Default rates in a loan book

Wahat Lending (a fictional consumer-finance company) approved 3,000 personal loans in 2023 and tracked each one for 24 months. 12.9% defaulted (90+ days past due). The credit committee is reviewing its approval policy and has asked the risk analytics team for:

1. the book's default rate with a confidence interval, and default rates by credit-score band and loan purpose;
2. a simple model of default risk that the committee can read (which factors matter, and how much);
3. a view on a proposed rule: decline every applicant with a credit score below 620.

The committee will also ask whether lending for business purposes is “riskier because of what the money is used for” — be careful with causal language.

The data. loans has one row per approved loan: purpose, verified monthly income, years with the current employer, credit score (300–850), debt-to-income ratio (%), amount, term, interest rate and a 0/1 default flag. It is synthetic, generated from a documented risk model; it describes no real lender or borrower.

Stakeholders. The credit committee (sets policy), the sales team (worried about turning away good customers), and, in this scenario, a regulator that expects lending decisions to be explainable and checked for unfair impact on protected groups.

Fairness. The file contains no protected characteristics (such as sex, age band or nationality), so this case cannot tell you whether a cut-off or a model treats groups differently. Variables like credit score, income or loan purpose can act as proxies for such characteristics; a real fairness review needs that data, an agreed definition of fairness and legal/compliance input.