Predict-Credit-Card-Payment-Default

Built a classification model to predict the likelihood of credit card default

1. Executive Summary

• Goal/Task: To develop a model to predict the likelihood of default
• Metric: Optimizing for ROC_AUC and Recall (since I assume False Negative is critical for our goal)
• Engineered two new features namely "credit_utilization_rate" and "revolving_balance_rate"
  • "credit_utilization_rate" is defined as the ratio of average statement balance to credit limit
  • "revolving_balance_rate" is defined as the ratio of average revolving balance to credit limit, where revolving ratio is defined as the difference between statement balance and payment amount
• Key Indicators of Default:
  • Model suggests, prior payment history/status as a key indicator in addition to "credit_utilization_rate" to predict the likelihood of default.
  • In the payment history, specifically the payment status of the recent few months are a strong indicator. For example, if you like to predict the default in October, then payment status is September, August, July are a key indicator.
• Ethical Considerations: Following, Equal Credit Opportunity Act, 1972, I personally choose not to include age, sex, marital status, & education as an attribute to predict default. So none of these attributes are included in the model building and are dropped.
• EDA Summary:
  • Statement balance are highly correlated to each other and hence can be condensed into one variable such as average statement balance over a defined period
  • By the numbers: Within the given dataset,
    • median credit limit is about 140K NT dollar
    • median statement balance is about 20K NT dollar
    • median payment is about 1500 to 2000 NT dollar
• Model Info:
  • Based on ROC_AUC, Logistic regression, Random Forest and Light GBM all perform similarly. ROC_AUC scores are about 0.7 with Recall of about 0.51 to 0.57. All these classifiers are capable of predicting a probability.
  • Hence depending on other business/technical needs one choose from Logistic Regression or Random Forest or Light GBM.
  • I'm personally choosing Logisitic Regression because of its less complexity and a slightly better Recall than others. ROC_AUC score of 0.7 & Recall score of 0.58.
• Business Recommendations:
  • By staying true to the mission mission to empower everyone achieve their financial goals, customers who are identified by this model as default risk, can be provided customized educational resources to help them improve their financial behavior.
    • By predicting the future, the future trajectory of those customers can be changed positively
  • In addition, such customers could be matched with secured credit card products to build their financial muscle and bring them to the prime customer category.
  • On the other side, this model could help prime customers ranked properly and matched to corresponding products.
  • Some of the customers have a high credit_utilization_rate but do make their payments on-time with good standing. So we could have a product that could recommend to increase their credit limit for such customers or suggest new credit cards, after possibly collecting additional info.
  • Also there are customers who are inactive. So there could be some reach out programs designed to follow up on these customers.

2. Next Steps

• Feature Engineering: Newly created features such as credit_utilization_rate and revolving_balance_rate seems to have high feature importance. Hence with more time, would continue to think of new features from exisiting variables
  • Feature engineering seems have to good effect on the model prediction
• Additional Data: From a product perspective, I would seek additional data such as income, past bankruptcy and other external data such as other loans/debt that can improve our model performance to predict the likelihood of default
  • Clearly as identified in the univariate analysis, some customers albeit their available features being equal to non-defaulters are indeed labelled as defaulters
• Hyper Parameter Tuning: Considering time, did a first pass hyper parameter tuning. With more additional time, it may be worth looking at it further, although it seems the effect is not significant
• To make the production worthy, next step is to build a pipeline for preprocessing the data