Mid-Bootcamp Project: Credit Card Revival

Objective

The project will focus in a classification problem: as a risk analyst of a bank I will analyse and understand the demographics and other characteristics of the bank customer's that accept or reject a credit card offer. The task is to build a model that will try to predict if the credit card offer is accepted, by using logistic regression and KNN classifier models.

Hypothesis

Can the models used predict if a customer will accept the credit card offer?

Data Set

Data from focused marketing study, with 18,000 current bank customers. Dataset included customer's demograpgics like income level, houses owened, household size and some customer's characteristics related to the bank like number of accounts opened, average balance, number of credits cards held, etc.

Data Processing and Data Cleaning

Data was quite clean, nevertheless I have decided to clean the headers, replace the null values by the average of the column, and setting the customer number as index as they were unique identifiers.

I used BoxCox transformation and Standscaler as data transformation tecnhiques. As data was imbalanced I have used some tecnhiques to deal with it: SMOTE, class weights, UpSampling and Tomek Links.

SQL & Tableau

This project contains two other analysis made with SQL and some visualizations in Tableau that can be foun in this repo. In the SQL analysis it can be found the number of unique values for relevant columns, customers with highest average balances, total average balances for all customers, average balance per income level, a selected view with some customers characteristics relevant for the business, etc. In the Tableau part, an exploratory analysis was made containing visuals regarding data distribution between offer accepted Yes/No (with percentages), trendline concerning the average balances throughout the year and the average balances for the different household size, etc.

Models used

For this problem I have decided to use Logistic Regression and KNN Classifier.

Results and Conclusions

Overall the models did not perform very well, having troubles to predict the minority class. The model with better performance was Logistic Regression with Up Sampling technique (accuracy: 0.65, recall: 0.76, F1-score: 0.20)

Main libraries used on the project

• Pandas
• Seaborn
• Matplotlib
• Sklearn