Loan Default Risk Prediction

1. Introduction

In this project, we build a machine learning pipeline to predict loan default risk using financial and demographic data. We apply EDA, preprocessing, Logistic Regression, and Random Forest models, and evaluate them using accuracy, precision, recall, and ROC-AUC.

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler, LabelEncoder
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score, confusion_matrix, classification_report, roc_auc_score, roc_curve

df = pd.read_csv("loan_default.csv")
print("Shape of dataset:", df.shape)
df.head()

Shape of dataset: (255347, 18)

	LoanID	Age	Income	LoanAmount	CreditScore	MonthsEmployed	NumCreditLines	InterestRate	LoanTerm	DTIRatio	Education	EmploymentType	MaritalStatus	HasMortgage	HasDependents	LoanPurpose	HasCoSigner	Default
0	I38PQUQS96	56	85994	50587	520	80	4	15.23	36	0.44	Bachelor's	Full-time	Divorced	Yes	Yes	Other	Yes	0
1	HPSK72WA7R	69	50432	124440	458	15	1	4.81	60	0.68	Master's	Full-time	Married	No	No	Other	Yes	0
2	C1OZ6DPJ8Y	46	84208	129188	451	26	3	21.17	24	0.31	Master's	Unemployed	Divorced	Yes	Yes	Auto	No	1
3	V2KKSFM3UN	32	31713	44799	743	0	3	7.07	24	0.23	High School	Full-time	Married	No	No	Business	No	0
4	EY08JDHTZP	60	20437	9139	633	8	4	6.51	48	0.73	Bachelor's	Unemployed	Divorced	No	Yes	Auto	No	0

Exploratory Data Analysis (EDA)

1. Target Variable Distribution (Default)

df['Default'].value_counts()
sns.countplot(x='Default', data=df)
plt.title("Class Distribution (Default vs Non-Default)")
plt.show()

2. Numeric Feature Distributions

numeric_cols = ['Age', 'Income', 'LoanAmount', 'CreditScore', 'MonthsEmployed', 'NumCreditLines', 'InterestRate', 'LoanTerm', 'DTIRatio']

df[numeric_cols].hist(figsize=(15,10), bins=30)
plt.suptitle("Numeric Feature Distributions")
plt.show()

3. Boxplots (Feature vs Default)

sns.boxplot(x='Default', y='Income', data=df)
plt.title("Income vs Default")
plt.show()

sns.boxplot(x='Default', y='CreditScore', data=df)
plt.title("Credit Score vs Default")
plt.show()

4. Categorical Features vs Default

categorical_cols = ['Education', 'EmploymentType', 'MaritalStatus', 'HasMortgage', 'HasDependents', 'LoanPurpose', 'HasCoSigner']

for col in categorical_cols:
    plt.figure(figsize=(6,4))
    sns.countplot(x=col, hue='Default', data=df)
    plt.title(f"{col} vs Default")
    plt.xticks(rotation=45)
    plt.show()

5. Preprocessing

We prepare the dataset by:

• Splitting into features (X) and target (y)
• Scaling numerical variables
• One-hot encoding categorical variables
• Creating a preprocessing pipeline

from sklearn.preprocessing import OneHotEncoder
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline

df = df.drop(columns=['LoanID'])

X = df.drop('Default', axis=1)
y = df['Default']

categorical_cols = X.select_dtypes(include=['object']).columns
numeric_cols = X.select_dtypes(exclude=['object']).columns

numeric_transformer = StandardScaler()
categorical_transformer = OneHotEncoder(handle_unknown='ignore')
preprocessor = ColumnTransformer(
    transformers=[
        ('num', numeric_transformer, numeric_cols),
        ('cat', categorical_transformer, categorical_cols)
    ])

6. Train-Test Split

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)

5. Logistic Regression (with Class Weight Balancing)

Logistic Regression serves as a simple baseline.
We add class_weight="balanced" to handle the imbalance between defaults and non-defaults.

log_reg = Pipeline(steps=[('preprocessor', preprocessor),
                          ('classifier', LogisticRegression(max_iter=1000, class_weight='balanced'))])

log_reg.fit(X_train, y_train)

y_pred_lr = log_reg.predict(X_test)
print("Logistic Regression Accuracy:", accuracy_score(y_test, y_pred_lr))
print(classification_report(y_test, y_pred_lr))

Logistic Regression Accuracy: 0.6764440963383591
              precision    recall  f1-score   support

           0       0.94      0.67      0.79     45139
           1       0.22      0.70      0.33      5931

    accuracy                           0.68     51070
   macro avg       0.58      0.69      0.56     51070
weighted avg       0.86      0.68      0.73     51070

6. Random Forest (with Class Weight Balancing)

Random Forest is an ensemble method that often performs well on tabular data.
We again add class_weight="balanced" to address class imbalance.

rf_model = Pipeline(steps=[('preprocessor', preprocessor),
                           ('classifier', RandomForestClassifier(n_estimators=100, random_state=42, class_weight='balanced'))])

rf_model.fit(X_train, y_train)

y_pred_rf = rf_model.predict(X_test)
print("Random Forest Accuracy:", accuracy_score(y_test, y_pred_rf))
print(classification_report(y_test, y_pred_rf))

Random Forest Accuracy: 0.8847464264734678
              precision    recall  f1-score   support

           0       0.89      1.00      0.94     45139
           1       0.70      0.01      0.03      5931

    accuracy                           0.88     51070
   macro avg       0.79      0.51      0.48     51070
weighted avg       0.86      0.88      0.83     51070

7. ROC Curve & AUC

We plot the ROC curve and calculate the AUC to evaluate the model's ability to discriminate between defaults and non-defaults.

y_pred_proba = rf_model.predict_proba(X_test)[:,1]
fpr, tpr, thresholds = roc_curve(y_test, y_pred_proba)

plt.figure(figsize=(6,4))
plt.plot(fpr, tpr, label=f"AUC = {roc_auc_score(y_test, y_pred_proba):.2f}")
plt.plot([0,1],[0,1],'--',color='gray')
plt.xlabel("False Positive Rate")
plt.ylabel("True Positive Rate")
plt.title("ROC Curve - Random Forest")
plt.legend()
plt.show()

8. Conclusion

• Logistic Regression (with balancing) achieved higher recall for defaults (70%), making it better for detecting risky loans.
• Random Forest achieved higher accuracy (88%) but struggled to identify defaulters due to class imbalance.
• AUC score of 0.73 shows Random Forest has reasonable predictive power.

In real-world credit risk analysis, recall for defaults is more important than overall accuracy, since missing a default is costly.
Future improvements could include SMOTE oversampling, feature selection, and hyperparameter tuning.