In the early years of PeopleFund, we spent approximately 4 billion won purchasing credit data. This was not a marketing decision. It was a scientific one.
We believed that the legacy credit scoring systems — built by credit bureaus using a narrow set of financial history signals — were leaving predictive signal on the table. The question was whether we could find it.
What a credit score actually measures
A traditional credit score measures a person's history of using credit. It captures whether you have borrowed money, whether you repaid it on time, how much credit you have outstanding relative to your limits. It does not capture much about the person.
This is not a criticism of the credit bureau model. It is a description of what that model was designed to do: minimize false negatives, prevent defaults, protect lender balance sheets. It was built to be conservative.
The consequence of that conservatism is that large populations of people who are genuinely creditworthy — in the sense that they will repay their loans — are denied credit or offered it at rates that reflect risk the lender has assumed rather than risk that is actually present.
The 1,600 variables
We assembled a dataset of 1,600 variables. Some were traditional: employment history, income, existing debt. Many were not.
We incorporated behavioral signals from mobile applications — time spent in specific categories, patterns of digital activity that correlated, in our data, with repayment behavior. We included consumption patterns, transaction histories, and signals from networks of relationships that traditional credit bureaus had no visibility into.
The goal was not to build a more sophisticated version of the existing model. It was to build a fundamentally different model — one that could identify creditworthiness in populations the existing system had defined as unworthy by default.
What the data showed
Our default rate was 2.5 percent. The industry average at the time was five percent. Our delinquency rate was 0.87 percent against an industry average of 4.2 percent.
These numbers have a direct economic consequence: we could offer credit at rates three to four percentage points lower than competitors and still achieve the risk-adjusted returns the business required. The borrower who was paying twenty percent interest to a competitor was paying sixteen percent to us, and repaying at a higher rate.
The credit score had been wrong. Not systematically wrong, but wrong at the margin — and the margin is where the people who most need credit live.
What this means beyond lending
I think about the 1,600 variable problem whenever I encounter a system that is using a small number of signals to make decisions about people who are more complex than those signals allow.
The pharmacy that treats all patients as transaction counterparties rather than people managing health conditions. The wellness brand that defines its customer by demographic rather than by what they are actually trying to accomplish. The hiring process that uses credential signals because measuring actual capability is harder.
In every case, there are more variables available than the system is using. The question is whether the cost of collecting and processing them is worth the improved decision quality.
At PeopleFund, the answer was yes — and the evidence was in the default rate.
I am still looking for the next 1,600 variables. I think I always will be.