Author: Alec Hing
Using a classification model, this project will focus on predicting whether an applicant is a "good" or "bad" customer who's applying for a credit card based on an application record and credit record dataset.
A credit card can be a useful tool in the daily lives of a consumer. It helps to build credit and trust for future loans one might take and allows someone to purchase something where they might not have direct cash on hand. It takes a responsible person to balance their budgets and know their spending limits. In the case where someone cannot pay off their amount owed by their due date, it results in their credit score to be negatively effected. If this case happens too frequently, they will most likely be deemed an at risk applicant for future applications to banks. Not being able to pay off their bills can be due to numerous factors someone goes through in life, which get taken into account when an applicant applies for a credit card (age, years employment status, income, and credit history to name a few). With all that being said, determining if a person is an at risk applicant can be hard with so many factors to consider.
The datasets used in this classification model are from Kaggle. Two datasets are provided, those being: application_record.csv and credit_record.csv which are found in the data folder.
The following featuring were used in total for this data analysis:
GenderOwn_carOwn_propertyWork_phonePhoneEmailUnemployedNum_childrenNum_familyAccount_ageTotal_incomeAgeYears_employedIncome_typeEducation_typeFamily_statusHousing_typeOccupation_type
This classification model project consisted of going through the scrubbing/cleaning process, preprocessing, developing a baseline model and then from there, numerous other models to hopefully improve the results of the baseline. Those other models that were created are: a baseline with SMOTE data, Grid Search CV, Random Forests, using undersampled data for the Random Forest model, XG Boost, and finally another XG Boost Model where the paramerters are tuned.
After each model there was a confusion matrix and classification report made. The values that were paid attention to the most were the number of False Negatives and the Recall value. Accuracy was also taken into account on the final selection of which model preformed the best. False negatives played a huge role in the finaly result for this project because it would be a worse outcome for a stakeholder if the model predicted a person was not a risky client when in fact they were. Conversely, false postives did not matter as much because predicting a person is a risk when they actually were not would not be as bad for a company.
In the end the 2 best models ended up being the Random Forest model and the model using SMOTE data. Each had around a recall value of ~60% and an accuracy of over 90%. Meaning overall the model predicted 90% correctly and was able to detect at risk clients 60% of the time. Most of the data removal from the datasets came from removing duplicates, which is why the dataset reduced to about 100,000 which is still a significant amount of data.
Out of all the actual people that owed money, thse 2 models predicted ~60% of them correctly which is way better than all the other models that had a recall under 20%. Therefore, due to these results I would recommend using the random forest or SMOTE model going forward. These model were able to get a high accuracy too which helps with identifying clients that can receive a credit card too.
The major feature that had the greatest impact on this model was a person's age. Other features that had an impact on the model were: a person's total income, their account age, and the years a person was employed for.
All in all, the best 2 models that came out of this analysis were the Random Forest and model with SMOTE data. Their recall values were over %60 and had accuracies over 90%.
Undersampling the data did give a better result in terms of recall; however there was a significant drop in useable data.
In the future to imporove these models, several things can be done. Number 1 would be to try and improve the overall recall value without undersampling. This could be done with more parameter tuning. There were also some features in these models that probably unnecessary like: email, phone, and work phone to name a few. Reducing noise in these dataset would have probably have a positive affect.
For any additional questions, please contact Alec & ahing619@gmail.com Full analysis of this project is in a Jupyter notebook or in the presentation
├── data <- Both sourced externally and generated from code
├── images <- Generated from code
├── students.ipynb <- Analysis in of this project in Jupyter notebook
├── Phase3_Presentation.pdf <- Non- Technical PDF version of project presentation
└── README.md <- Overview of what this project entails and its results