Machine Learning-Based Credit Risk Assessment for Predicting Loan Defaulters in Ethiopian Banking industry

doi:10.21203/rs.3.rs-3800266/v1

Download PDF

Research Article

Machine Learning-Based Credit Risk Assessment for Predicting Loan Defaulters in Ethiopian Banking industry

https://doi.org/10.21203/rs.3.rs-3800266/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Machine Learning is an AI technique, empowers organizations globally to gain insights from their data and achieve success. This study focuses on improving risk management of banking in Ethiopia by employing machine learning techniques to detect and predict loan defaulters, ultimately mitigating bad loans. The research aims to distinguish borrowers who repay loans promptly from those who don't, forecast potential defaults, and assess the creditworthiness of prospective customers without overhauling existing systems and data. To showcase the study's results, classification models are constructed to reveal significant patterns among customer attributes. The primary objective is to apply cutting-edge machine learning algorithms to classify bank loan defaulters. Several machine learning algorithms, including Random Forest, Decision Tree, Gradient Boosting (GB), XGBoost, and Multi-Layer Perceptron (MLP), are employed in the process. Various data split ratios, such as 70/30, 60/40, and 90/10, are explored, alongside the 80/20 data split, to build classification models for each classifier algorithm. Based on the experiment's outcomes, the 70/30 split is selected for its superior accuracy. The study evaluates the performance of these algorithms using metrics like accuracy, precision, recall, and F1-score. The results indicate that the Gradient Boosting classification technique surpasses other algorithms, achieving a training accuracy of 98.7% and a testing accuracy of 97.8%.

Theoretical Computer Science

Classification

Loan Defaulter Prediction

Machine Learning

Artificial intelligence plays a significant role in the area of banking industry[1]. In an increasingly global and competitive banking landscape, financial institutions must adopt creative and proactive strategies to gain and retain market share. The banking sector, offering a diverse range of services to customers, is fundamentally dependent on its clientele. Without customers, the banking sector would cease to exist, making customer satisfaction a top priority. Understanding and serving their customers effectively is vital for the sector's survival and success.

As we look ahead, the banking industry will increasingly rely on knowledge as a key asset rather than just financial resources. Information systems, particularly in the form of decision support systems (DSS), play a crucial role in aiding management to make informed decisions, allocate resources efficiently, remain competitive, and enhance customer satisfaction.

Machine learning, a powerful technology, is the process of extracting meaningful patterns and insights from various data sources, including databases, documents, and the internet[2]. It leverages methods for pattern recognition and statistical analysis. In the realm of data mining, several challenges need to be addressed, such as dealing with noisy data, missing values, data dynamics, data sparsity, data relevance, algorithmic efficiency, and the sheer size and complexity of the data.

The data available to banking institutions is often vast and noisy, characterized by inaccuracies and complex structures. Machine learning, with its ability to handle large-scale and complex datasets, offers a solution where traditional statistical approaches fall short. It has become an invaluable tool for analyzing extensive datasets efficiently.

The proliferation of database management systems has further facilitated the collection and storage of vast quantities of data. However, the mere retrieval of information is no longer sufficient for informed decision-making. Extracting and making data readily available is essential.

The knowledge and insights extracted from data have a wide range of applications, including market analysis, fraud detection, customer retention, production control, and scientific exploration. Machine learning empowers the banking sector to focus on critical information within the data collected about the behavior of their current and potential creditors[3]. It plays a pivotal role in predicting creditor status, loan defaults, medical coverage, and payment patterns.

The banking industry stands to gain significant competitive advantages by successfully implementing machine learning. Areas where machine learning can be applied within the banking sector include identifying risk factors that impact profits, claims, and losses; creditor-level analysis; marketing and sales analysis; product development; financial analysis; and estimation.

In this rapidly evolving landscape, the application of machine learning is poised to revolutionize the way banks operate and serve their customers, ultimately driving growth and efficiency in the industry.

In the field of banking and financial services, the application of data mining techniques for predicting loan defaults has become increasingly important. This approach plays a pivotal role in the development of robust credit scoring models and the automation of lending processes. It allows for the exploration of the intricate relationships between borrower characteristics (input) and loan default status (output)[4]. One study conducted by M.J. Xavier, Sundaramurthy, P.K. Viswanathan, and G. Balasubramanian demonstrates the feasibility of this approach using a dataset from a finance company that provides small loans in rural areas. The study employs factor analysis to generate input variables for a neural network algorithm, resulting in a significant enhancement in predictive accuracy[5].

Neural networks (NN) are a potent tool for forecasting outcomes based on input variables. They offer substantial potential for application in loan default prediction and the automation of lending through credit scoring models. However, the integration of factor analysis plays a crucial role in simplifying the complexity by reducing the number of variables to a manageable dimension. When these factor values are utilized as inputs for the neural network, a remarkable improvement in predictive accuracy is achieved. It is worth noting that factor analysis alone does not achieve a 98 percent prediction accuracy, but when combined with neural networks, this level of accuracy can be attained, as demonstrated in the study using real data from a financial services company. The future may see a combination of linear and non-linear techniques in predictive modeling[6].

The strength of this research lies in the careful selection of the neural network algorithm. Neural networks are a suitable choice for predicting bank loan defaulters due to their performance characteristics. In the study, neural network classification is performed using multilayer perceptrons in Weka. The advantage of neural networks for prediction is their ability to learn from examples and discover hidden and strongly non-linear dependencies, even in the presence of significant noise in the training dataset.

This research also has implications for the field of auditing. The study suggests that other data mining techniques, such as classification mining for risk assessment and addressing customer churn issues, warrant further exploration to develop models that can benefit data owners.

Furthermore, research in the area of credit risk assessment using data mining techniques has been conducted in Ethiopia. These studies have shown that data mining technology can effectively address credit risk assessment challenges for banks. For instance, Askale conducted research on the application of data mining technology to support loan disbursement activities at Dashen Bank S.C. in Ethiopia. The research aimed to develop a model that would aid in the loan decision-making process and reduce the high default rate. The study focused on creating a classification model for customer repayment behaviors (Regular, Loss, or Default) to support credit risk assessment. The resulting predictive model assists in determining whether a potential borrower is likely to default, thereby guiding credit decision processes. This research primarily relied on the collection of data, including credit reports and loan approval data, related to individual borrowers of the bank to extract predictor features for the dependent class known as repayment behaviors. These features included security types (e.g., building, vehicle, personal guarantee) and collection-related attributes, such as the term of repayment and the loan disbursement date, as inputs for model development.

Predicting bank loan defaulters using machine learning involves several key steps. Firstly, data collection is essential, gathering historical customer information, including financial, employment, and credit data. Then, data preprocessing tasks such as data cleaning, feature engineering, and handling missing values are performed. Next, a suitable machine learning algorithm is chosen, and the dataset is split into training and testing sets for model training and evaluation. Feature selection and model tuning are critical to optimize performance. Techniques Random Forest, Decision Tree, Gradient Boosting (GB), XGBoost, and Multi-Layer Perceptron (MLP), were employed. Evaluation metrics such as accuracy, precision, recall, and F1-score are used to assess the model's performance. Regular monitoring and updating of the model are necessary to adapt to changing customer behaviors and economic conditions. The researcher followed the experimental research design methodology[7][8].

In the stydy the researcher used 12815 Dataset within 13 columns. In the analysis of the four different machine learning models applied to a classification task, namely Random Forest, Decision Tree Classifier, XGBoost Algorithm, and MLP (Multi-Layer Perceptron), several key performance metrics were evaluated.

Random Forest exhibited impressive training accuracy at 98.0%, although the testing accuracy was slightly lower at 97.6%. This model displayed excellent precision (0.98) for identifying defaulters ("Bad" class), but its recall (0.94) indicated that it missed some defaulters, resulting in a balanced F1 score of 0.96. However, it struggled when classifying non-defaulters ("Good" class), with a low recall of 0.00, which adversely impacted precision and the F1 score (0.98).

The Decision Tree Classifier achieved a training accuracy of 95.0% and a testing accuracy of 94.0%. It demonstrated strong precision (0.97) for defaulters and a decent recall (0.93), resulting in a balanced F1 score of 0.95. Similar to the Random Forest, it faced challenges with the "Good" class, where it had a low recall (0.00) and subsequent decreases in precision and the F1 score (0.97). XGBoost Algorithm outperformed the other models in terms of testing accuracy at 98.0%. It excelled in precision (0.98), recall (0.98), and F1 score (0.98) for both defaulters and non-defaulters, demonstrating its ability to identify both classes effectively.

The MLP model achieved a training and testing accuracy of 91.2%. It displayed good precision (0.96) for defaulters but had a lower recall (0.87), resulting in a reasonably balanced F1 score of 0.91. Impressively, the model correctly classified non-defaulters with good precision (0.88), high recall (0.97), and a high F1 score (0.92).

In summary, while the Random Forest and Decision Tree models exhibited strong performance in identifying defaulters but struggled with non-defaulters, the XGBoost Algorithm achieved high accuracy and balanced performance for both classes. The MLP model, with a slightly lower accuracy, demonstrated good capabilities in both defaulters and non-defaulters, striking a balance between precision and recall as shown in Table below.

Table 1

Result comparison of the four Machine Learning algorithm
Algorithm	Accuracy (Training)	Accuracy (Testing)	Precision (Bad)	Recall (Bad)	F1 Score (Bad)	Precision (Good)	Recall (Good)	F1 Score (Good)
Random Forest	98.0%	97.6%	0.98	0.94	0.96	0.94	0.00	0.98
Decision Tree Classifier	95.0%	94.0%	0.97	0.93	0.95	0.93	0.00	0.97
XGBoost Algorithm	97.8%	98.0%	0.98	0.98	0.98	0.98	0.00	0.98
MLP	91.2%	91.2%	0.96	0.87	0.91	0.88	0.97	0.92

As we illustrated from Fig. 2 our model is XGBoost is outperforming the remaining models. The accuracy of the model is 98%.

This study explores the application of Machine Learning technology to predict loan default patterns at Awash Bank by developing a classification model that categorizes contracts as high-risk or low-risk. The research adopts an experimental research approach, focusing on data preparation and modeling. This process involves several key steps, including data understanding, data selection, data pre-processing, data cleaning, data transformation, data mining, interpretation, and result validation.

The dataset, comprising a total of 18,308 records, undergoes rigorous preprocessing efforts. Subsequently, the data is split into training and testing datasets to develop a classification model. Various machine learning algorithms, including Random Forest, Decision Tree, XGBoost, and MLP, are employed to conduct experiments on this prepared dataset. The results reveal that the XGBoost algorithm yields impressive classification accuracy, with 97.8% accuracy in the training dataset and 97.4% accuracy in the testing dataset. These findings provide valuable insights into the characteristics of loan customer behavior, especially in identifying loan defaulters. Domain experts find the generated rules for defaulter and non-defaulter contracts to be acceptable.

In conclusion, this research suggests that the banking sector can benefit from data mining applications, with XGBoost proving to be an effective tool for understanding the loan customer behavior, particularly in predicting loan default patterns.

“(11) APPLICATIONS OF IMAGE PROCESSING IN MACHINE LEARNING |
“1-s2.0-S1877050920315830-main.pdf.” .
Y. Lu, Y. Wang, Y. Zhang, Y. Lu, X. Yu, and Y. Wang, “ScienceDirect ScienceDirect ScienceDirect ScienceDirect Risk Model Based on Machine Assessment of Credit Risk Model Based on Machine Learning Learning A Comparative Assessment of Credit Risk Model Based on Machine A Comparative Assessment Credit Risk loan Model Based on Machine Learning —— a case of study of bank data —— a case study of bank loan data Learning case loan data,” Procedia Comput. Sci., vol. 174, pp. 141–149, 2020, doi: 10.1016/j.procs.2020.06.069.
D. B. Hailu, “St . Mary ’ s University Faculty of Informatics St . Mary ’ s University Faculty of Informatics,” no. June, 2019.
“data-08-00169.pdf.” .
L. Review, “Machine Learning for Credit Risk Prediction : A Systematic Literature Review,” pp. 1–17, 2023.
K. McDonough, “Experimental research methods,” Routledge Handb. Instr. Second Lang. Acquis., no. May, pp. 562–576, 2017, doi: 10.4324/9781315676968.
E. Design, “Experimental Research Designs,” 2008.

The authors declare no competing interests.

Download PDF

Version 1

posted

You are reading this latest preprint version

Machine Learning-Based Credit Risk Assessment for Predicting Loan Defaulters in Ethiopian Banking industry

Status:

Version 1

Abstract

Figures

I. Introduction

II. Related work

III. Methodology

IV. Result and Discussion

V. Conclusion

References

Additional Declarations

Status:

Version 1