Machine learning prediction workflow

The problem

A lender wanted earlier warning of repayment problems to offer support before default.

Findings

• • The model clearly outperforms the baseline on held-out data
• • Recall at the chosen threshold fits the follow-up team's capacity
• • Performance is checked across loan types before use

Practical impact

Shows a responsible path from model to operational decision.