Dear credit and lending team,

The project of mine closest to your work is The $150,000 cliff. It uses all 1.69 million SBA 7(a) loans to measure how one guarantee rule shapes which loans get made, and what that does to a default model trained on them. Three things from it matter for lending.

Policy shapes your training data. A default model fit on loans from one guarantee era predicted 6.5% five-year charge-offs for the next era; 7.9% happened.

Some fields are written after the outcome. The public SBA file rewrites loan terms after a default, which pushed my first model's AUC to 0.98. How I caught it.

Sorting isn't screening. Loans at the cutoff charge off 3.4 points more over five years, but the gap mostly disappears once you account for which loans get pushed there.

Before graduate school I co-developed a weighted investment risk index at KPMG Risk Consulting, separating policy, market and execution risk for overseas energy projects, and at Youde Investment Management I ran downside scenarios that put about $5.6 million of potential losses on an office-property book. The time machine problem, about point-in-time data, matters for any credit backtest.

If any of this is close to what your team works on, I'd be glad to talk. I'm looking for a part-time, remote internship, 10 to 20 hours a week: adamzhao795@gmail.com.

Junming