Credit scoring has been around for over 25 years and millions of lending decisions are being made in the UK every year with the process being that customers are reviewed based on a credit rating that takes into account their past history to predict how likely they are to be able to pay back that loan. These rules are then applied to everyone in the same way, which means that people will be denied when they are perfectly capable of paying that back.

To see if Machine Learning can beat credit score at predicting if a customer is likely to default on a loan or credit card.

The Kortical models were built off the customer account information that had been anonymised and prepared for the FCA.

They gave the full transaction history for a small portion of the bank’s 20 million customers to train the models on with indicators of when default events occurred.

Getting data from large banks is notoriously challenging, we had to beat the top score on a public data-science competition for credit scoring, with 924 teams that entered, to prove we had a market leading solution and that it was worth investing in getting us the data.

Being transaction level data with CRM data, there were over 220 million rows of data. The platform does a lot of the heavy lifting in terms of cleaning and data transformation so the data scientists were able to iterate quickly on highly bespoke feature engineering to improve the model.

The results were much better than anyone anticipated, catching 83% of bad debt not caught by credit score. Which drastically reduces the bank’s exposure to credit risk. Should they want to keep the same rate of default, they could lend to 77% more people.

Using the explainability functionality uncovered really interesting insights, notably the customer segments behaved radically different and had contrasting drivers, something that a linear model would not be able to pick up on:

• The key indicator of a young person going into financial difficulty, was going into their unarranged overdraft (typical to what they had seen in the past)
• For an older person (45-65) going into their unarranged overdraft was not an indication, but spending between midnight and 6am was a large driver
• Late night spending for young people had no impact and in fact the top 15 drivers were different across the 2 groups

Being able to find different customer segments automatically and their drivers from the data is the reason why machine learning is so much more effective than linear models and rules-based decision engines.

There is a Swedish bank that is applying AI to credit decisions with great success. This technology is enabling them to answer credit applications much quicker than a human reviewing it. They have proved it is more accurate also, which means they are able to capture more of the lucrative lending market and be a stronger business.

If any of these points interest you, do get in touch with us via the form below, we have so much we can share around AI and finance.