This study uses a synthetic, realistically structured dataset comprising 5,000 borrower-level loan records. The dataset was designed to simulate credit-risk characteristics associated with commercial banking practice in Bangladesh. Each observation represents a hypothetical borrower or loan account and includes demographic, financial, loan-related, and repayment-behavior attributes.

The dataset incorporates simulated patterns inspired by financial institutions in Bangladesh, including private commercial banks, such as BRAC Bank, City Bank, and Dutch-Bangla Bank; Islamic banks, such as Islami Bank Bangladesh; and state-owned banks. This design was used to ensure diversity in borrower profiles, institutional characteristics, and credit-risk behavior. However, these

Figure 1. Proposed framework flowchart.

bank names were used as examples for simulation design. No borrower-level records, customer information, loan accounts, or transaction data from these institutions were used. The dataset was constructed using domain knowledge of banking practice and credit-risk variables used in credit scoring. The main variables include monthly income, loan amount, loan tenure, employment status, credit history length, debt-to-income ratio, previous default history, late payment count, collateral value, account balance, bank type, loan type, and regional classification. These variables were selected because they are associated with repayment capacity, credit exposure, and default risk. The target variable is loan default status, where 1 indicates default, and 0 indicates non-default. In this study, “default” is operationally defined as a simulated borrower failing to meet repayment obligations based on the generated latent credit-risk profile. The default variable was not directly hard-coded from one or two predictors. Instead, it was generated probabilistically based on multiple borrower-level characteristics, including debt-to-income ratio, prior default history, late-payment count, loan amount, monthly income, credit history length, collateral value, account balance, and employment status. The dataset was generated for research and simulation purposes. It was designed to simulate borrower behavior while preserving data privacy. No confidential borrower information, publicly available customer-level data, or restricted bank records were used. Therefore, the findings should be interpreted as simulation-based evidence demonstrating the feasibility of an explainable machine learning framework for credit risk prediction in Bangladesh.

The synthetic dataset was generated using a fixed random seed of 42 to ensure reproducibility. A total of 5000 borrower-level observations were created, where each observation represents a hypothetical loan applicant. The feature structure was developed using common credit-risk variables and plausible lending characteristics relevant to commercial banking practice in Bangladesh. Continuous variables were generated within realistic bounded ranges. Monthly income was simulated between BDT 10,000 and BDT 120,000, loan amount between BDT 50,000 and BDT 500,000, and loan tenure between 6 and 60 months. The debt-to-income ratio was bounded between 0.10 and 0.90, while the credit history length was generated between 0 and 20 years. Collateral value was simulated between BDT 0 and BDT 800,000, and account balance was generated between BDT 0 and BDT 300,000. Late payment count was generated as a discrete count variable ranging from 0 to 10.

Categorical variables were generated using predefined probability proportions. Bank type was simulated as 60% private commercial banks, 25% Islamic banks, and 15% state-owned banks. Regional classification was simulated as 50% urban, 30% semi-urban, and 20% rural. Employment status was simulated across salaried, self-employed, business, and unemployed borrower groups, with higher simulated risk assigned to unstable employment categories.

The default label was generated probabilistically rather than being directly hard-coded from a small number of predictors. First, a latent credit-risk score was calculated using multiple borrower characteristics, including debt-to-income ratio, previous default history, late payment count, loan amount, monthly income, credit history length, collateral value, account balance, and employment status. Higher debt-to-income ratio, previous default history, higher late payment count, larger loan burden, and unstable employment status increased the latent risk score. In contrast, higher monthly income, longer credit history, stronger collateral value, and higher account balance reduced the latent risk score.

The latent risk score followed the general form:

Risk Score = β₁ + β₁ (Debt-to-Income Ratio) + β₂ (Previous Default) + β₃ (Late Payment Count) + β₄ (Loan Amount) – β₅ (Monthly Income) – β₆ (Credit History Length) – β₇ (Collateral Value) – β₈ (Account Balance) + β₉ (Employment Risk).

The latent score was transformed into a default probability using a logistic function. Finally, the binary default label was sampled from this probability. This procedure preserved uncertainty in borrower repayment behavior and avoided a deterministic relationship between only a few predictors and default status. The final simulated class distribution was approximately 72% non-default and 28% default.

Before model development, the dataset underwent preprocessing to ensure consistency and suitability for machine learning analysis. Missing values were handled using mean imputation for numerical variables and mode imputation for categorical variables. Categorical variables such as employment status, loan type, bank type, and region were converted into numerical form using encoding techniques. Feature scaling was applied using Standard Scaler, particularly for models sensitive to feature magnitude, such as Support Vector Machine and Logistic Regression. The dataset was divided into training and testing subsets using a stratified 80:20 split. Therefore, 4000 observations were used for training and validation, while 1000 observations were reserved as an independent test set. The test set was kept completely unseen during model development. Five-fold cross-validation was applied only to the training set for model selection and hyperparameter tuning. To avoid data leakage, preprocessing steps, feature selection, and hyperparameter tuning were performed inside the cross-validation process.

Imputation, encoding, scaling, and feature selection were fitted only on the training fold and then applied to the corresponding validation fold. After selecting the best hyperparameters, the final model was retrained on the full training set and evaluated once on the independent 1000-record test set. Classification decisions were made using a probability threshold of 0.50. A borrower was classified as default if the predicted probability of default was greater than or equal to 0.50; otherwise, the borrower was classified as non-default. The dataset contained an imbalanced class distribution of approximately 72% non-default and 28% default cases. No oversampling or undersampling technique was applied. However, model performance was evaluated using precision, recall, F1-score, AUC-ROC, and the confusion matrix in addition to accuracy because false negatives are particularly important in credit screening.

Statistical analysis was conducted using the Statistical Package for the Social Sciences (SPSS) to complement the machine learning models and provide a baseline understanding of the dataset. This step aims to examine the underlying statistical properties of the variables before applying advanced predictive techniques.

The analysis includes descriptive statistics, frequency distribution, correlation analysis, and logistic regression. Descriptive statistics are used to summarize key financial variables such as monthly income, loan amount, and debt-to-income ratio. Frequency analysis is performed to examine the distribution of default and non-default cases within the dataset.

Furthermore, correlation analysis is conducted to identify relationships among variables and determine the most influential factors affecting credit risk. Logistic regression is applied as a traditional statistical model to evaluate the significance and impact of predictor variables on the probability of default.

Overall, the SPSS-based statistical analysis provides valuable insights into the structure of the dataset and supports the findings obtained from machine learning models, ensuring the robustness and reliability of the proposed framework.

Feature selection is conducted to enhance model efficiency and minimize overfitting. Initially, a correlation analysis is performed using a heatmap to examine relationships among variables. Highly correlated or redundant features are carefully evaluated and removed where necessary.

Additionally, feature importance scores are computed using tree-based models such as Random Forest and XGBoost to identify the most influential predictors of credit risk to improve model efficiency and reduce overfitting.

Figure 2. Correlation structure heatmap.

The correlation structure among variables is visualized in Figure 2, which highlights the relationships between financial and behavioral features.

To ensure a comprehensive evaluation, multiple machine learning models are implemented and compared.

Logistic Regression is used as a baseline model due to its simplicity and interpretability.Random Forest is applied as an ensemble learning method capable of capturing nonlinear relationships and reducing overfitting.XGBoost is selected as the primary model due to its superior performance in classification tasks and its ability to handle imbalanced datasets effectively.Support Vector Machine (SVM) is included to evaluate classification performance in high-dimensional feature space.

This multi-model framework enables a robust comparison and identification of the most effective model for credit risk prediction ([4]).

To address the black-box nature of machine learning models, this study incorporates explainable artificial intelligence techniques.

The primary explainability approach used is SHAP (Shapley Additive Explanations), which provides both global and local interpretability by quantifying the contribution of each feature to model predictions.

Global interpretation identifies the overall importance of features influencing model decisions.Local interpretation explains individual predictions, enabling case-level analysis of borrower risk.

Additionally, LIME (Local Interpretable Model-agnostic Explanations) is considered as a complementary technique for validating local explanations. SHAP is used to provide consistent and interpretable explanations of model predictions. Explainability techniques are particularly important in banking applications, where model transparency is required for trust, accountability, and regulatory compliance.

Model performance is evaluated by using multiple classification metrics to ensure a comprehensive assessment.

Accuracy evaluates overall prediction correctness;Precision measures the proportion of correct positive predictions;Recall assesses the ability to detect actual defaults;F1-score provides a balance between precision and recall;AUC-ROC evaluates the model’s ability to distinguish between classes.

Additionally, a confusion matrix is used to visualize classification outcomes.

The dataset was divided using a stratified 80:20 train-test split. From the 5000 borrower records, 4000 observations were used for training and validation, while 1000 observations were kept as an independent test set. The test set was not used during model training, validation, or hyperparameter tuning ([2]). Five-fold cross-validation was applied only to the training set. To avoid data leakage, preprocessing steps, including missing-value imputation, categorical encoding, feature scaling, and feature selection, were performed separately within each training fold and then applied to the corresponding validation fold. Hyperparameter tuning was conducted using grid search for Logistic Regression, Random Forest, Support Vector Machine, and XGBoost. For Logistic Regression, the regularization strength parameter C was searched using values of 0.01, 0.1, 1, and 10, with l2 penalty and both liblinear and lbfgs solvers. The final selected Logistic Regression setting was C = 1, l2 penalty, and lbfgs solver.

For Random Forest, the grid search considered 100, 200, and 300 estimators; maximum depth values of None, 5, 10, and 15; minimum sample split values of 2, 5, and 10; and minimum sample leaf values of 1, 2, and 4. The final selected Random Forest configuration used 200 estimators, a maximum depth of 10, a minimum samples split of 2, and a minimum samples leaf of 1. For the Support Vector Machine model, the search range included C values of 0.1, 1, and 10; linear and radial basis function kernels; and gamma values of scale and auto. The final selected SVM setting was C = 10, radial basis function kernel, and gamma = scale. For XGBoost, the grid search considered 100, 200, and 300 estimators; maximum depth values of 3, 5, and 7; learning rates of 0.01, 0.05, and 0.10; subsample values of 0.8 and 1.0; and colsample_bytree values of 0.8 and 1.0. The final selected XGBoost configuration used 200 estimators, a maximum depth of 5, a learning rate of 0.05, a subsample of 0.8, and a colsample_bytree of 0.8. After selecting the best parameter settings, each final model was retrained on the full training set and evaluated once on the independent test set using accuracy, precision, recall, F1-score, AUC-ROC, and confusion matrix. These settings were selected based on average cross-validation performance on the training data. The independent test set was not used during parameter selection. This procedure ensured a fair comparison among the models and reduced the risk of overfitting.

4.1.1. Descriptive Statistics

Table 2. Descriptive statistics of key variables.

Table 2 summarizes the descriptive statistics of key financial variables used in the dataset. The descriptive statistics provide an overview of the key financial characteristics of borrowers included in the dataset. The results indicate that the average monthly income is 45,000, while the mean loan amount is 250,000. The debt-to-income ratio has an average value of 0.42, reflecting a moderate level of financial obligation among borrowers. The observed variation in these variables suggests that the dataset captures diverse borrower profiles, which is essential for developing a reliable credit risk prediction model. The distribution of monthly income is shown in Figure 3, which illustrates the income variability among borrowers.

Figure 3. Histogram of monthly income.

The histogram illustrates the distribution of borrower income across the dataset. The distribution is slightly right-skewed, indicating that the majority of borrowers fall within the low- to middle-income range, while relatively fewer borrowers belong to the higher-income category.

4.1.2. Frequency Analysis

Table 3. Distribution of default status.

Table 3 shows the frequency distribution of default and non-default cases in the dataset. The frequency analysis shows that 72% of borrowers are classified as non-default, while 28% fall into the default category. This indicates that although the majority of borrowers are considered low-risk, a substantial proportion represents potential credit risk, emphasizing the importance of accurate prediction models. The distribution of default and non-default borrowers is presented in Figure 4.

The bar chart presents the distribution of default and non-default borrowers. It clearly shows that non-default cases dominate the dataset; however, the presence of a considerable number of default cases highlights the need for effective credit risk assessment strategies.

4.1.3. Correlation Analysis

Table 4 presents the correlation matrix among the key study variables. The

Figure 4. Default vs non-default distribution.

Table 4. Correlation matrix.

correlation analysis reveals significant relationships among the variables. The debt-to-income ratio exhibits the strongest positive correlation with default (0.60), indicating that a higher financial burden increases the likelihood of default. In contrast, monthly income shows a negative correlation (−0.45), suggesting that higher income reduces credit risk. Loan amount also demonstrates a moderate positive relationship with default, indicating its influence on borrower risk. Figure 5 shows the correlation heatmap among key financial variables, highlighting their relationships with credit default status.

The heatmap visually represents the strength and direction of relationships among variables. It confirms that the debt-to-income ratio is the most influential factor in determining credit risk, followed by loan amount and income.

4.1.4. Logistic Regression Analysis

Table 5 provides the logistic regression model summary including key performance indicators. The logistic regression results indicate that the debt-to-income ratio and previous default history are statistically significant predictors of credit

Figure 5. Feature selection correlation heatmap.

Table 5. Logistic regression model summary.

Table 6. Logistic regression coefficients.

risk. The logistic regression model summary indicates moderate baseline predictive capability. The model achieved a classification accuracy of 0.82 and an AUC-ROC value of 0.84. The Nagelkerke R² value of 0.45 suggests that the model explains a moderate proportion of variation in the simulated default outcome. These results provide a useful statistical benchmark before applying more advanced machine learning models.

Table 6 presents the logistic regression coefficients and their statistical significance. The logistic regression coefficients indicate that the debt-to-income ratio and previous default history are important predictors of simulated credit risk. The positive coefficients for loan amount, debt ratio, and previous default history suggest that increases in these variables are associated with a higher probability of default. Conversely, monthly income has a negative coefficient, indicating that higher income reduces the likelihood of default. These findings align with credit-risk theory and support the importance of financial capacity and repayment history in determining borrower risk. The performance of the logistic regression model is evaluated using the ROC curve shown in Figure 6.

Figure 6.ROC curve of logistic regression model.

The ROC curve evaluates the classification performance of the logistic regression model. The curve demonstrates that the model has acceptable discriminative ability in distinguishing between default and non-default borrowers.

Overall SPSS-based: The SPSS-based statistical analysis confirms that key financial variables, including income level, loan amount, and debt-to-income ratio, play a significant role in determining credit risk. The results obtained from traditional statistical methods are consistent with the findings of the machine learning models, thereby validating the robustness and reliability of the proposed framework.

The proposed models were implemented using the same preprocessing and validation conditions to ensure consistency and fairness in performance comparison. Four machine learning models were considered in this study: Logistic Regression as the baseline model, Random Forest, Support Vector Machine, and XGBoost. Hyperparameter tuning was applied using grid search with five-fold cross-validation on the training set only. After tuning, each final model was retrained on the full training data and evaluated on the independent test set.

The performance of each model was evaluated using widely accepted classification metrics, including accuracy, precision, recall, F1-score, and AUC-ROC.

Table 7. Model performance evaluation.

Table 7 compares the performance of all machine learning models used in this study. The results show that all machine learning models outperform the baseline Logistic Regression model. XGBoost achieved the strongest overall classification performance in terms of accuracy, precision, recall, and F1-score. Specifically, XGBoost achieved an accuracy of 0.96, precision of 0.95, recall of 0.94, and F1-score of 0.95. However, Random Forest achieved the highest AUC-ROC value of 0.94, indicating stronger class-separation ability. Therefore, XGBoost is selected as the preferred model based on overall classification performance, while Random Forest is recognized as the best model in terms of AUC-ROC.

The Receiver Operating Characteristic curve in Figure 7 is used to evaluate the

Figure 7. ROC curve.

classification performance of the models by illustrating the trade-off between the True Positive Rate and False Positive Rate. The Area Under the Curve provides a single metric to summarize each model’s discriminative performance. From the analysis, the Random Forest model achieved the highest AUC value of 0.94, indicating excellent classification performance and a strong ability to distinguish between default and non-default cases. XGBoost and Support Vector Machine also demonstrated strong performance, with AUC values of 0.90 and 0.89, respectively. In contrast, Logistic Regression showed comparatively lower performance with an AUC of 0.84, although it still maintained acceptable predictive power. The ROC curves further confirm these findings, as the Random Forest curve is closest to the top-left corner of the plot, indicating superior class-separation ability.

Overall, the results indicate that ensemble-based models outperform traditional approaches. XGBoost achieved the best overall performance in terms of accuracy, precision, recall, and F1-score, while Random Forest obtained the highest AUC value. This indicates that both models are effective, with XGBoost providing superior classification accuracy and Random Forest excelling in class separation.

The confusion matrix in Figure 8 evaluates the classification performance of the XGBoost model on the independent test set of 1000 observations. The model correctly classified 618 non-default cases and 342 default cases. It misclassified 18 non-default borrowers as default and 22 default borrowers as non-default. Therefore, the model achieved an overall accuracy of 96%, with precision of 0.95, recall

Figure 8. Confusion matrix.

of 0.94, and F1-score of approximately 0.95. These results are consistent with the XGBoost performance reported in Table 7.

The relatively low number of false negatives is important in credit screening because missed risky borrowers may create financial losses for banks. However, since the data are synthetic, the confusion matrix should be interpreted as simulation-based performance rather than confirmed real-world operational performance.

Figure 8 illustrates the confusion matrix, showing strong classification performance with a high number of correct predictions. However, the performance may be influenced by the structured nature of the synthetic dataset, and further validation using real-world data is recommended.

Based on the evaluation metrics, XGBoost was selected as the preferred model because it achieved the highest accuracy, precision, recall, and F1-score among the compared models. The model achieved 96% accuracy, 0.95 precision, 0.94 recall, and 0.95 F1-score on the independent test set. However, Random Forest achieved the highest AUC-ROC value of 0.94, indicating stronger overall class-separation ability. Therefore, the findings suggest that XGBoost provides the strongest overall classification performance, while Random Forest offers the strongest discriminative performance based on AUC-ROC. The strong performance of XGBoost can be attributed to its ability to capture nonlinear relationships, handle structured financial data effectively, reduce overfitting through regularization, and combine weak learners into a strong predictive model. These characteristics make XGBoost suitable for credit risk prediction under controlled simulation conditions. Nevertheless, the high performance should be interpreted cautiously because the dataset is synthetic and may be more structured than real-world banking data.

The SHAP summary plot provides a global view of feature importance. The results show that:

Debt-to-income ratioMonthly incomeLoan amountPrevious default history

These are the most influential features in determining credit risk.

The SHAP summary plot provides a global view of feature importance. The results show that debt-to-income ratio, monthly income, loan amount, previous default history, and late payment count are among the most influential features in determining predicted credit risk.

As shown in Figure 9, the SHAP summary plot presents the overall contribution of the most influential features to the model’s credit-risk predictions. Features with positive SHAP values increase the likelihood of default, while features with negative SHAP values reduce the predicted risk. Among the key factors, previous default history has a strong impact, suggesting that borrowers who have defaulted before are more likely to be classified as high-risk ([1]). Similarly, higher debt-to-income ratio, larger loan amount, and more frequent late payments are associated with increased credit risk because they reflect greater financial pressure and weaker repayment behavior. In contrast, higher monthly income and stronger repayment capacity tend to reduce the probability of default.

Figure 9.SHAP summary plot.

The SHAP individual explanation plot in Figure 10 explains a specific borrower’s prediction by showing how each feature contributes to the final predicted outcome. Positive SHAP values push the prediction toward default, while negative SHAP values reduce the predicted risk. For example, a high debt-to-income ratio, previous default history, or frequent late payments may increase the predicted risk, whereas stable income, lower loan burden, stronger account balance, or longer credit history may reduce the predicted risk. This case-level explanation helps loan officers and decision-makers understand why a borrower is classified as high-risk or low-risk, thereby improving transparency and trust in the model.

Figure 10 illustrates how individual features contributed to a specific borrower’s predicted credit risk outcome.

Furthermore, the SHAP individual plot provides a detailed explanation of how the model arrives at a prediction for a specific borrower by breaking down the

Figure 10.SHAP individual explanation plot.

contribution of each feature. Each feature either pushes the prediction toward a higher risk of default or pulls it toward a lower risk. For example, factors such as a high debt-to-income ratio, previous default history, or frequent late payments typically increase the predicted risk, as they indicate financial instability or poor repayment behavior. On the other hand, features like a stable income level or lower loan burden may reduce the risk and move the prediction toward a non-default outcome. This step-by-step contribution helps in understanding the reasoning behind the model’s decision for an individual case. As a result, it becomes easier for loan officers and decision-makers to justify and trust the prediction, making the model more practical and useful in real-world banking applications.

The improved performance of the proposed approach can be explained by several key factors that collectively enhanced the effectiveness of the model. First, the use of ensemble learning techniques, particularly XGBoost and Random Forest, played a crucial role in achieving higher accuracy. These models are well known for their ability to capture complex nonlinear relationships and interactions among features, which are often overlooked by traditional statistical methods.

Second, the quality and relevance of the selected features significantly contributed to the model’s performance. The inclusion of both financial and behavioral variables, such as income level, debt-to-income ratio, and repayment history, ensured that the model was trained on realistic and meaningful indicators of credit risk. This allowed the model to better understand borrower behavior and make more accurate predictions.

Another important factor is the application of proper data preprocessing techniques. Handling missing values, transforming categorical variables, and applying feature scaling improved the overall consistency and stability of the dataset. These steps reduced noise and ensured that the models could learn patterns more effectively.

Furthermore, hyperparameter tuning played a vital role in optimizing model performance. By carefully adjusting model parameters using techniques such as grid search, the study was able to reduce overfitting and improve generalization capability. This ensured that the model performs well not only on training data but also on unseen data. In addition, the integration of explainability techniques such as SHAP enhanced the overall reliability of the model. By providing clear insights into feature importance and prediction behavior, SHAP helped validate that the model’s decisions were logical and aligned with real-world expectations. This not only improved transparency but also increased confidence in the model’s outcomes.

However, it is important to note that the high performance achieved in this study may be partially influenced by the structured nature of the synthetic dataset. In real-world banking environments, data is often more complex, noisy, and imbalanced, which can reduce model performance. Therefore, the reported results should be considered as an upper-bound estimate rather than a direct reflection of real-world outcomes. Despite this limitation, the consistency of performance across multiple evaluation metrics suggests that the proposed model is robust and reliable under controlled conditions. Future research should focus on validating the model using real-world datasets to further confirm its practical applicability.