Three Reasons Why AI Enables You to Find Better Borrowers
It is no secret that Artificial Intelligence (AI) improves credit underwriting – however, we often receive the question why AI improves credit underwriting. This blog post includes three reasons why AI outperform traditional credit underwriting (e.g scorecard) development.
1) 1: Linear vs Non-Linear Models
When using AI, you use non-linear models compared to linear models which today are the most common models used in underwriting.
Above you can see two pictures both including bad loans (blue points) and good loans (green points). When using a linear model, a straight line is used to differentiate a good/bad loan. As can be seen in the left picture is it impossible to draw a straight line capturing all good loans on one side and all bad loans on the other side. Which is more easily done with the non-linear model.
To be more concrete, let’s say we have the variable age and we can see that people over 35 years old are in general better payers compared to ones below 35. This information can be used in a linear model. However with a non-linear model we can go deeper into the data. A non-linear model enables us understand that not all under 35 are necessary bad payers – by combining age with where you live, it could actually be positive to be below 35 compared to being over 35. By using a non-linear model we can capture these rare cases by looking at variables in combination and thereby find better borrowers.
2: Change-over-time vs. Point-in-Time
Most credit underwriting models use a point-in-time solution, meaning that you analyse questions such as: How many payment defaults have you had the last 3 years? or How many bankruptcies have you been involved within the last 2 years? These answers and predictors are very useful but limited, since we don’t have the ability to understand trends over time. By using change-over-time solution, you can also analyse when the payment defaults occurred or when the bankruptcies occurred.
If you think about it who would you prefer to grant a loan to?
A borrower who missed a few payments three years ago but has had a perfect record ever since or a borrower who has never missed a payment until the past few months, and missed a bunch in a row?
3: Vastly More Data
AI is an extremely good technology for analysing huge amounts of data and understanding variables from different sources and with different distributions. In order to concretise how vastly more data enables higher accuracy will we use an example below. In the example will we build a simple model that will determine if a person is a man or a women.
The first variable we choose is height, since men (in general) are taller than women. This is not true for everyone since it exist short men and tall women. Therefore we choose our next variable as weight, since men in general are heavier than women. The problem is that now our model thinks that all kids are women, which of course not is true. Therefore the third variable will be age, and our model become quite accurate.
If I had asked you if age could be used for determining gender one minute ago, you would probably have said that we were going nuts. Since, only age is a terrible predictor when trying to understand a person’s gender. When you are using a linear model such as logistic regression, the model itself will interpret each variable separately and a linear model will give you an answer which is not correct when using variables that are dependent on each other. Today when developing scorecards, this problem is solved by calculating the importance of each variable and then you are able to use these variables together.
However if you would have used 200 variables it would be impossible to put a scorecard on top of all those variables unless you have a huge analytics team that works day and night for creating one scorecard.
Artificial intelligence enables you to get this correlation for free, the model itself understand that age is only a good variables when it is combined with weight and age. The best part is that AI can understand these relations when you are working with hundreds or thousands of variables – resulting in a credit decision model that truly understands payment behaviour in the details and enables you to find better borrowers compared to traditional techniques.
The Real Challenge – Explainability
Utilizing AI in your credit model is hard but manageable. The real challenge is that an credit model has to be understood so it can pass your existing risk management committee and compliance requirements. This requires you to dig deep into the complex math behind the AI models – our recommendation is to find a partner who already has addressed these types of challenges before since it is really time-consuming, math-heavy and complex processes. At Evispot we have developed a solution, called Traits, customized to give creditors the transparency and explainability required to take an AI model into production.
If you’re ready to capture the benefits AI can give your credit risk models, let’s have a chat!