Loan Default Prediction

An ML investigation on the defaulting of loans.

Description

When an borrower defaults on a loan, they fail to repay their debt according to the initial arrangements set by the debtor -- a failed repayment proves to be an issue with both parties involved. Our goal was to predict whether a borrower would default on a loan, given their outlined financial statements using Principal Component Analysis, and K-nearest Neighbors.

Role | Member

Working on the project, I acted as a team member alongside two other students to follow through on our goal of accurately predicting a borrower's defaulting behavior. As teammates, we all followed through on our strengths, collaborating on the direction of the project and ensuring representative results with regular code reviews.

Creating a Model

As an overview of our methodology, we took a training, testing, and validation approach to predicting the defaulting of loans based upon 50 features provided a pruned sample of individual financial statements. With all of our implementations with custom algorithm configurations in Python, we tuned our hyperparameters for PCA to include a dimensionality reduction of 6, minimizing our error rate of ~8.4%. As for our K-nearest Neighbors implementation, we minimized our error rate over sample sizes of 100, with k-values of 2 and 4 respectively, at ~10%.

Key Takeaways

As one of my first involved applications of m achine learning, I found the importance of not only understanding the underlying models and algorithms, with our custom implementations, but the importantance of cross-validating our own implementations among ourselves, helping to reduce compute time by getting it right the first time, and pushing for accurate and representative results.