In this project, the risk of loan default is predicted using various Machine Learning sampling techniques and Random classifiers.

• numpy
• pandas
• pathlib
• warnings
• collections
• sklearn.metrics
• sklearn.linear_model
• sklearn.model_selection
• sklearn.preprocessing
• imblearn.over_sampling
• imblearn.under_sampling
• imblearn.combine
• imblearn.ensemble

• Github
• Gitbash
• Gitlab
• Slack
• Jupyter lab
• Microsoft CSV
• WinZip

• Loan data of customers from Jan 2019- Mar 2019. Source (LendingClub)

• credit_risk_resampling (Predictions by resampling the data)
• credit_risk_ensemble (Predictions by Random classifiers)

Wherever applicable, the following values are preset throughout the model:

• The random_state is 42
• The number of estimators is 100

For both the Jupyter notebooks, the data preparation and cleaning methodology is the same.

• The final two rows of the CSV contains redundant data, thus, its omitted.
• All the null value columns are dropped.
• The loans that have 'Issued' status are removed.
• The '%' symbol from interest rate is removed. The value is converted to a float datatype.
• The target value is transformed to a binary state. The loans having 'Current' status is changed to 'low_risk', while all the other loan statuses are assumed as 'high_risk'.
• The target variable (y) is the loan status, while the feature variables(X) are everything else.
• The X variables that contain strings are binary encoded via dummy values, so that they can be scaled.

• Target: Loan Status
• Feature: All columns except loan status
• Positive is the number of predicted low-risk loans
• Negative is the number of predicted high-risk loans

• The feature and target data is randomly split into testing and training. A random_state of 42 is assigned for all instances.
• The target values in the training data is heavily disproportionate, the low-risk loans far outweighs the high-risk loans.
• The scaler function is fit to the training data.
• The testing and training data is sclaled to eliminate bias in the prediction.

• Naive oversampling
• Synthetic Minority Oversampling Technique (SMOTE)
• Cluster Centroid (CC) undersampling
• Synthetic Minority Oversampling Technique Edited Nearest Neighbours (SMOTEENN)

• Balanced Random Forest Classifier
• Easy Ensemble Classifier

For each of the prediction methods listed above, the following metrics are computed.

• Balanced Accuracy score
• Confusion Matrix
• Imbalanced classification report

• Naive: 0.82
• SMOTE: 0.80
• CC: 0.78
• SMOTEENN: 0.81
• Among the resampling methods, the Naive method has the top accuracy score, narrowly ahead of SMOTEENN and SMOTE methods.

• False positive will have the greatest damage
• Naive: 21 False Positives
• SMOTE: 27 False Positives
• CC: 19 False Positives
• SMOTEENN: 2261 False Positives (largest).

• Naive: 0.85
• SMOTE: 0.88
• CC: 0.76
• SMOTEENN: 0.87
• Among the resampling methods, the SMOTE method has the top recall score, narrowly ahead of SMOTEENN and Naive.

• Naive: 0.82
• SMOTE: 0.81
• CC: 0.79
• SMOTEENN: 0.18
• Naive method has the best Geometric mean score

• BRFC: 0.77
• EEC: 0.93

• BRFC: 35 False Positives
• EEC: 8 False Positives

• BRFC: 0.88
• EEC: 0.94

• BRFC: 0.76
• EEC: 0.93

• SMOTEENN method has the highest amount of False Positives and the lowest geometric mean among all the models.
• Naive method is apparently the best among resampling methods.
• SMOTE method fares well in many of the calculated metrics.
• CC method has the most false-negatives (4066), though it is not undesirable. Yet it is unreliable.
• Top 3 features in BRFC are:
  • 'total_rec_prncp'
  • 'total_pymnt_inv'
  • 'total_rec_int'

• EEC method comes out as the most reliable among all, with a substantial accuracy, recall, geometric mean and least amount of false-positives.

• Satheesh Narasimman

• Khaled Karman, Bootcamp tutor

• https://imbalanced-learn.org/stable/generated/imblearn.metrics.classification_report_imbalanced.html

• https://imbalanced-learn.readthedocs.io/en/stable/api.html#module-imblearn.combine

• https://imbalanced-learn.readthedocs.io/en/stable/api.html#module-imblearn.under_sampling

• https://imbalanced-learn.readthedocs.io/en/stable/api.html#module-imblearn.over_sampling

• https://imbalanced-learn.readthedocs.io/en/stable/generated/imblearn.ensemble.BalancedRandomForestClassifier.html

• https://imbalanced-learn.readthedocs.io/en/stable/generated/imblearn.ensemble.EasyEnsembleClassifier.html

• https://scikit-learn.org/stable/modules/generated/sklearn.metrics.confusion_matrix.html

• https://scikit-learn.org/stable/modules/model_evaluation.html