Predictive Modeling for Credit Approval¶

Project Proposal¶

Project Goal¶

The objective of this project is to develop a predictive model that forecasts credit card approval based on applicant data. We aim to identify the key factors influencing credit card approval decisions by analyzing various applicant characteristics. We will focus on model interpretability to understand the importance of different variables and explore potential biases related to age, gender, or other demographic factors. Mitigating these biases is essential to develop a fairer model.

Research Questions¶

  1. Can we accurately predict credit card approval based on applicant data?
  2. Which features most strongly influence the credit card approval decision?
  3. Are there any biases related to age, gender, or other demographics? How can we mitigate these biases?

Dataset Overview¶

  • Source: Kaggle: Credit Card Approval Prediction Dataset
  • Variables: This dataset includes information such as age, gender, income, marital status, education, and employment status, along with the target variable indicating credit approval.

The dataset has categorical and numerical variables, which will require pre-processing. Specifically, we will handle missing values, encode categorical features, and scale numerical variables for optimal model performance.

Methods¶

  1. Data Preprocessing: Handling missing values, encoding categorical variables, and feature scaling.
  2. Exploratory Data Analysis (EDA): Analyzing variable distributions, relationships, and identifying any bias.
  3. Modeling: Implementing various classification algorithms (e.g., Logistic Regression, Decision Trees, and Random Forest) to predict credit approval.
  4. Evaluation Metrics: Using accuracy, precision, recall, and AUC as evaluation metrics to assess model performance.
  5. Bias Mitigation: Evaluating model biases (e.g., based on age and gender) and applying bias mitigation techniques such as re-weighting or re-sampling.

Code to Load Dataset¶

Below is a sample code snippet to load the dataset and preview the data:

In [9]:
# Import necessary libraries
import pandas as pd

# Load the dataset
application_record = pd.read_csv('/Users/shats/Documents/VSCode/everything/STAT451/Project/application_record.csv')
credit_record = pd.read_csv('/Users/shats/Documents/VSCode/everything/STAT451/Project/credit_record.csv')
In [11]:
application_record.head()
Out[11]:
ID CODE_GENDER FLAG_OWN_CAR FLAG_OWN_REALTY CNT_CHILDREN AMT_INCOME_TOTAL NAME_INCOME_TYPE NAME_EDUCATION_TYPE NAME_FAMILY_STATUS NAME_HOUSING_TYPE DAYS_BIRTH DAYS_EMPLOYED FLAG_MOBIL FLAG_WORK_PHONE FLAG_PHONE FLAG_EMAIL OCCUPATION_TYPE CNT_FAM_MEMBERS
0 5008804 M Y Y 0 427500.0 Working Higher education Civil marriage Rented apartment -12005 -4542 1 1 0 0 NaN 2.0
1 5008805 M Y Y 0 427500.0 Working Higher education Civil marriage Rented apartment -12005 -4542 1 1 0 0 NaN 2.0
2 5008806 M Y Y 0 112500.0 Working Secondary / secondary special Married House / apartment -21474 -1134 1 0 0 0 Security staff 2.0
3 5008808 F N Y 0 270000.0 Commercial associate Secondary / secondary special Single / not married House / apartment -19110 -3051 1 0 1 1 Sales staff 1.0
4 5008809 F N Y 0 270000.0 Commercial associate Secondary / secondary special Single / not married House / apartment -19110 -3051 1 0 1 1 Sales staff 1.0
In [12]:
credit_record.head()
Out[12]:
ID MONTHS_BALANCE STATUS
0 5001711 0 X
1 5001711 -1 0
2 5001711 -2 0
3 5001711 -3 0
4 5001712 0 C
In [13]:
application_record.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 438557 entries, 0 to 438556
Data columns (total 18 columns):
 #   Column               Non-Null Count   Dtype  
---  ------               --------------   -----  
 0   ID                   438557 non-null  int64  
 1   CODE_GENDER          438557 non-null  object 
 2   FLAG_OWN_CAR         438557 non-null  object 
 3   FLAG_OWN_REALTY      438557 non-null  object 
 4   CNT_CHILDREN         438557 non-null  int64  
 5   AMT_INCOME_TOTAL     438557 non-null  float64
 6   NAME_INCOME_TYPE     438557 non-null  object 
 7   NAME_EDUCATION_TYPE  438557 non-null  object 
 8   NAME_FAMILY_STATUS   438557 non-null  object 
 9   NAME_HOUSING_TYPE    438557 non-null  object 
 10  DAYS_BIRTH           438557 non-null  int64  
 11  DAYS_EMPLOYED        438557 non-null  int64  
 12  FLAG_MOBIL           438557 non-null  int64  
 13  FLAG_WORK_PHONE      438557 non-null  int64  
 14  FLAG_PHONE           438557 non-null  int64  
 15  FLAG_EMAIL           438557 non-null  int64  
 16  OCCUPATION_TYPE      304354 non-null  object 
 17  CNT_FAM_MEMBERS      438557 non-null  float64
dtypes: float64(2), int64(8), object(8)
memory usage: 60.2+ MB
In [25]:
application_record.isnull().sum()
Out[25]:
ID                          0
CODE_GENDER                 0
FLAG_OWN_CAR                0
FLAG_OWN_REALTY             0
CNT_CHILDREN                0
AMT_INCOME_TOTAL            0
NAME_INCOME_TYPE            0
NAME_EDUCATION_TYPE         0
NAME_FAMILY_STATUS          0
NAME_HOUSING_TYPE           0
DAYS_BIRTH                  0
DAYS_EMPLOYED               0
FLAG_MOBIL                  0
FLAG_WORK_PHONE             0
FLAG_PHONE                  0
FLAG_EMAIL                  0
OCCUPATION_TYPE        134203
CNT_FAM_MEMBERS             0
dtype: int64
In [19]:
credit_record.isnull().sum()
Out[19]:
ID                0
MONTHS_BALANCE    0
STATUS            0
dtype: int64
In [24]:
# Create correlation matrix for numerical features
import seaborn as sns
import matplotlib.pyplot as plt
import numpy as np

# Select numerical columns
numerical_cols = application_record.select_dtypes(include=['int64', 'float64']).columns

# Create correlation matrix
plt.figure(figsize=(12, 10))
#exclude flag mobil
correlation_matrix = application_record[numerical_cols].drop(columns=['FLAG_MOBIL']).corr()

# Create heatmap with customized appearance
sns.heatmap(correlation_matrix, 
            annot=True,  # Show correlation values
            cmap='coolwarm',  # Color scheme
            center=0,  # Center the colormap at 0
            fmt='.2f',  # Round to 2 decimal places
            square=True,  # Make cells square
            linewidths=0.5)  # Add gridlines

plt.title('Correlation Matrix of Numerical Features', pad=20)
plt.tight_layout()
plt.show()

# Print strongest correlations (optional)
print("\nStrongest correlations:")
# Get the upper triangle of the correlation matrix
upper = correlation_matrix.where(np.triu(np.ones(correlation_matrix.shape), k=1).astype(bool))
# Stack the correlations and sort by absolute value
strongest_correlations = upper.unstack()
strongest_correlations = strongest_correlations[strongest_correlations != 0].sort_values(key=abs, ascending=False)
print(strongest_correlations.head(10))

Strongest correlations:
CNT_FAM_MEMBERS  CNT_CHILDREN        0.884781
DAYS_EMPLOYED    DAYS_BIRTH         -0.617908
DAYS_BIRTH       CNT_CHILDREN        0.349088
CNT_FAM_MEMBERS  DAYS_BIRTH          0.306179
FLAG_PHONE       FLAG_WORK_PHONE     0.290066
DAYS_EMPLOYED    CNT_CHILDREN       -0.241535
CNT_FAM_MEMBERS  DAYS_EMPLOYED      -0.234373
FLAG_WORK_PHONE  DAYS_EMPLOYED      -0.232208
                 DAYS_BIRTH          0.171829
DAYS_EMPLOYED    AMT_INCOME_TOTAL   -0.141291
dtype: float64