Team Coursework Project / B
Predicting Business Credit Applicants
Machine Learning / Customer Prioritisation / Business Analytics
A team coursework project using 100,000 anonymised customer records to connect model evaluation with a practical customer-prioritisation decision.
XGBoost captured approximately 80.27% of positive applicants within the top 30% of ranked customers.
01 · Context
Universal Plus Bank wanted to identify customers who were more likely to apply for business credit so that engagement resources could be focused on higher-potential customers.
The dataset contained 100,000 anonymised customer records with 200 numerical features.
02 · Problem
The target variable was highly imbalanced at approximately 90:10.
The objective was therefore not simply to maximise prediction accuracy, but to balance precision and recall while answering a practical business question:
How many potential applicants can the bank identify if it only contacts a limited share of customers?
03 · Approach
The workflow connected data preparation, class imbalance handling, six classification models, threshold optimisation and evaluation with customer prioritisation.
- 100K customer records
- Data cleaning & preparation
- Class imbalance handling
- Six classification models
- Threshold optimisation
- Model evaluation
- Customer prioritisation
Models evaluated: Logistic Regression / SVM / Random Forest / Decision Tree / XGBoost / LightGBM
04 · Key Decision
Although LightGBM achieved slightly higher recall, XGBoost provided the strongest overall balance across precision, F1 and customer-ranking performance.
The cumulative gain analysis was particularly important because it translated model performance into a resource-allocation decision.
Model selection was a trade-off, not a single-metric competition.
| Model | Precision | Recall | F1 |
|---|---|---|---|
| Logistic Regression | 0.428133 | 0.490155 | 0.457049 |
| SVM | 0.482919 | 0.511955 | 0.497013 |
| Random Forest | 0.420343 | 0.482419 | 0.449247 |
| Decision Tree | 0.171267 | 0.479958 | 0.252451 |
| XGBoost | 0.526278 | 0.510549 | 0.518294 |
| LightGBM | 0.467651 | 0.523558 | 0.494028 |
05 · Business Translation
Instead of contacting the full customer portfolio, the model could support a prioritised outreach strategy: rank customers by predicted propensity and concentrate engagement on the highest-potential segment.
The model therefore becomes not only a prediction tool, but a mechanism for allocating sales and customer-management resources.
06 · Reflection
Machine-learning performance becomes valuable only when it can be connected to a concrete business decision.