>enough to solve this problem

The goal isn't to solve the problem, as in predict with 100% accuracy. The goal is to improve by a few percentage points.

Since categorization is what ML does well and credit histories are great annotated training sets, it seems like a perfect fit.

A bank in Belarus scrapped ML model that was used to issue consumer loans after a year. Initially the model gave better returns, but after a year it was underperforming. The bank went back to a weighted average of few factors like it was before with humans reviewing the weights. In retrospect this is not surprising. The model tries to predict long-term behaviour, but besides few stable factors like profession, income, education human life still rather unpredictable even after averaging hundreds of thousands cases.

It would definitely seem like ML could improve performance here, if they fed it all the right features.

I feel like outstanding debts (amount, time since last payment) could be really helpful in this regard so people can't rack up huge debts from multiple lenders, but that would also be a privacy nightmare.

I thought credit reporting agencies already provide this data?

It’s almost as if it’s impossible to rationalize irrational behavior.

No amount of machine learning or artificial intelligence can change that.

In the context of data collection and prediction, what does it mean to be irrational? Truly random? Do people actually behave in truly random ways? Or are they predictable?

In theory you’d want the algorithm to determine how rational someone is likely to be, assuming that is correlated with debt repayment. If you can say with some degree of confidence that a person is likely to act irrationally and that people who act irrationally don’t pay back debt as often, then that’s a signal to consider when deciding to lend. Not sure exactly how you’d determine that but I guess the hope of collecting such varied data points as phone battery level, typing speed, and food delivery preferences are meant to provide some amount of signal there.

What do rational/irrational mean, quantitatively speaking? Before we even write the algorithm, we've got to come up with a formal definition and I can't even begin to think about how you would do that.

One can be predictable and irrational at the same time

Future behavior isn’t entirely predictable by past behavior. Sure. External events might put a particular individual in arrears, even though she/he has never defaulted in her/his life.

Machine learning however, can use past behavior to make an educated guess, regarding how/when/if an individual will repay and/or how vulnerable said individual is to unforeseen external events.

Machine learning might not be able to go as deep as evaluating someone’s chances of getting hit by lightning, for the purposes of credit ratings. However it does figure out if an individual has the chance to get out of sticky situations.

Summarizing, it’s not about predicting irrational behavior. It’s about using past behavior to make a better educated guess than was possible before.