Delinquency is a condition that arises when an activity or situation does not occur at its scheduled (or expected) date i.e., it occurs later than expected.
Create a delinquency model which can predict in terms of a probability for each loan transaction, whether the customer will be paying back the loaned amount within 5 days of insurance of loan (Label ‘1’ & ’0’)
- Used various permutations of the following models
- Random Forest Classifier
- XGBoost
- The training data was highly skewed, so tried various resampling methods
- Oversampling of majority class
- Downsampling of minority class
- SMOTE
- Also used Recursive Feature Elimination (RFE) to extract only the relavant features to train the model on
- Among the models the one with the highest accuracy (90.58%) was the following
- Random Forest Classifier, Downsampling Minority, RFE with 7 features extracted
- Although, upon analyzing the complete classification report we find that the model doesn't generalize well to the minority class.
- The recall values for all the models for the
negative (false/0)case is pretty low. - This clearly indicates there is a need to improve the performance of the model
- Given more time, I would probably try out a few more classifiers, tune the hyperparameters and attempt to generate more useful features through feature engineering.
- However for now, this is what I could come up with in a couple of days! Will try to improve upon the model as soon as possible :)