Manual underwriting meets ML automation

The problem was that a lot of new loan applications were being referred to an underwriter team for manual review and decision. We needed to solve the issue at its source.

It all started with a question

We set out to develop a solution which would enhance the acquisition process for everyone — improving customer experience, reducing costs, and obtaining more consistent decisions. The solution would also need to be implemented in line with current regulations and compliance requirements.

But before we got started, we asked ourselves… could machine learning (ML) hold the answer? Could it be the secret to boosting automation and achieving better compliance

Phase 1 — Find the inefficiencies

Our first step was to optimise the current decision strategy by combining traditional analytical techniques and Natural Language Processing. This helped us identify potential process efficiencies and consistent underwriter behaviour — resulting in a 30% reduction in underwriting effort.

We also wanted to identify which decisions were too complex or inconsistent. To do so, we created a comprehensive and transparent database from which the target population for the data science exercise could be identified from the remaining referrals.

Phase 2 — Find the right model

The second phase was a race between several ML algorithms, with the winner being the model structure that best described the relationships in the data to predict the target (Accept/Reject). 

A key advantage of using ML techniques over more traditional methods was the ability to handle large volumes of data effectively. Our solution could therefore outperform existing policy rules in the identification of desired customers.

A bespoke tool we could count on

Following our initial tests, we agreed with our client on which model to take forward according to predetermined criteria. We then used our bespoke tool to facilitate stakeholder review and sign-off of the final ML model and strategy.

Our flexible R- and Python-based tool allowed the user to explore a model and identify relationships in the data. Regardless of complexity, users could understand the model output at both aggregate and individual case levels. By focusing on outcome interpretation, our tool made it easier to understand how certain profiles influence the modelling target, making it a comprehensive model monitoring solution too. 

One of its most unique properties was its ability to explore individual cases and identify the reasons underlying model decisions and recommendations. This enhances transparency and understanding, while supporting compliance in heavily regulated industries. That means customers receive a fairer outcome too, thanks to a higher level of consistency than human decisions.

More benefits than one

Overall, we demonstrated that underwriting effort could be reduced by up to 80%. As well as this, we delivered a streamlined and efficient credit risk underwriting process, resulting in:

• Improved customer experience through faster application processing.
• More informed decisions due to the ML model’s ability to utilise more extensive data.
• Consistent decisions that were resilient to human biases and oversights.
• Simplified decision processes, reducing underwriter effort in processing applications that required manual checks.
• A flexible approach that can easily be adapted to changing scenarios.

Underwriting is a time-consuming and costly part of the application decision process. Automating this has been a huge advantage to our client and its customers alike.

Can 4most support your organisation in a similar way? Let us know how we can help you by completing our short contact form.