My AI Suggested a Technique from Cancer Research

The Problem with Averages

Ask most lenders about their default rate and they'll give you a number. “Our portfolio defaults at 8%.” That number is technically true and practically useless.

It tells you nothing about timing. Do loans default in the first 90 days or after 18 months? Is the risk front-loaded or does it creep in late? Are there specific windows where defaults cluster? An average can't answer any of those questions. It flattens the entire story into a single point.

We were building analytics for a specialty lending portfolio — roughly $40M across hundreds of merchant cash advance loans. The risk team needed more than a default rate. They needed to see the shape of default over time.

An Unexpected Suggestion

I asked AI for survivorship analysis techniques. Survivorship analysis is a common concept in machine learning and statistics — modeling how long something lasts before an event occurs. I expected a textbook recommendation.

AI came back with Kaplan-Meier curves. The technique was developed in 1958 by Edward Kaplan and Paul Meier to study patient survival rates in clinical trials. It's the standard method in oncology for answering the question: given a cohort of patients, what's the probability of surviving past any given point in time?

The connection was immediate. Replace “patients” with “loans.” Replace “survival” with “performing.” Replace “death” with “default.” The math is identical. And it solves the exact problem averages can't: it shows you the trajectory.

What the Curve Shows

A Kaplan-Meier curve starts at 100% and steps down each time a loan defaults. The x-axis is time since origination. The y-axis is the percentage of loans still performing.

Where the curve drops steeply, defaults are clustering. Where it flattens, loans that made it past that point tend to keep performing. The shape tells you things a bar chart never could:

• The danger window.If the curve drops sharply in months 3–6, that's where your underwriting needs to be tightest. Resources spent monitoring loans after month 12 might be wasted.
• Partner comparison.Overlay curves from different origination partners. One partner's loans might survive longer than another's — even if their overall default rates look similar.
• Vintage analysis.Compare loans originated in Q1 vs Q3. If Q3 loans default faster, something changed — in the market, in the underwriting, or in the borrower pool.

Why It Works for Lending

Kaplan-Meier has a specific feature that makes it perfect for loan portfolios: it handles censored data. In oncology, “censored” means a patient dropped out of the study or the study ended before they died. In lending, it means a loan was paid off early, refinanced, or is still active.

Most simple default analyses throw out active loans or treat them as non-defaults. Kaplan-Meier keeps them in the calculation for as long as they're observed, then correctly accounts for the fact that we don't know their final outcome. This gives you a more accurate picture, especially for newer vintages where most loans are still performing.

It's also non-parametric — it doesn't assume defaults follow a normal distribution or any other shape. It lets the data speak for itself. That matters in specialty lending where default patterns can be highly irregular.

The Co-Pilot Knew a Route I Didn't

I knew what I needed: a way to see default risk over time, not flattened into an average. I had the domain knowledge to frame the problem. What I didn't have was the specific technique from biostatistics that solved it perfectly.

That's one of the underappreciated modes of working with AI. The pilot/co-pilot dynamic isn't always the human steering and AI executing. Sometimes the co-pilot knows a route you've never flown. The skill is recognizing when the suggestion is right — and you can only do that if you understand the problem deeply enough to evaluate the answer.