International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 04 | Apr 2025
p-ISSN: 2395-0072
www.irjet.net
Enhancing Loan Default Prediction and Fraud Detection with Ensemble Learning Adwait Mandge1, Rohan Fatehchandka2, Kunal Goudani3, Tanaya Shelke4, Prof. Pramila M. Chawan5 1,2,3,41B. Tech Student, Dept of Computer Engineering, and IT, VJTI College, Mumbai, Maharashtra, India 5Associate Professor, Dept of Computer Engineering, and IT, VJTI College, Mumbai, Maharashtra, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Loan risk assessment is crucial for financial
aggregates the outcomes to increase prediction accuracy. Features are randomly selected at every node(feature bagging), and each tree in the forest is trained on an arbitrary subclass of the data (bootstrapping). Random Forest reduces variation and overfitting and averages the outcomes of these trees, which enhances generalization. The advantages of this algorithm are handling large datasets, reducing overfitting, and maintaining good accuracy even in the presence of missing data.
security. This study combines ensemble learning for loan default prediction and time series analysis for ghost borrower detection. Using PyCaret, we optimize model selection to identify high-risk borrowers. Additionally, ARIMA, LSTMs, and anomaly detection techniques analyze transaction patterns to flag fraudulent behaviors like sudden withdrawals and post-loan inactivity. By integrating predictive modeling and anomaly detection, we enhance early fraud detection. This approach provides financial institutions with a comprehensive risk management framework, improving decision-making and reducing potential losses.
XGBoost (Extreme Gradient Boosting) is a scalable and extremely effective gradient boosting implementation. It constructs decision trees sequentially, aiming to fix the mistakes caused by preceding trees with each new tree. XGBoost uses gradient descent to minimize the overall loss function in order to optimize the model. Performance and speed are well-known for XGBoost, particularly with tabular or structured data.The advantages of this algorithm are its speed, scalability, handling of missing data, and regularization to minimize overfitting.
Key Words: Machine Learning, Deep Learning, Ensemble, Loan Default Prediction, Ghost Borrower Detection, TCN
1.INTRODUCTION Accurate loan default prediction is vital for financial institutions to mitigate risks. Traditional credit assessment methods often overlook hidden patterns, making machine learning a powerful alternative. This study utilizes PyCaret to compare ensemble techniques like bagging, boosting, and stacking for loan default prediction. By analyzing borrower demographics, financial history, and loan details, we evaluate model performance using accuracy, precision, recall, and F1-score, identifying the most effective approach for credit risk assessment.
2.2 Deep Learning Algorithms Neural Networks replicate the composition of the human brain. Neural networks comprise multiple bands of neurons connected by edges with weights that are updated during training. Neural networks are widely used in deep learning processes as they can interpret occurrences of perplexing repetitive sequences. The advantages of this algorithm are its flexibility, it can interpret information from huge datasets, and its potential to replicate lateral relationships.
To enhance fraud detection, we address ghost borrowers—fraudsters who manipulate financial records to evade repayment. We integrate time series analysis using TCN (Temporal Convolution Networks) to identify suspicious transaction patterns. By combining predictive analytics with fraud detection, our study provides actionable insights to improve lending decisions and reduce financial losses.
Multi-Layer Perceptron (MLP) is used in deep learning tasks. MLPs consist of multiple layers which are interconnected to each other, every neuron of one layer acts as an input to the neuron in the next layer. MLPs are used in classification, regression, and serve as the foundation for more intrinsic neural networks. Its capability to handle both regression and classification problems, and its role as a foundation for more advanced neural networks is advantageous.
2. LITERATURE REVIEW 2.1 Machine Learning Algorithms
Temporal Convolutional Networks (TCNs) are a type of deep learning architecture designed for sequence modeling tasks, offering an alternative to recurrent neural networks (RNNs) like LSTMs and GRUs. TCNs leverage 1D
Random Forest learning technique is used to solve various problems of regression and classification problems. It builds several trees during training and
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 157