Lending Club Loan Approval

Data Information

• Lending Club Loan Data 2007-2011 containing loan status and relevant financial information.

• 42,538 entries, 115 features

• Target feature (loan_status) reclassified/binarized as loan_type.

  • 1 (Good Loan): Fully Paid
  • 0 (Bad Loan): Charged Off, Does Not Meet the Credit Policy

Data Wrangling/Feature Engineering

1. Data Cleaning

   • Removed empty columns/null values.

2. Feature Selection

   • Eliminated redundancy/data leakage.
   • Retained relevant/useful features.

3. Features Converted to Numerical Data Type

   • Feature containing numerical values

     • revol_util: revolving line utilization rate.

   • Features containing ordinal values

     • grade: loan grade assigned by Lending Club.
     • emp_length: employment length in years.

4. Newly Added Feature

   • fico_range_avg: mean value of fico_range_low and fico_range_high.

5. Target Feature Engineering

   • Excluded loans in-progress.
   • Target feature reclassified into two groups (Good Loan, Bad Loan), then binarized.

Exploratory Data Analysis

Pearson Correlations

• Notable loan_type correlations

  • grade: loan grade assigned by Lending Club.
  • fico_range_avg: mean value of fico_range_low and fico_range_high.
  • revol_util: revolving line utilization rate.
  • inq_last_6mths: the number of inquiries in past 6 months (excluding auto and mortgage inquiries).

Term/Loan Amount

• Loan amount tends to increase with longer term regardless of loan type.

Influence of Applicant Location

• States with highest # of good loans ≈ States with highest # of bad loans
• Applicant location (state) is not a good predictor for loan type.

Loan Type Distribution

• Grade description: 0 as worst, 6 as best
• Ideally, the number of bad loans should decrease with higher grade.

Modeling

Overview

• Nature of task

  • Supervised learning
  • Binary classification of highly imbalanced data

• Machine learning tools used

  • Scikit-Learn
  • Imbalanced-Learn
  • XGBoost

Procedure

I. Data Preprocessing

1. One-hot encoding on categorical features
2. Conversion of datetime objects to ordinal numeric
3. Training/Test split (80%: 20%)
4. Standardization of features with StandardScaler
5. Minority class (Bad Loan) oversampling with SMOTE from Imbalanced-Learn

II. Hyperparameter Tuning with Randomized Search

• Stratified K-Fold with cv = 5
• n_iter = 30
• scoring = ‘roc_auc’

III. Training with Tuned Hyperparameters

IV. Performance Evaluation

• Evaluation metric

  • F1 scores
  • ROC AUC

• Models trained and evaluated

  • Logistic Regression
  • Random Forest
  • Support Vector Machine
  • eXtreme Gradient Boosting

Model Comparison

• Best model: eXtreme Gradient Boosting

Features of Importance

• Primary features of importance

  • last_credit_pull_d: the most recent month in which Lending Club pulled credit for this loan.
  • grade: loan grade assigned by Lending Club.
  • inq_last_6mths: the number of inquiries in past 6 months (excluding auto and mortgage inquiries).

Assumptions/Limitations

• Assumption

  • Data collected from loan applicants is genuine.

• Limitations

  • Outdated data.
  • Large volume of missing features (only 24 out of 115 features usable).
  • Limited sample size (only 39,239 entries used).

Conclusion

• Best model: eXtreme Gradient Boosting

• Primary features of importance: last_credit_pull_d, grade, inq_last_6mths

• Prospective improvements

  • Up-to-date dataset
  • Hyperparameter tuning with different techniques
  • Alternative classifier algorithms: Neural Network, Deep Learning, etc.