Credit decisioning: Finding Customers and Managing Risk

Client information / Background

The client is in the credit space, providing short-term small loans to customers. For example, a $100 loan to purchase clothes or general merchandise. Their market is the mid- to lower-class market in South Africa.

The Challenge

They aim to provide as many loans as possible with a specific risk management approach. The better they determine the risk of a future or existing customer the more informed they will be when considering whether, or not, they will provide the individual with a loan and what the appropriate loan amount should be.

Before we got involved they had one model combining two industry credit models' scores. They could only score "offline". A data scientist needed to physically run their model on a batch file and then send it via email. Modelling was used to assess the risk of new applicants.

To improve their value offering to their customers, they needed models readily available that could score in real-time, generate better credit scores and score for various points in the credit life cycle, from pre-scoring a prospective customer, to scoring them continuously on various risk bands.

The Solution

We developed and operationalised multiple models throughout the customer lifecycle stages and integrated the scores back into their operational customer management system.

There are 15 machine learning models in operation. Each model’s metrics are meticulously tracked and any deviation will require retraining of the existing model or a challenger model to be implemented.

For brevity, we will discuss the two extreme examples, pre-scoring leads for marketing and scoring a customer on the likelihood of non-payment.

Pre-scoring model

The client uses the pre-scoring model to improve sales and marketing efforts. Everyday thousands of leads are pre-scored and then if they qualify an offer is made to them through one of the digital channels. Another machine learning model recommends the channel and time of day the offer is communicated. In addition, the client uses a variety of strategies based on the type of lead and the lead's behaviour.

Likelihood of non-payment

All existing customers have scores associated with them, depending on the bucket of delinquency they fall into.

A machine learning model recommends the next appropriate action. That is how to deal with a delinquent or potentially delinquent customer. This could be just a gentle SMS reminder, multiple phone calls through the call centre or transferring this customer to the collections unit. 

All existing customers are ranked according to importance, based on their likelihood of non-payment and associated risk throughout their lifecycle. For example, a customer with a higher loan amount that has not paid for 4 months, but used to be a reliable payer would be listed higher than one with a lower loan amount that has not paid for a month. Those listed higher will receive more attention to recover the loan.

Spatialedge’s Role

Not only did we develop the additional models, all the models and data pipelines run on our core platform, available in the cloud. 

We provide a managed service to operationally support and maintain all of the machine learning models. We track all model metrics, such as the mean, median, kurtosis, skewness, standard deviation, performance, etc. When one of the metrics starts to deviate, then the team will assess and eventually develop an improved challenger model to maintain a healthy level of performance.

Copyright © Spatialedge, all rights reserved.

Head Office

Trumali House, Stellenbosch

UK Office

27 Old Gloucester Street, London