Boosting credit risk assessments in banking with AI

The assessment of credit risk demands a great deal of manual effort. A discussion with Simone Mensink (Director of Banking at IG&H) and Hein Wegdam (Head of ING Real Estate) on the credit assessment of the future, which combines expert knowledge with artificial intelligence.

At banks, the asset-based finance department often deals with credit checks high in complexity. The department deals with both low-risk and high-risk customer profiles that require a lot of manual handling.

In recent years, several financial institutions started using ‘rule-based’ decision trees to process credit checks more efficiently. These trees are automated for the most part and are run to determine whether a credit request will be granted.

But, defining the details per rule and upkeep of the system are very labor-intensive. Wegdam: “The rule-based model became too complex over time. The many rules made maintenance difficult, and subtle patterns were not recognized, ultimately making the results less accurate.”

While machine learning may promise some relief, an artificial intelligence (AI) model based on historical data does not perform well when data is limited or of poor quality. In addition, these types of models are slow to adapt to new circumstances or policies because they can only learn from historical results.

Customized AI thanks to expert input

“When giving one credit application to two different experts, we see that they can come to different conclusions. By capturing the knowledge of multiple experts in an AI model, credit decisioning becomes more efficient and consistent,” explained Mensink. This model is designed to work objectively, reducing the likelihood of human error and bias.

Nonetheless, guidance from people remains crucial. The experts identify the relevant variables, create a training set and provide approximately 500 representative examples with an objective risk scoring. This reduces dependence on historical data and makes the model more flexible to adapt to changing (market) conditions.

Pointing at a practical example, Wegdam said: “Together with the data science team of IG&H we developed unique decision models for the real estate financing market for loan reviews, extensions and applications. 80% of reviews and 50% of loan extensions were automated. We create added value by using our real estate financing knowledge where specific expertise is needed, like risk exceptions.”

Checks and balances pave the road to success

The most important success factor for an organization involves overcoming objections and prejudices. Wegdam: “We believe we can service regulations better and faster using these models. They should therefore be well thought out, tested and monitored. Dashboard reports improve the management team’s understanding of processes and model performance. The model also notifies the user when it is unable to provide an appropriate answer and leaves the manual evaluation to the user.”

There are several ways in which the model is checked and stays current. For each case, it is explained why the model arrived at this outcome. It shows the three most important variables that contributed to the outcome in question and can thereby be checked by a human expert. The risk department also builds in checkpoints to test and retrain the model if, for example, a changing market circumstance requires it.

Cooperation between the Front Office and Risk Management departments is crucial. By taking ownership of the model, the Risk Management department can play a leading role in adoption within the organization. “The company’s data scientists themselves will need to be involved to work with the model, but also to be able to check and adjust it,” explained Mensink.

“At ING we did so by working together on the design and content of the model so that the process provides transparency and grows trust. If employees are comfortable and open to experiencing this way of working for themselves, success will follow.”