My task is to make a machine learning model for predicting the probability of default when taking a loan (binary classification).

But I also want to predict the loan amount. Something like: if the client is very good, then you can give him more money than we usually give. Now the amount is billed based on experience and intuition by the loan portfolio analysts. The problem with adding the "loan amount" feature to the model and training it again is that historically we have given more money to clients who had a low probability of default. Something like:

group max Probability of default max money
1 0.05 100
2 0.1 50
3 0.15 25
4 1 0 (reject)

Because of this, the model will observe an 'inverted' correlation. The more money, the more likely the customer will pay back! Although the real situation is exactly the opposite. The more money we give the client, the less likely he is to pay it back, so only good clients should be given large sums.

Can you advise me how to correct this situation, what area of scientific knowledge should be studied, what A/B tests should be performed?


1 Answer 1


Typically this problem is solved by looking at loans in terms of percentage loss rates. IE if we loan out \$100, how many dollars is lost on defaults. If we loan 300 dollars and $15 is lost, our rate is 15/300 or 5%. This solves the problem of using a raw number. Loaning lots of money would invariably increase losses in raw terms but that's not inherently bad.

Now you have default rates by risk category with risky groups usually having a higher rate. You can then look at the interest rate you charge them and see if you are breaking even or not.

In short, probablity of default is no longer a useful metric here. If I give a riskier person slightly less money or charge a better interest rate, it might become a better or worse deal.

Update per comment "It is the amount of the loan and possibly the interest rate I want to predict."

If this is the case, I would talk to the people writing the loans for a bit more insight. It would appear that the upper limit of the size of the loan is credit worthiness. The lower limit is minimum money needed for a successful loan. Ie, you aren't going to get a car loan for 12k if the car costs 35k.

Linear regression is what I would use. Working in this area, most places in the world require high levels of explainablity so most ML is out. So internal risk managment can feel safe and for legal reasons such as discrimination and avoiding run away economic chaos.

  • $\begingroup$ It is the amount of the loan and possibly the interest rate I want to predict. Just looking at the average loss and setting the maximum for the entire risk group is the current suboptimal approach. I want to predict the amount for each client individually using the ml model. The current approach is essentially a discrete breakdown based on some expert considerations. And you can either split the group further e.g. best of the best and worst of the worst. Or predict the amount of credit. $\endgroup$
    – Andrew
    Oct 25, 2022 at 13:45
  • $\begingroup$ added an addendum $\endgroup$
    – user70889
    Oct 25, 2022 at 14:31
  • $\begingroup$ The name of the ml method did not help me much in solving the problem. Again, the problem is that historically the data is such that the more money we give to the client, the less likely it is to default. Because we funded more of those who were better ranked in terms of default probability. And I can't just add the "amount of money" variable to my dataset. $\endgroup$
    – Andrew
    Oct 31, 2022 at 15:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.