This was an independent project that I worked on during the Summer of 2019. I am self taught in Python, with no formal instruction. I used Scikit Learn's reference page to help with the implementation and also read a couple of blog posts & articles.

In general, the most influencing factors on a debtors propensity to default were their payment history (do they pay their credit card bills on time) and their current balance (how big is it and how long have they been carrying it). There were plenty of other demographic and financial features that had little to no correlation with default.

I started with the SVC model and read up on parameter tuning until I found out about GridSearchCV. Understanding the C and gamma parameters and different kernel types was helpful, and I eventually got my raw accuracy to the low 80's. I did notice a substantial number of Type II (false negative) errors, which is problematic because it means lost dollars in this case.

Then I moved to the Decision Tree Classifier, which was an easier implementation.
It seemed too simple at first, but when I grew the number of leafs and branches on the tree, I realized that the model was overfitting. So, I limited it to two levels (three nodes total) and was again able to achieve low 80's accuracy with a very simple model.

I wanted to find a way to get the probability of a client defaulting vs paying their debt, so SVR was the third and final model. I first had some trouble with a lack of differentiation in the answers (a ton of test predictions with the same probability for each example), and the realized that my gamma value was way too high. I decreased it, and acheived a score of 86-87%.