Predicting Credit Card Defaults in Taiwan

Author: Mitchell Krieger

Overview

To optimize returns, Banks must manage their risk when providing credit cards to customers. This project attempts to create a model that predicts if a client will or will not default on their loan. This project was a competition with classmates to see who could achieve the model with the best F-1 scores that predicts people who will default on a holdout dataset. My best model was a random forrest and had an F-1 score of .56 on the test portion of the dataset. When holdout competition dataset was run through the model it performed 3rd best in the class.

Data:

This datset comes from the UCI Machine Learning Repository

Methods:

First, insights into trends in the data were found using exploratotry visual analysis. Based on these insights, additional features were engineered:

• Percentage Use of Limit
• High limit, greater than $310,000 (above 85th percentile)
• Binary Late Payment or Not
• Binary Paid in full or not
• Percentage of Bill Paid
• Young (<30 years old)

Then dummy variables were added for categorical variables like sex, marital status and education level.

After spliting data into a train and test set, multiple classification techniques were attempted including:

• Logistic Regression
• Decision Trees & Radom Forest
• K Nearest Neightbors
• XGBoost

Additional models were attempted using ensemble voting.

Exploratory Data Analysis

Default rates varied slightly based on gender, education level and marital status:

The closer to a customer's bill was to their credit limit, the more defaults appeared (green dots along the $y=x$ line or the upper region of the triagnular area):

Similarly on average people were only likely to pay their card minimum (~ $5000) which was much less than their card bill:

Results

A random forest model after a tuning with grid search was the best performer using the F-1 score. The best paramenters were:

Hyperparameter	Values
Max Depth	6
Max Features	7**
Max Lead Nodes	35
Max Samples	0.75
Min Sample Split	25
Num of Esitmators	100

This model generated an F1 score of .56 on the test set.

Conclusions

On the holdout set, this model came in 3rd place of 30 competitng models.

Next Steps

Next steps are to:

• Tune models further and idenitfy better feature selection processes
• Add additional feature interactions and potentially polynomial features
• Conduct research on cultural and economic differences in Taiwan

Repository Structure

├── data                      <- directory containing data used for project modeling
├── images                    <- directory containg images of plotting
├── Holdout.ipynb             <- Jupyter Notebook containing predictions of the best model on holdout test set.
├── bakeoff_instructions.md   <- instructions for the competition
├── defaults-in-taiwan.ipynb  <- Narrative Jupyter Notebook containing EDA and modeling processes
└── preds_Mitch_Krieger.csv   <- a csv file containing default predictions on a holdout set of data