How can machine learning be used to improve credit underwriting?

LexisNexis Risk Solutions is a leader in providing analytics solutions worldwide that integrate non-traditional data, such as Digital ID, Device, and Behavior intelligence, improving the depth and diversity of credit profiles. This approach is especially beneficial for individuals with limited credit histories. By applying machine learning algorithms, LexisNexis effectively cleans and processes this data, identifying patterns and insights that traditional credit assessments might miss. This leads to more accurate risk assessments, helping clients make better-informed lending decisions and potentially increasing access to credit for underserved populations.

Data comes from various systems and the cleaning job involves "triage". For example, in one system it could Ram Krishnan and in a another system it could be Ramkrishnan. We need to look at other data points like ID, address etc. to ensure we have only 1 name. Many cases, address details will be missing and we need to have some augmentation to fill the right data for modeling.

Machine Learning (ML) enhances credit underwriting by analyzing vast amounts of data to make more accurate lending decisions. For Example: Imagine a bank evaluating a loan application. Instead of relying solely on traditional factors like income, ML considers a broader range of data such as spending habits, online behavior, and payment history. This holistic analysis helps the bank predict the applicant's creditworthiness more accurately, leading to fairer and more informed lending decisions.

Supervised learning is the most common in this type of modeling. However, newer techniques like auto-encoders can help is isolating groups. For example if you train an auto-encoder on all your "good" customers with good credit rating and then apply it on all the data, the outliers will start showing up. Training auto-encoders is painful, since you have to do layer by layer, but worth it - since it is also explainable.

Most loans are disbursed through a co-lending model or through a shared risk model. The risk is often shared disproportionately among the anchors (80-20 ratio usually). An anchor (like an NBFC) might often work with multiple other lenders (FinTechs) but may not want to share their underwriting policies as they are proprietary to them. To allow seamless collaboration anchors can wrap their credit policies as scores through regression models. The score based outcome may work at par with "if-then-else" conditions but gets neatly wrapped into a single score which can give a go-no-go through a numeric threshold that can be exposed to the channel partners. For Ex: A regression model that encapsulated bureau and bank statement checks.

Most times, simple explainable solutions are preferred over more accurate solutions - because this is a regulated area and the models have to be explained to the regulators. Logistic regression is very commonly used.

CX or customer experience is one of the key areas. It is easy to explain to the customer that your credit rating was "xxx" and so we can/cannot proceed. So the credit worthiness of the customer is reduced to some number which is easy to explain. When you look into so many different factors, it is not easy to explain why we are denying/agreeing to some decision.

This is a regulated area and so you have to ensure only variables which are fair need to be considered. For example, you cannot discriminate on race or color. Although you may have this data