A regression-based ML project that identifies the key drivers of credit card spending across 5,000 bank customers — enabling data-driven credit limit calculation for new applicants.

A global bank wants to understand what factors drive credit card spend across its customer base. This understanding is critical for:

• Setting accurate credit limits for new applicants
• Identifying high-value customer segments
• Reducing credit risk by predicting spend behaviour upfront

The bank surveyed 5,000 customers, collecting demographic, financial, and behavioural data. This project builds a predictive model to:

1. Identify the top drivers of total credit card spend (Primary + Secondary card combined)
2. Predict the total spend for new applicants given their profile
3. Use those predictions to recommend credit limits at the individual level

• Source: Internal bank customer survey
• Size: 5,000 customers
• Format: .xlsx with accompanying data dictionary
• Target Variable: total_spent (Primary Card spend + Secondary Card spend, log-transformed)

Note: Data is numerically encoded. Categorical variables require careful separation before modelling.

Predicting-Credit-Card-Spend/
├── data/
│   ├── credit_card_data.xlsx        # Raw customer survey data
│   └── Data_Dictionary.xlsx         # Variable definitions
├── outputs/
│   ├── Decile_analysis_train.csv    # Decile analysis on training set
│   ├── Decile_analysis_test.csv     # Decile analysis on test set
│   └── python_FA.xls                # Factor analysis output
├── Predicting Credit Card Spend.ipynb   # Full EDA + Modelling notebook
├── requirements.txt
└── README.md

Key challenges tackled:

• Numerical encoding masks categorical variables — required careful type separation
• Missing value treatment: variables with >25% missing dropped; remainder imputed
• Outlier treatment using 1st–99th percentile clipping to reduce skew

# Outlier clipping at 1st and 99th percentile
credit[numerical_var] = credit[numerical_var].apply(
    lambda x: x.clip(lower=x.dropna().quantile(0.01),
                     upper=x.quantile(0.99))
)

Variables dropped due to multicollinearity:

drop_col = ['addresscat', 'agecat', 'cardtenurecat', 'card2tenurecat',
            'cardtenure', 'commutecat', 'edcat', 'equipmon', 'inccat',
            'lnlongten', 'longten', 'spoused', 'spousedcat', 'tenure',
            'tollmon', 'wiremon']

• Log transformation on target variable (total_spent_ln) to normalize right-skewed spend distribution
• Dummy variable creation for categorical variables (excluding boolean variables which are already binary)
• Missing value analysis — calculated % missing per variable to guide imputation strategy

Missing_values = pd.DataFrame(
    credit_conti_var.apply(continuous_var_summary).T.round(1)['NMISS'] / 5000 * 100
)
# Drop variables with >25% missing values
M_V = Missing_values[Missing_values['NMISS'] > 25]

• Linear Regression as the primary model — interpretable, directly maps to business insight
• Train/test split for out-of-sample validation
• Focus on coefficient interpretation to identify key spend drivers

A key differentiator of this project — Decile Analysis validates model performance in business terms:

# Group predictions into deciles and compare predicted vs actual spend
Predicted_avg = train[['Deciles', 'pred_tot_spend']].groupby(
    train.Deciles).mean().sort_index(ascending=False)['pred_tot_spend']

Actual_avg = train[['Deciles', 'total_spent_ln']].groupby(
    train.Deciles).mean().sort_index(ascending=False)['total_spent_ln']

This shows whether the model correctly ranks customers from highest to lowest spenders — critical for credit limit assignment.

• Income category is the single strongest predictor of total spend — higher income customers spend significantly more
• Card tenure (how long the customer has held the card) positively correlates with spend — loyal customers spend more
• Secondary card usage is a strong signal — customers using both primary and secondary cards show 2x higher total spend
• Customers with longer address tenure tend to be more financially stable with predictable spend patterns
• Log transformation of the target variable significantly improved model fit by reducing skew

# Clone the repository
git clone https://github.com/vicky60629/Predicting-Credit-Card-Spend.git
cd Predicting-Credit-Card-Spend

# Install dependencies
pip install -r requirements.txt

# Launch the notebook
jupyter notebook "Predicting Credit Card Spend.ipynb"

The model's business performance is validated through decile-level comparison of predicted vs actual spend — a standard technique used in banking and insurance to evaluate regression models.

A well-performing model shows monotonically decreasing averages from Decile 1 to 10, confirming it correctly ranks customers by spend potential.

What worked well:

• Decile analysis goes beyond standard RMSE/R² — validates the model in business language that stakeholders understand
• Careful missing value and outlier treatment significantly improved model stability
• Log transformation of spend target made residuals more normally distributed

Future enhancements:

• Try XGBoost / LightGBM for potentially better predictive accuracy
• Add SHAP values for better feature importance explainability
• Build a Streamlit dashboard for interactive credit limit prediction
• Experiment with Ridge / Lasso regression for automatic feature selection
• Extend to credit limit recommendation engine using predicted spend buckets
• Track experiments with MLflow

Vicky Gupta — Data Engineering Analyst @ Accenture (4.5 years) | Aspiring Data Scientist

Experienced in PySpark, ETL pipelines, and building end-to-end ML solutions across banking, retail, and NLP domains.

🔗 LinkedIn | GitHub

This project is licensed under the MIT License — see the LICENSE file for details.

⭐ If you found this useful, please consider starring the repository!