Credit card plays an important part of everyone's life as it is not only the safer option than carrying cash but also protects an individual from fraud or scams, where the maximum amount of money or credit that a financial institution can lend to a credit card customer or client at once is known as the credit limit. These banks consider a number of crucial aspects when determining a client's credit limit, including income, payment history, and debts. As a result, they grant larger credit limits to lower-risk borrowers and vice versa. It is necessary to develop a system that can predict the client's credit limit based on a variety of features because the credit limit is susceptible to vary depending on the client's behavior and expenditure which could make this process challenging and time consuming.
The project aims to solve the regression problem by applying machine learning models to predict credit card limit. Below is the breif overview of the project workflow:
2). Cleaning the dataset to remove irrelevant and missing information using data cleaning and manipulation techniques.
3). Visualising the dataset using graphs, charts and figures to gain better understaning of the data.
5). Feature selection using backward elimination approach and performing transformation on response variable(credit limit).
6). Applying multiple machine learning algorithms to train the models and evaluate their performance using different metrics and visualisation.
The data file Credit_Card_Dataset.csv is a data set of credit card customers from a large bank. The data dictionary is given below.
| Variable | Description |
|---|---|
CLIENTNUM |
Client ID (not for analysing!) |
Attrition_Flag |
Whether or not the customer has left the bank in the last 12 months(yes or no) |
Customer_Age |
Age of customer in years |
Gender |
Gender of customer (male or female) |
Dependent_count |
Number of dependents (children) of the customer |
Education_Level |
Education level attained by the customer(primary, secondary or third level) |
Marital_Status |
divorced, married/living with partner or single |
Income_Category |
Yearly income in Euro (to nearest 1k, grouped <30k, 31-50k,51-70k,71-110k or 111k+) |
Card_Category |
Type of credit card (Blue or Gold) |
Months_on_book |
Number of months as credit card customer |
Credit_Limit |
Credit card limit |
Avg_Utilization_Ratio |
A measure of how often the credit card is used |
Pay_on_time |
Whether or not the monthly balance on the credit card paid off( yes or no) |
Random_numbers |
Random Numbers |
Importing some essential libraries, such as data manipulation libraries that offer a wide variety of capabilities to help with a range of actions on data, such as organizing or performing mathematical operations.Similarly, data visualization libraries assist in gaining important insight into the data through various graphical representations by revealing hidden patterns and trends and data modelling librabries which helps in building different statistical or machine learning models.
Importing the dataset to the notebook using the github link:
Cleaning the dataset:
1). Upon exploring the dataset, it was found that the null values present in the dataset has string datatype and thus need to be converted to null values.
2). 217 rows found to have missing values and thus needs to be removed using dropna() function.
3). Columns like CLIENTNUM that is unique for each customer , Random_numbers and Attrition_Flag does not add any value to credit limit analysis and thus these irrelevant columns can be dropped.
4).One duplicate value was found and dropped from the dataframe and datatypes of the features were found correct which later will need label encoding due to the presence of objects during the model fitting.
1). From the bar chart of the categorical variables it can be seen that almost all the variables are found to be deeply imabalanced such as Marital Status, Card Category and payment on time, excpet for the Gender variable as there is slight difference between the number of male and female customers.
2). The Pie chart shows the percentage of customers lying in different income categories where 39% customers earn less than 30k and only 8% customers earn more than 111k+ and the histogram displays the distribution of customer's age according to which the average age of a customer is found to be 46 years approximately.
3). The histogram below shows the distribution of credit limit(response/dependent variable) that is positively skewed indicating higher number of customers have lower income and tells the direction of outiers towards the left.
However,there can be seen a spike on the very right of the distribution as well.On further investigation about customers earning more than 30,000 income it was found that hike is mostly for customers with higher income category, males, a significantly large number of gold card holder out of total gold card holders.
1).The correlation matrix for numerical variables shows that a postive high correlation exists between Customer Age and Months_on_book(Number of months as credit card customer)with a value of "0.79" ,i.e with the increase in age of customer their number of months as credit card customer increases and vice versa. (Multicollinearity is present between 2 independent variable). Thus only one out of the two variable should be used in the model.
Moreover, a negative moderate correlation exists between Avg_utilization ratio and Credit Limit of "-0.48" which shows that with the increase in avg_utilization ratio (measure of how often the credit card is used) the Credit Limit for that customer decreases moderatly and vice versa.(Inclusion of avg_utilization ratio variable will dpend on it's significance added to the model)
2). Categorical variable against Credit Limit(Response variable).
(a). Gender and Credit Limit:
The boxplot shows that the males customers have much higher median credi limit(8791.0) when compared to female customer's median credit limit(2819.0).
(b). Marital Status and Credit Limit:
The median credit limit difference between customers of different marital status is not huge with the median value of single(4511.0),married(4138.5) and divorced(4701.0), inspite of imbalanced number of customers belonging to different marital status.
(c). Income Category and Credit Limit:
The box plot for the income category against credit limit displays a directly proportional relationship between the variables because with the increase in Income, Credit Limit increases as well.
d). Card Category and Credit Limit:
The median credit limit for the customers who holds a gold card category is much higher(27977.5) despite being very less in number. On the other hand, median credit limit of the blue card holders are much less (3978.5).
e). Pay on time and Credit Limit:
There is no significant difference between the median credit limit for customers who pay on time and who do not pay on time.
1). The histogram of the credit limit in the univariate analysis showed the presence of outliers on either sides of the distribution, thus these outliers needs to be removed. The technique used in this project is called trimming. After trimming 1% of the total data from both the sides, a decrease in the skewness can also be seen.
2).Before fitting the model the categorical variables needs label encoding.
Finally, the dataframe below with 5999 observations shows the trimmed and label endoded data which is ready to be used for model building.
1). Firstly, separating the independent and dependent variable into training(80%) and testing(20%) data using hold out cross validation split and performing log transformation on dependent variable(credit limit) to make the distribution close to symmetrical or normally distributed.
2). On building the model with linear regression approach, the regression results shows an R-sqaured value of 0.661. The results also shows the p values of all the independent variables , where a varaible is considered insignificant with a p value greater than 0.05 and vice versa.
Performing backward elimination approach to remove all the insignificant variables an rebuilding the model.The results shows no noticeable change in R-sqaured value.
3).Building a user defined function for all the machine learning regression models and using best hyperparamter in each model and plotting scatter plot for actual versus predicted value.
Support Vetor machine has out-performed with an R-squared value oof 0.755, MSE value of 0.030 and RMSE value of 0.172 when comapred to all the other models.On the other hand, linear regression model performed the lowest amongst all.
