@misc{misc,
  author = {Joshua Mvula and Lighton Phiri},
  title = {Credit Scoring System for Multiclass Loan Applicant Classification Using Machine Learning},
  abstract = {Credit assessment at many lenders in Zambia still depends on manual judgement and on formal credit histories that large parts of the population do not have. This project developed a credit scoring system that classifies loan applicants into four ordered risk categories (low, medium, high and very high) using machine learning. Loan and repayment records from the PostgreSQL database of Hobbiton Investment Management were extracted, consolidated at borrower level and transformed into six behavioural features: average loan amount, average number of instalments, loan count, average days late, maximum days late and overdue loan count. An audit of the database showed that demographic, employment, education and alternative-data fields were absent or unreliable, which limited the feature set to repayment behaviour. The study followed CRISP-DM and Scrum. Four classifiers (Logistic Regression, Decision Tree, Random Forest and XGBoost) were trained on 293 borrowers and tested on 126 under identical conditions. XGBoost performed best, with a test accuracy of 85.71\%, a macro-averaged F1-score of 85.11\%, a micro-average ROC-AUC of 0.98 and a cross-validated accuracy of 87.59\% ({\textpm}3.06\%). The selected model was packaged as a Flask REST service hosted on Render and is called through a PHP REST gateway and a PHP MVC web interface hosted on InfinityFree. For each applicant the prototype returns a risk category with class probabilities and reports failures with distinct error codes. Because the risk labels were derived from repayment rules rather than observed defaults, future work should validate the model against real default outcomes and extend the feature set when richer data become available.},
  year = {2026},
  pages = {104},
  publisher = {University of Zambia},
  address = {Lusaka, Zambia},
}