Latent Representations of Loan Delinquencies from Neural Networks for Decisioning Customers and Assessing Phone Call Strategies

During the previous Mitacs PhD internship (IT11500), a long short-term memory encoder-decoder (LSTM-ED) neural network survival model was developed to predict time to debt repayment for ATB collections customers. This model is novel from both methodological and practical perspectives. The LSTM-ED simultaneously predicts the probability that debt will be repaid within the first 90 days of delinquency and when in the future debt will be repaid (given that repayment occurs within the first 90 days), which is not present in the academic literature. This internship will focus on novel applications of latent representations learned by the LSTM-ED.
In the first subproject, we will use the latent representations of delinquencies to develop thresholds for determining whether debt will be repaid within the first 90 days. Delinquency-specific thresholds will be generated by passing the latent representations as input to a feedforward neural network (FNN) and applied to the predicted probabilities ???? ??(???? = 90) output by the LSTM-ED. The extended version of the LSTM-ED will automatically decision collections customers into two classes (does not repay within 90 days and repays within 90 days). This will make it easier to deploy and integrate the model into ATB’s daily collections process. The second subproject

Faculty Supervisor:

Bei Jiang

Student:

Partner:

ATB Financial

Discipline:

Mathematics

Sector:

Finance and Insurance

University:

University of Alberta

Program: