The business provides loans to SME’s on a short-term basis (6 months) as a working capital solution. They used a credit risk score to do underwriting decisions prior to Zinia. The loan is repaid on a daily basis, thereby the repayment behaviour is very noisy and difficult to forecast. The credit score used by the business was not predictive enough to capture the volatile repayment behaviour. Thereby, the underwriting decision was not optimised to capture the net (of bad debt) lending opportunity available.
First, Zinia automatically cleans and standardizes the data fields and also selects the data variables for the AI model. In this case, Zinia selected the following types of data variables (about 8 final model variables).
|Product features||Other Variables||Loan performance|
|Loan term in days||Deposits||Default/ Paid in full/Early settlement|
|Loan amount||Daily cash|
|Internal credit score|
Then, Zinia builds a set of supervised models. It builds deep neural networks for customer behaviors such as price sensitivity to take up the loan, customer default and how much they would pay back. Then Zinia combines these models to automatically produce a daily repayment forecast.
Zinia also builds an alternative KNN based forecast (semi supervised approach) and finds that the best final daily forecast was produced by blending supervised and semi supervised approaches (Neural networks & KNN)
Zinia forecast achieved 97% accuracy against actuals, which translated into an improvement of about 20% net present value for the balances lent (in comparison to their previous modelling)