKDD International Conference), H. Kauderer, G. Nakhaeizadeh et. al. discuss a data mining approach to optimizing the risk adjusted allocation of collection resources in a phone collections department. The basic idea is to use historical customer information to develop a predictive decisioning system. The assumption is that detailed information of customer behavior is a better predictor of risk than the simple delinquent days approach.

The first phase of model development considered the allocation issue as a classification problem. The days delinquent levels were replaced by three distinct risk classes defined as follows:

Level 1 – today’s delinquent customer has to stay under the 30-day delinquency threshold in each of the following three months.
Level 2 – today’s delinquent customer violates the 30-day threshold but stays below the 60-day threshold in each of the following three months
Level 3 – today’s delinquent customer violates the 60-day delinquency threshold in at least one of the following three months.

Using these predefined discrete target risk levels a classification algorithm (C5.0) was first applied to learn a classification model. The observed accuracy rate was promising, but also resulted in huge trees and rule sets that were not practical for implementation. Consequently, three alternative modeling techniques (linear regression, regression tree, and neural networks) were used and compared to the classification model. These techniques required the use of continuous-value target risk levels that were also derived based on linear and nonlinear transformation of risk associated with activities of the three following months.

In summary, the new models outperformed the traditional days delinquent model in the higher risk classes, while only the classification model outperformed the traditional model in the lowest risk class. Based on the results, the regression tree model was introduced into the phone collections workflow. The experience after 12 months of operation showed stable performance and a robustness that allowed for adjustments to score cutoff points to address fluctuating workload and capacity constraints.

More Stats in Industry