MACHINE LEARNING · GRADUATE COURSEWORK (ISM 6136)

Modeling Default Risk

An end-to-end pipeline predicting loan defaults, and using the results to propose an actual lending strategy, not just a model.

After EDA, I ran k-Means clustering (k=4, silhouette 0.514) to segment borrowers, then compared Decision Trees against neural networks for the prediction task. I selected a reduced neural network on recall (0.83) and F1 (0.72), because in default prediction, missing a defaulter costs far more than flagging a good borrower, so recall matters more than raw accuracy.

The clusters mapped to default rates ranging from 7.6% to 73.7%, which let me propose a two-stage lending pipeline: segment first, then score, turning the model into a decision the business could act on.