Transforming loan acceptance predictions using artificial intelligence
Demonstrating how an AI model closely mirrored the accuracy of a complex manual probability calculator, resulting in significant time savings and improved decision-making.
AI-Powered Probability Predictions
Implement an AI model capable of accurately predicting loan approval probabilities, leveraging deep learning and probabilistic regression to outperform traditional LMS predictions.
Synthetic Data Generation for Rapid Testing
Create a synthetic data generation tool to quickly produce plausible borrower enquiries, enabling fast and efficient proof of concept testing without the need for real lender data.
Enhanced Data-Driven Decision Making
Integrate a supervised learning algorithm into the existing LMS, allowing for continuous improvement and validation of loan predictions, ultimately enhancing decision-making accuracy and reducing manual effort.
Results
Even with our initial model, which was tested using synthetic rather than real data and applied to a limited number of enquiries, we successfully mirrored the complex manual probability calculator with remarkable accuracy. This achievement demonstrates the potential of our AI model to replicate or even improve upon existing methods.
By leveraging AI, we anticipate significant savings in time, as the model automates the probability calculation process that previously required extensive manual effort. This automation not only accelerates decision-making but also reduces the risk of human error. Moreover, the model’s ability to handle large volumes of data efficiently ensures scalability, supporting future growth and adaptation to varying market conditions.
Next steps
Given this hugely encouraging proof point the client is now working with lenders and partners to obtain more extensive data for more prolonged testing, to see if the probability accuracy can exceed the existing manual calculator.
As always with AI, it comes down to the data and associated quantity, quality and related privacy concerns. Some risks that in particular we flagged to the client to be aware of:
- The banks do not suitably anonymise borrower data and client is able to see PII (personally identifiable information) data.
- The banks provide data which is too old to reflect current behaviours, or just generally of poor quality.
- The examples that are provided by the banks are not suitably exhaustive and representative user profiles are missed from the training.
- Client doesn’t keep sufficient awareness of the actual lender decisions, so cannot track the accuracy and effectiveness of our model which would therefore degrade.
Impact
At this early stage, any return on investment is still to be proven, but is likely to be a combination of the following:
- Possible improved matching beyond the existing LMS will result in happier lenders (they only receive enquiries which are highly likely to be approved, thereby not wasting their time) and happier borrowers (not wasting time with lenders that ultimately reject them). This will of course over time make the service more attractive to parties who then direct more lending activity to the platform, therefore increasing revenues.
- Efficiency and error reduction savings from removal of human reliance on algorithm updates.
