An Approach to Credit Risk Assessment through Multimodal Data Fusion

Yash Khandagale, Manish Kapal, Arya Shinde, Sanika Wani

BE Sem VII- 2024-25

Guide
Prof.Avinash Gondal

Department Of Computer Engineering

Watumull Institute Of Electronics And Technology
 01 Introduction

02 Motivation

03 Literature
Review
04 Objective

05 Problem
Statement
Conte 06 Proposed
Architecture
nts 07 Dataset and Technology

08 Proposed Methodology

09 Gantt
Chart
10 Conclusion

1
1 Reference
 Introduction

• Credit risk modeling is a critical function for financial institutions, helping them assess the
likelihood of loan applicants defaulting.
• The ability to predict an applicant’s risk accurately helps banks and other lending
institutions minimize non-performing loans (NPLs) and optimize loan approval decisions.
• This project focuses on classifying loan applicants into risk categories (P1-P4) using
advanced machine learning (ML) and deep learning (DL) techniques, based on internal bank
data and external credit data (e.g., CIBIL).

Department Of Computer Engineering,Watumull

Institute,
 Motivation

The financial industry faces significant risks due to loan defaults. Misclassification of
credit risk can lead to substantial financial losses, as high-risk customers may be
approved for loans while low-risk customers are denied. Traditional credit scoring models
often struggle with complex, non-linear data and do not account for dynamic changes in
borrower behavior. By applying machine learning and deep learning models, we can
improve the accuracy and robustness of credit risk predictions, helping banks make better
decisions, minimize defaults, and optimize their lending processes.

Department Of Computer Engineering,Watumull

Institute,
 Literature Review

• Credit risk prediction has evolved from traditional models like Logistic
Regression to more advanced methods such as XGBoost and Random Forest,
which offer better accuracy but still face challenges with class imbalance and
transparency.
• Techniques like SMOTE improve data balance, and tools like SHAP value
enhance model interpretability. While deep learning shows potential, its
complexity limits practical use.

Department Of Computer Engineering,Watumull

Institute,
Sr. No. Title Abstract Limitations Future Scope
1 Credit Risk Prediction using This paper presents a model that combines Genetic Algorithms for feature Computational complexity and Enhance scalability and optimize Genetic
Genetic Algorithm and Stacking selection with stacking ensemble methods, achieving high accuracy for credit scalability issues when applied to Algorithm efficiency for larger datasets
Ensemble risk prediction by optimizing feature selection and model integration. larger datasets. and real-time applications.

2 Explainable AI Using H2O AutoML This paper focuses on applying H2O’s AutoML for model tuning and uses Lacks detailed performance metrics Incorporate class imbalance techniques
and Robustness Check in Credit SHAP values for interpretability. It emphasizes the need for transparency in and does not fully address class and provide a more comprehensive
Risk Management credit risk models to ensure regulatory compliance while maintaining high imbalance. evaluation with detailed metrics.
accuracy.

3 Customer Credit Risk: Application A comparative analysis of machine learning (SVM, Random Forest) and deep SVM struggles with large datasets, and Explore hybrid models combining ML
and Evaluation of Machine learning models (DNN, CNN) for credit risk prediction. SVM and XGBoost are complex models may suffer from and DL techniques to improve
Learning and Deep Learning identified as the top-performing algorithms in terms of accuracy. overfitting without proper tuning. performance on large and imbalanced
Models datasets.

4 Optimizing Loan Approval This paper uses ensemble methods to optimize loan approval decisions, Challenges in deploying stacking Test deep learning methods and improve
Decisions providing insights on different machine learning algorithms and their ensembles in real-time systems, with the deployment of ensemble models in
performance in predicting credit risk categories. limited exploration of deep learning real-time credit risk environments.
models.

5 Financial Risk Prediction Using Provides an extensive comparison of various machine learning algorithms Insufficient handling of class imbalance Integrate interpretability tools (like
Machine Learning (Logistic Regression, Decision Trees, XGBoost) for financial risk prediction. It and lack of model interpretability. SHAP) and better address class
focuses on practical implementation in the financial industry. imbalance through resampling
techniques.

6 Novel Approach to Optimize Proposes an enhanced SVM model for credit risk prediction, achieving high Limited dataset size raises concerns Apply the model to larger datasets and
Credit Risk Prediction accuracy (96.92%) and targeted improvements to the algorithm. about generalizability, and struggles enhance its handling of class imbalance
with imbalanced data were noted. using techniques like SMOTE.

7 An Automatic Deep This paper introduces a Deep-Q Network (reinforcement learning) approach Low accuracy (~10%) due to class Improve accuracy and scalability by
Reinforcement Learning Based for credit scoring. It focuses on adaptive learning and real-time feedback for imbalance and complex tuning tuning the Deep-Q network and
Credit Scoring Model using Deep- improved credit risk classification. requirements, limiting practical incorporating data balancing techniques.
Q application.

8 Credit Risk Assessment Explores the use of HistGradientBoosting for credit risk assessment, achieving Needs further validation in real-world Test model robustness in real-world
high accuracy (89.08%) and precision (94.54%) while efficiently handling large scenarios, with a risk of overfitting if environments and apply stronger
datasets. not properly regularized. regularization to avoid overfitting.
 Objective

The objective of this project is to develop an accurate, scalable, and interpretable credit
risk prediction model that categorizes loan applicants into four risk categories: P1 (lowest
risk) to P4 (highest risk).
The model will leverage machine learning and deep learning techniques, combining both
internal bank data (loan accounts, payment histories) and external credit data (CIBIL
scores).
The primary goals are:
• To minimize financial losses by accurately predicting high-risk applicants.
• To ensure model interpretability for regulatory compliance.
• To balance model performance and computation time, making it scalable for real-world
deployment.

Department Of Computer Engineering,Watumull

Institute,
 Problem Statement
Financial institutions face challenges in predicting the risk of loan default accurately, as
traditional models struggle to handle large volumes of data, class imbalance, and complex
relationships between borrower characteristics.
Inaccurate credit risk classification can lead to either loan defaults or missed lending
opportunities.
Therefore, the problem is to build a robust machine learning model that:
1. Predicts the likelihood of default and classifies applicants into risk categories.
2. Handles imbalanced datasets, ensuring minority high-risk categories are accurately
identified.
3. Provides model transparency and interpretability for regulatory compliance.

Department Of Computer Engineering,Watumull

Institute,
 Proposed Architecture
The architecture for this project consists of several key stages:
1. Data Collection: Collect internal bank data (loan, account history,
etc.) and external data (CIBIL).
2. Data Preprocessing:
• Handle missing values.
• Encode categorical features and scale numerical ones.
• Apply feature selection using Genetic Algorithms or Principal
Component Analysis (PCA).
3. Data Balancing: Address class imbalance using SMOTE.
4. Model Training:
• Train multiple models: Decision Tree, Random Forest, and
XGBoost.
• Perform cross-validation and hyperparameter tuning.
5. Evaluation and Explainability:
• Evaluate the model using metrics like Accuracy, Precision,
Recall, and AUC-ROC.
• Implement explainability using SHAP values.
6. Deployment: Deploy the best-performing model for real-time loan
approval predictions.

Department Of Computer Engineering, Watumull

Institute.
Dataset and Technology

Department Of Computer Engineering,Watumull

Institute,
 Dataset and Technology
Technology:-
• Programming Language: Python
• Libraries/Frameworks: -
• Scikit-learn: For machine learning algorithms.
• XGBoost: For implementing gradient boosting models.
• Pandas & NumPy: For data preprocessing and manipulation.
• H2O AutoML: For hyperparameter tuning and model selection.
• SHAP: For model interpretability
• Hardware:
• Intel i7 processor (or equivalent).
• 16GB RAM.
• Optional: GPU (for faster training of deep learning models).
Department Of Computer Engineering,Watumull
Institute,
 Proposed Methodology
1. Data Preprocessing:
• Handle missing values through imputation.
• Encode categorical variables using One-Hot Encoding.
• Scale numerical features to ensure uniformity across different ranges (e.g., loan
amounts, credit scores).
2. Feature Selection:
• Use Genetic Algorithms (GA) to reduce dimensionality and identify key predictive
features.
3. Model Selection:
• Train multiple machine learning models:
• Decision Tree: For simplicity and interpretability.
• Random Forest: For improved accuracy and robustness.
• XGBoost: For high accuracy and efficiency in handling imbalanced data.
Department Of Computer Engineering,Watumull
Institute,
 Proposed Methodology
4. Evaluation:
• Use metrics such as Accuracy, Precision, Recall, F1-Score, and AUC-ROC to evaluate
model performance.
• Perform cross-validation to ensure the model generalizes well to unseen data.
5. Explainability:
• Implement SHAP values to provide interpretability for model predictions, ensuring
compliance with financial regulations.
6. Deployment:
• Deploy the best-performing model to the bank’s loan approval system for real-time
predictions.

Department Of Computer Engineering,Watumull

Institute,
 Gantt Chart

Department Of Computer Engineering,Watumull

Institute,
 Conclusion

• This project presents a comprehensive approach to credit risk modeling, leveraging both
internal and external datasets. By applying machine learning algorithms like XGBoost and
ensemble methods like Random Forest, the project aims to provide a robust solution for
classifying loan applicants into different risk categories.
• The model will not only improve prediction accuracy but will also provide interpretability
using SHAP, ensuring that the system meets regulatory requirements and delivers clear insights
into the decision-making process.
• The expected result is a highly accurate model (85-90% accuracy) that balances performance
with transparency, allowing financial institutions to make informed, data-driven lending
decisions while minimizing risk.

Department Of Computer Engineering, Watumull

Institute.
 References
1. Barthe, S., Sanyal, S., Biswas, S. K., & Purkayastha, B. (2023). "Credit Risk Prediction
Using Genetic Algorithm and Stacking Ensemble." Journal of Financial Analytics.
2. Sikha, S., Vijayakumar, A., & Yemmanuru, P. K. (2023). "Explainable AI Using H2O
AutoML in Credit Risk Management." International Journal of Data Science.
3. Yeboah, J., & Nti, I. K. (2024). "Customer Credit Risk: Application of Machine Learning
and Deep Learning Models." Financial Risk Management Journal.
4. Vohra, N., & Goyal, D. (2023). "Optimizing Loan Approval Decisions with Ensemble
Learning." Financial Technology & Decision Systems.
5. Dhawan, K., & Jayakumar, N. (2023). "A Novel Approach to Optimize Credit Risk Using
Enhanced SVM." Journal of Risk and Model Optimization.

Department Of Computer Engineering,Watumull

Institute,
 THANKYOU

Department Of Computer Engineering,Watumull

Institute,