For the past few years, we have been hearing about predictive analytics that can identify a borrower who is going to default before the default actually happens. 

Defaults aren't a prediction problem that technology will eventually eliminate. Defaults are an uncertainty that lending produces, by design. What lenders need is not machinery that assumes this uncertainty can be removed, but machinery designed to operate effectively under it. 

You cannot predict at an individual borrower level who is going to default. You can, at best, predict, over a sufficiently large cohort of borrowers, what the default rate might be. 

The behaviour of large groups can be predicted statistically. The behaviour of a particular individual cannot. 

If you have one million borrowers, you can make remarkably good statements about what will happen to the portfolio. Narrow it to 100,000 borrowers in Karnataka, and you can still make useful predictions. Narrow it further by repayment history, geography, employment, credit score and other risk characteristics, and you can create increasingly meaningful cohorts. 

But as the cohorts get smaller, the certainty of those predictions starts decreasing. 

Eventually, you arrive at an individual borrower, say Ashok.Ashok's salary didn't arrive yesterday because his employer had a cash-flow problem. Your model cannot know that. It cannot know that unless you invade the privacy of Ashok and everyone he interacts with. Which you absolutely should not do. 

This is the fundamental limitation of trying to predict defaults at an individual level. 

For example, if we look at borrowers onboarded in Karnataka who have paid 60% of their EMIs on time and have a credit score of 650 or above, we may be able to say with reasonable confidence that 90% of them will repay on time in April.

Source link