I built the Loan Default Prediction project from scratch in NumPy, implementing logistic regression, a decision tree, and a random forest without any ML libraries. The code lives in a GitHub repo my GitHub profile and powers a live Streamlit demo.
| Model | AUC | Precision | Recall | F1 |
|---|---|---|---|---|
| Logistic Regression | 0.636 | 0.104 | 0.517 | 0.173 |
| Decision Tree | 0.837 | 0.544 | 0.107 | 0.179 |
| Random Forest | 0.861 | 0.342 | 0.568 | 0.427 |
Random forest is the final model because it outperforms the others on every metric.
At the default 0.5 threshold logistic regression scores 93 % accuracy while catching zero of 2,020 real defaulters.
Tuning logistic regression to a threshold of 0.07 raises recall to 52 %, capturing 1,045 of 2,020 defaulters.