Banks can now predict loan defaults using data from richer datasets — a workaround for scarce default data

A new machine learning method lets banks forecast loan recovery rates using information from other loan portfolios, even when the data structures don't match exactly. This matters because actual defaults are rare events — banks don't have enough of their own data to build accurate models alone, and this approach lets them borrow statistical power from other sources.

Loan recovery prediction is how banks calculate how much cash they must hold in reserve against potential losses — the numbers feed directly into regulatory capital requirements. Until now, banks with thin default histories faced a hard choice: build models on insufficient data (leading to either overly cautious or reckless reserve estimates) or use generic industry benchmarks that don't fit their portfolio. This method creates a middle path. It lets a bank train on another bank's richer default history while accounting for structural differences in how loans are structured and which borrowers they serve. The practical win is narrower but real: better capital estimates, potentially freed-up cash for lending, and more honest risk measurement when you can't wait for the actual defaults to accumulate.