Table of Content

• 1. Introduction
• 2. Data Preparation and Analysis
  • 2.1. Analysis
• 3. Feature Engineering
• 4. Model Training
• 5. Results
• 6. Conclusion

import pandas as pd
import lightgbm as lgb
from sklearn.preprocessing import OrdinalEncoder
from sklearn.model_selection import cross_val_score, train_test_split
from sklearn.metrics import roc_auc_score
from catboost import CatBoostClassifier
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

1. Introduction

Project Title: Credit Risk Analysis for Personal Loan Applications

Introduction

In this project, I conducted an in-depth analysis of a loan application dataset to build a model that predicts the likelihood of a loan default. This dataset provides details about applicants, including their age, income, employment history, and loan-related features like loan amount, interest rate, and loan intent. The objective was to develop a model that could help financial institutions assess credit risk more accurately, minimizing potential losses due to defaults.

Dataset Overview

The dataset includes information on various applicant and loan characteristics:

• Applicant Information: Features such as person_age, person_income, person_home_ownership, and person_emp_length (years of employment) describe each applicant’s profile.
• Loan Information: Features like loan_intent, loan_grade, loan_amnt (loan amount), and loan_int_rate (interest rate) offer insights into each loan’s purpose and terms.
• Credit History: The dataset includes cb_person_cred_hist_length (credit history length in years) and cb_person_default_on_file, a binary indicator of past default history.
• Target Variable: loan_status, where 0 indicates no default and 1 indicates a default.

Project Goals

1. Data Cleaning and Exploration: Initial analysis to understand data distributions, handle missing values, and perform feature engineering.
2. Feature Engineering: Convert categorical variables into numerical formats suitable for modeling, such as using one-hot encoding for loan_intent and loan_grade.
3. Modeling and Evaluation: Develop and evaluate a classification model (e.g., logistic regression, decision tree, or random forest) to predict loan_status, focusing on metrics such as accuracy, precision, recall, and AUC (area under the curve).
4. Insights and Recommendations: Generate insights on factors associated with higher risk of default, supporting data-driven decision-making in loan approval processes.

Technologies and Tools

• Data Processing: Python (pandas, numpy)
• Visualization: matplotlib, seaborn
• Modeling: scikit-learn, CatBoost (or another machine learning library for handling categorical data)
• Evaluation Metrics: Accuracy, AUC-ROC, Precision, Recall

This project showcases my skills in data processing, feature engineering, machine learning, and financial risk analysis. By building an effective predictive model, I aim to demonstrate the potential for data science to improve risk assessment and lending decisions in finance.

Data Preparation and Analysis

train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')

train.head(6)

	id	person_age	person_income	person_home_ownership	person_emp_length	loan_intent	loan_grade	loan_amnt	loan_int_rate	loan_percent_income	cb_person_default_on_file	cb_person_cred_hist_length	loan_status
0	0	37	35000	RENT	0.0	EDUCATION	B	6000	11.49	0.17	N	14	0
1	1	22	56000	OWN	6.0	MEDICAL	C	4000	13.35	0.07	N	2	0
2	2	29	28800	OWN	8.0	PERSONAL	A	6000	8.90	0.21	N	10	0
3	3	30	70000	RENT	14.0	VENTURE	B	12000	11.11	0.17	N	5	0
4	4	22	60000	RENT	2.0	MEDICAL	A	6000	6.92	0.10	N	3	0
5	5	27	45000	RENT	2.0	VENTURE	A	9000	8.94	0.20	N	5	0

train.describe()

	id	person_age	person_income	person_emp_length	loan_amnt	loan_int_rate	loan_percent_income	cb_person_cred_hist_length	loan_status
count	58645.000000	58645.000000	5.864500e+04	58645.000000	58645.000000	58645.000000	58645.000000	58645.000000	58645.000000
mean	29322.000000	27.550857	6.404617e+04	4.701015	9217.556518	10.677874	0.159238	5.813556	0.142382
std	16929.497605	6.033216	3.793111e+04	3.959784	5563.807384	3.034697	0.091692	4.029196	0.349445
min	0.000000	20.000000	4.200000e+03	0.000000	500.000000	5.420000	0.000000	2.000000	0.000000
25%	14661.000000	23.000000	4.200000e+04	2.000000	5000.000000	7.880000	0.090000	3.000000	0.000000
50%	29322.000000	26.000000	5.800000e+04	4.000000	8000.000000	10.750000	0.140000	4.000000	0.000000
75%	43983.000000	30.000000	7.560000e+04	7.000000	12000.000000	12.990000	0.210000	8.000000	0.000000
max	58644.000000	123.000000	1.900000e+06	123.000000	35000.000000	23.220000	0.830000	30.000000	1.000000

train.isnull().sum()

id                            0
person_age                    0
person_income                 0
person_home_ownership         0
person_emp_length             0
loan_intent                   0
loan_grade                    0
loan_amnt                     0
loan_int_rate                 0
loan_percent_income           0
cb_person_default_on_file     0
cb_person_cred_hist_length    0
loan_status                   0
dtype: int64

status_counts = train['loan_status'].value_counts()
status_percentage = (status_counts / len(train)) * 100
round(status_percentage, 0)

loan_status
0    86.0
1    14.0
Name: count, dtype: float64

Key Observation so Far on train dataset

1. Missing Values: There are no missing values.
2. Outliers: person_emp_length and person_age look like they contain outliers with max values of 123 which seems far-fetched. So they will require further examination.
3. Target Feature: The dataset is not balanced. It has a ratio of 86:14.

columns = train.select_dtypes(include = "object").columns

for col in columns:

    print(f"Value counts for column: {col}")
    print(train[col].value_counts(), "\n")

Value counts for column: person_home_ownership
person_home_ownership
RENT        30594
MORTGAGE    24824
OWN          3138
OTHER          89
Name: count, dtype: int64 

Value counts for column: loan_intent
loan_intent
EDUCATION            12271
MEDICAL              10934
PERSONAL             10016
VENTURE              10011
DEBTCONSOLIDATION     9133
HOMEIMPROVEMENT       6280
Name: count, dtype: int64 

Value counts for column: loan_grade
loan_grade
A    20984
B    20400
C    11036
D     5034
E     1009
F      149
G       33
Name: count, dtype: int64 

Value counts for column: cb_person_default_on_file
cb_person_default_on_file
N    49943
Y     8702
Name: count, dtype: int64 

Encoding

1. Type of Encoding to be used: Given the nature of the data two types of encoding have been chosen to encode the object data type.

               - **They are:** Ordinal Encoding for person_home_ownership, loan_grade and cb_person_default_on_file as these features have some order to it.
                               One-hot encoding for loan_intent as the feature has no order to it.

2. The encoding has been done in the encoding function below.

def encoding(data, columns):

    data = data.copy()
    #using a combination of freq and ordinal encoder

    # Mapping dictionary
    mapping = {'RENT': 3, 'MORTGAGE': 2, 'OWN': 1, 'OTHER': 4 }
    data['person_home_ownership'] = data['person_home_ownership'].map(mapping)
    
    for col in columns:

        if col == 'loan_intent':

            data = pd.get_dummies(data, columns = ['loan_intent'])
        else:
            # Initialize the OrdinalEncoder
            ordinal_encoder = OrdinalEncoder()
            
            data[col] = ordinal_encoder.fit_transform(data[[col]])

    return data 

clean_train = encoding(train, columns)
clean_test = encoding(test, columns)

Analysis

I will perform the following analysis.

1. Does age have any impact on loan approval?
2. Look at the overall loan intent and see if there is any correlation to loan_status
3. Explore if there is any relationship between a person's income and any previous default on file, i.e. does a higher income mean less default and Vice versa?

1. Does age have any impact on loan approval?

train_analysis = train.copy()

bins = [15, 24, 34, 44, 54, 64, 100]
labels = ["16-24", "25-34", "35-44", "45-54", "55-64", "65+"]
train_analysis["age_bin"] = pd.cut(train_analysis["person_age"], bins = bins, labels = labels)

plot_data = train_analysis.groupby(['age_bin','loan_status']).size().reset_index(name='count')

plt.figure(figsize=(12, 6))
sns.barplot(data = plot_data, x='age_bin', y='count', hue='loan_status')

plt.title('Loan Status Counts by Age Group')
plt.xlabel('Age Group')
plt.ylabel('Count of Loan Status')
plt.legend(title='Loan Status')
plt.show()

Observations

1. The majority of loan applicants are below the age of 35.
2. They are also the applicants with the most loan defaults on file

2. Look at the overall loan intent and see if there is any correlation to loan_status

plot_data = train_analysis.groupby(['loan_intent','loan_status']).size().reset_index(name='count')

plt.figure(figsize=(12, 6))
sns.barplot(data = plot_data, x='loan_intent', y='count', hue='loan_status')

plt.title('Loan Status Counts by Loan Intent')
plt.xlabel('Loan Intent')
plt.ylabel('Count of Loan Status')
plt.legend(title='Loan Status')
plt.show()

Observations

1. The loan intent is just about evenly distributed with home improvement been the least reason and education been the highest.
2. Medical and debt consolidation had the highest approval rate which is a bit surprising, I was expecting education or home improvement to have the highest rate

3. Explore if there is any relationship between a person's income, age group and any previous default on file, i.e. does a higher income mean less default and Vice versa?

plt.figure(figsize = (12,6))

sns.lineplot(train_analysis, x = 'age_bin', y = 'person_income', hue = 'cb_person_default_on_file' )
plt.show()

Observations

1. The non-default rate shows stability across all age groups, even with a salary increase.
2. However, the default rate does spike from the age of 64 and older, despite the increase in salary, which I find interesting.

3. Feature Engineering

Given limited domain knowledge, only two new features will created from existing features to try to improve model performance

1. Debt-to-Income Ratio: This represents the ratio of the loan amount to the person's income. A higher DTI could indicate a higher risk of default.
2. Interest Rate Category: To classify interest rates into categories with 1 = low, 2 = medium, and 3 - high.

#feature engineering
def feature_eng(data):

    data['debt_to_income_ratio'] = data['loan_amnt'] / data['person_income']
    data['interest_rate_category'] = pd.cut(data['loan_int_rate'], bins=[0, 8, 12, 24], labels=[1, 2, 3])

    return data

clean_train = feature_eng(clean_train)
clean_test = feature_eng(clean_test)

# Convert specified columns to categorical type for train
clean_train['loan_grade'] = clean_train['loan_grade'].apply(lambda x: int(x) if x.is_integer() else x).astype('category')
clean_train['person_home_ownership'] = clean_train['person_home_ownership'].apply(lambda x: int(x) if x.is_integer() else x).astype('category')
clean_train['cb_person_default_on_file'] = clean_train['cb_person_default_on_file'].apply(lambda x: int(x) if x.is_integer() else x).astype('category')


# Convert specified columns to categorical type for test
clean_test['loan_grade'] = clean_test['loan_grade'].apply(lambda x: int(x) if x.is_integer() else x).astype('category')
clean_test['person_home_ownership'] = clean_test['person_home_ownership'].apply(lambda x: int(x) if x.is_integer() else x).astype('category')
clean_test['cb_person_default_on_file'] = clean_test['cb_person_default_on_file'].apply(lambda x: int(x) if x.is_integer() else x).astype('category')

Model Training

X =  clean_train[clean_train['person_age'] < 100].drop(columns = ['id', 'loan_status'])
y = clean_train[clean_train['person_age'] < 100]['loan_status']

ids = clean_test['id']
clean_test_data = clean_test.drop(columns = ['id'])

count = y.value_counts()
class_weights = [(round((count.min()/count.sum())*100, 0))/ 10 , (round((count.max()/count.sum())*100, 0))/10]
cat_features = ['interest_rate_category', 'person_home_ownership', 'loan_grade', 'cb_person_default_on_file']
best_params = {'iterations': 529, 'depth': 3, 'learning_rate': 0.23354815782201432, 'l2_leaf_reg': 4.401132273867116,
               'bagging_temperature': 0.6457881123448875, 'random_strength': 8.71241029080795, 'border_count': 243}

model = CatBoostClassifier(**best_params, cat_features = cat_features, random_state = 42, eval_metric = 'AUC', class_weights = class_weights)
model.fit(X,y)

pred = model.predict_proba(clean_test_data)[:,1]
output = pd.DataFrame({'id': ids, 'loan_status': pred})
output.to_csv('submission2.csv', index=False)

X =  clean_train[clean_train['person_age'] < 100].drop(columns = ['id', 'loan_status'])
y = clean_train[clean_train['person_age'] < 100]['loan_status']

ids = clean_test['id']
clean_test_data = clean_test.drop(columns = ['id'])

count = y.value_counts()
scale_pos_weight = count.max() / count.min()

model = lgb.LGBMClassifier(random_state = 42, objective = 'binary', metric = 'auc', scale_pos_weight = scale_pos_weight)
model.fit(X,y)

pred = pred = model.predict_proba(clean_test_data)[:,1]
output = pd.DataFrame({'id': ids, 'loan_status': pred})
output.to_csv('submission.csv', index=False)

Results

1. Cat Result on Kaggle = 0.96273
2. Lightgbm Result on Kaggle = 0.95845

Conclusion

CatBoost gives a better a result on unseen data then lightgbm