Creditworthiness Prediction for Credit Card Issuance

Binary classification model to predict customer creditworthiness, built for a fictional bank (Pro National Bank) to support credit card application decisions.

Overview

Given anonymized customer data, the model predicts whether a customer has a high credit rating (TARGET=1) or not (TARGET=0). The task presents a significant class imbalance: only 8.78% of observations belong to the positive class.

The project evaluates 8 base classifiers across 6 feature selection strategies, resulting in 291 model configurations compared systematically.

Dataset

~338,000 customer records with 19 features including demographics, employment, income, and contact information.

https://proai-datasets.s3.eu-west-3.amazonaws.com/credit_scoring.csv

Key preprocessing decisions:

• DAYS_BIRTH converted to positive values (DAYS_BIRTH_POSITIVE)
• FLAG_MOBIL dropped - 100% of values identical, zero predictive value
• DAYS_EMPLOYED anomalies handled - values exceeding human lifespan (~1,000 years) identified as Pensioners and treated separately
• OCCUPATION_TYPE NaN values (13.44% of data) replaced with "unknown" - neither deletion nor imputation was justified after systematic analysis across all other features

Pipeline

Preprocessing (ColumnTransformer):

• Numerical features → StandardScaler
• Nominal categorical → OneHotEncoder
• Ordinal categorical (NAME_EDUCATION_TYPE) → OrdinalEncoder with explicit category ordering

Feature selection strategies evaluated:

1. Base (all features)
2. VarianceThreshold
3. Correlation Filtering
4. VIF (Variance Inflation Factor)
5. SelectKBest (k = 5, 10, 15, 20, 25, 30, 35, 40)
6. RFE - Recursive Feature Elimination (k = 5–40)
7. Embedded Methods - hyperparameter tuning with RandomizedSearchCV (regularization, tree depth, class weighting)

Models

291 configurations evaluated (8 classifiers × 6 feature selection strategies + hyperparameter tuning). Final ranking filtered by Accuracy > 0.90 and sorted by Recall to minimize false negatives - failing to identify a creditworthy customer is the costliest error in this context.

Model	Accuracy	F1	Precision	Recall	ROC-AUC	FN
Random Forest emb	0.96	0.823	0.699	1.000	0.979	0
Gradient Boosting rfe k:35	0.96	0.823	0.699	0.999	0.979	5
Gradient Boosting CorrelationFiltering	0.96	0.823	0.699	0.999	0.979	5
Gradient Boosting rfe k:30	0.96	0.823	0.699	0.999	0.979	5
Gradient Boosting rfe k:20	0.96	0.823	0.699	0.999	0.979	5

Top 5 of 291 configurations. Full results in sorted_results_df.csv.

Best model: Random Forest (Embedded) - the only configuration achieving perfect Recall (1.000, 0 false negatives) with F1 = 0.823. Total training time: ~226 minutes.

Key Findings

• Tree-based ensemble methods (GBM, Random Forest, XGBoost) consistently outperform linear models on this dataset
• Feature selection strategies did not significantly improve performance over the base configuration - the full feature set was already informative
• Logistic Regression achieves high Recall only when Accuracy is sacrificed (0.828), making it unsuitable for production despite its interpretability

Tech Stack

Python 3.9 · pandas 2.2.2 · NumPy 2.0.0 · scikit-learn 1.5.1 · XGBoost 2.1.0 · SciPy 1.13.1 · statsmodels 0.14.2 · Matplotlib 3.9.1 · Seaborn 0.13.2 · tqdm 4.66.4

How to Run

The notebook runs in a local Anaconda environment.

conda env create -f environment.yml
conda activate creditworthiness_project
jupyter notebook

Download the dataset manually and place it in the project folder:

https://proai-datasets.s3.eu-west-3.amazonaws.com/credit_scoring.csv

Open Prediction_of_creditworthiness_for_issuing_a_credit_card.ipynb and run cells sequentially.

Project Structure

creditworthiness/
├── Prediction_of_creditworthiness_for_issuing_a_credit_card.ipynb   # Main notebook
├── environment.yml                                                    # Conda environment
├── results_df.csv                                                     # All 401 model results
└── sorted_results_df.csv                                              # Final ranked results