Comparing the Effectiveness of Machine Learning and Deep Learning Models in Student Credit Scoring: A Case Study in Vietnam

Abstract

In emerging markets like Vietnam, where student borrowers often lack traditional credit histories, accurately predicting loan eligibility remains a critical yet underexplored challenge. While machine learning and deep learning techniques have shown promise in credit scoring, their comparative performance in the context of student loans has not been thoroughly investigated. This study aims to evaluate and compare the predictive effectiveness of four supervised learning models—such as Random Forest, Gradient Boosting, Support Vector Machine, and Deep Neural Network (implemented with PyTorch version 2.6.0)—in forecasting student credit eligibility. Primary data were collected from 1024 university students through structured surveys covering academic, financial, and personal variables. The models were trained and tested on the same dataset and evaluated using a comprehensive set of classification and regression metrics. The findings reveal that each model exhibits distinct strengths. Deep Learning achieved the highest classification accuracy (85.55%), while random forest demonstrated robust performance, particularly in providing balanced results across classification metrics. Gradient Boosting was effective in recall-oriented tasks, and support vector machine demonstrated strong precision for the positive class, although its recall was lower compared to other models. The study highlights the importance of aligning model selection with specific application goals, such as prioritizing accuracy, recall, or interpretability. It offers practical implications for financial institutions and universities in developing machine learning and deep learning tools for student loan eligibility prediction. Future research should consider longitudinal data, behavioral factors, and hybrid modeling approaches to further optimize predictive performance in educational finance.

1. Introduction

In an era of constant global economic fluctuations, access to financial resources has become a crucial factor for the sustainable development of individuals and communities. For students who must cover tuition fees and living expenses, securing loans from financial institutions is essential (Organisation for Economic Co-operation and Development 2025). As a result, accurately predicting students’ loan eligibility is a critical step in helping lenders mitigate credit risk while ensuring that students can access funding in a stable and efficient manner. However, traditional credit evaluation methods often struggle to handle large, complex datasets that contain nonlinear relationships (Mestiri 2024).

The rapid advancement of Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL) techniques, has opened up new prospects for credit risk management. Numerous studies have demonstrated that ML-based models outperform traditional statistical methods in predicting loan defaults (Abbas and Hussein 2024; Golbayani et al. 2020; Schmitt 2022). Algorithms such as Linear Discriminant Analysis, Random Forest, Logistic Regression, Decision Tree, Support Vector Machine, and Deep Neural Networks have been widely applied, yielding promising results in credit scoring and risk assessment (Sayed et al. 2024).

The application of ML and DL techniques in credit scoring has shown promising results in recent years. However, the majority of existing studies have focused on traditional borrowers with well-established credit histories, while the prediction of loan eligibility for students remains largely underexplored. This lack of attention to student-specific financial behavior and credit profiles has resulted in a shortage of tailored models that accurately predict the loan eligibility of this demographic. Although studies from Shukla et al. (2023) and Wang et al. (2015) evaluated the performance of ML and DL algorithms in general credit scoring, comprehensive comparative research focusing specifically on student borrowers is still limited.

Consequently, the unique financial patterns, limited credit histories, and alternative risk factors associated with student borrowers remain underexplored. Additionally, the scarcity of student-specific datasets restricts the applicability of existing findings to educational loan contexts (Wang et al. 2015). Without datasets that reflect student income sources, repayment behavior, and academic performance, the predictive power of these models for predicting student loan eligibility remains questionable.

Another underexplored aspect in previous research is the lack of a comprehensive comparative analysis between ML and DL approaches in student credit scoring applications. Rather than offering broad benchmarking across multiple algorithms, many studies have focused narrowly on a limited subset of classifiers (Shukla et al. 2023), thereby hindering a deeper understanding of their relative strengths and weaknesses, especially in the context of student credit risk. Furthermore, few studies have addressed the trade-offs between model accuracy, interpretability, and computational efficiency, which are crucial factors for financial institutions when adopting ML and DL credit scoring models.

This study aims to compare the effectiveness of ML and DL models in predicting student loan eligibility. Specifically, we evaluate and compare the performance of traditional ML models (e.g., Logistic Regression, Decision Trees, and Random Forest) and DL models (e.g., Deep Neural Networks) based on key performance metrics such as accuracy, precision, recall, and F1-score. Additionally, we analyze the advantages and limitations of each approach to determine their practical applicability in student credit scoring systems. By addressing these research gaps, this study seeks to provide a comprehensive evaluation of ML and DL methodologies in student loan eligibility prediction, with a particular focus on the case of Vietnam.

Vietnam represents a typical case of a developing economy in the credit scoring landscape, where the adoption of ML and algorithmic credit scoring is accelerating but remains constrained by regulatory and operational challenges. Local banks are increasingly applying ML models such as Logistic Regression and Random Forest to improve credit risk management using both financial and non-financial data (Van Trung and Vuong 2024). However, the lack of a comprehensive legal framework raises concerns over fairness, transparency, and data privacy (Lainez and Gardner 2023). Like many emerging markets, Vietnam faces data limitations, including imbalance and bias, which undermine the reliability of credit assessments (Fejza et al. 2022). The absence of a unified national credit database further exacerbates these issues (T. Le 2017). Despite these challenges, Vietnam offers a promising testing ground for innovative approaches in credit scoring, with the potential to inform similar initiatives in other developing countries.

By exploring the intersection between ML and DL techniques in credit risk assessment and student loan repayment prediction, this study contributes to the expanding body of literature on data-driven credit scoring models. The results are expected to provide practical guidance for financial institutions, policymakers, and student loan providers in optimizing credit risk models while promoting fair, transparent, and inclusive access to financial support for students. This study not only enriches the literature on ML and DL credit scoring but also provides concrete implications for designing effective and equitable loan eligibility prediction in emerging economies such as Vietnam.

2. Literature Review

2.1. Conceptual Framework of Credit Scoring and Student Loan Repayment

Credit scoring is a systematic approach employed by lenders to assess the creditworthiness of potential borrowers, including students applying for educational loans. This process involves analyzing a range of data points to estimate the likelihood of loan repayment, an increasingly critical issue amid the growing volume of student loan debt and rising default rates in countries such as the United States (Lochner and Monge-Naranjo 2016). A thorough understanding of credit scoring and its implications for student loan repayment is essential for both borrowers and lending institutions. Credit scoring typically leverages algorithmic models to evaluate an individual’s credit risk based on historical financial data, including payment history, debt-to-income ratios, and outstanding obligations (Kiviat 2019). These models help lenders determine loan eligibility, interest rates, and repayment conditions, thereby shaping the overall structure of the lending landscape (Malik and Hermawan 2018).

Students as borrowers often encounter significant challenges in loan repayment due to escalating tuition fees and an uncertain job market, both of which contribute to elevated default rates (Lochner and Monge-Naranjo 2016). Effective credit scoring mechanisms can assist in designing loan products that balance access to education with realistic and manageable repayment terms, thereby reducing the risk of default (Luo et al. 2018). However, while credit scoring offers a structured and data-driven method for assessing borrower risk, it may inadvertently perpetuate inequality. Luo et al. (2018) stated that individuals from disadvantaged socioeconomic backgrounds often face systemic barriers to obtaining loans, regardless of their true repayment potential. This underscores the need for more inclusive and equitable lending practices that go beyond traditional credit evaluation metrics.

2.2. Traditional and Data-Driven Approaches in Credit Scoring

Credit scoring methodologies can be broadly categorized into traditional statistical approaches and more recent data-driven techniques leveraging ML and DL techniques (illustrated with details in Figure 1). While traditional methods rely on fixed assumptions and structured data, modern techniques allow for greater flexibility and predictive power through the use of complex, high-dimensional data.

Traditional methods, such as logistic regression and linear discriminant analysis, rely heavily on historical credit data and assume linear relationships, which often fail to capture complex borrower behavior (Adesanya 2024; Hussain et al. 2024). These models also struggle to adapt to changing economic conditions and borrowers’ circumstances, especially in cases with limited credit histories like students (Adesanya 2024; Hussain et al. 2024). While Logistic Regression originates from this statistical tradition, its widespread use and ability to provide probabilistic outputs often lead to its inclusion as a baseline model within ML comparisons, bridging the gap between traditional and algorithmic approaches.

In contrast, data-driven approaches leverage ML techniques, such as decision trees, ensemble methods, and neural networks, to analyze large-scale data and detect intricate patterns missed by traditional models (Hussain et al. 2024; Adesanya 2024). ML and DL systems enable real-time processing, allowing for timely and accurate credit evaluations, and can incorporate alternative data sources to dynamically update risk assessments (Adesanya 2024). While these approaches offer enhanced precision and adaptability, concerns regarding data privacy and algorithmic bias remain critical and must be addressed to ensure ethical implementation (Kotb and Proaño 2023). These advanced techniques enable financial institutions to assess credit risk more effectively, particularly in regulated environments (Shukla and Gupta 2024).

Beyond personal credit scoring, ML and DL techniques have also been applied in the domain of corporate credit rating. Several studies, such as that of Golbayani et al. (2020), have compared the predictive performance of models including Neural Networks, Support Vector Machine, and Decision Trees in the context of corporate credit evaluation.

Moreover, the shift towards complex ML/DL models necessitates more comprehensive evaluation frameworks. Recent literature highlights the need to assess models across multiple dimensions beyond predictive performance, often summarized by principles like Sustainability (Robustness), Accuracy, Fairness, and Explainability (SAFE) (Giudici 2024; Babaei et al. 2025).

2.3. Criteria in Personal Credit Scoring

2.3.1. General Criteria in Personal Credit Scoring Models

Traditional credit scoring models, significantly influenced by benchmarks like the FICO score, typically utilize established criteria derived from an individual’s financial history and behaviors. Commonly included demographic and status variables encompass gender, age, marital status, number of dependents, phone ownership, education level, occupation, duration of residency, and credit card possession (Šušteršič et al. 2009; Hand et al. 2005; Lee and Chen 2005; Sarlija et al. 2004; Banasik et al. 2003; Chen and Huang 2003; Lee et al. 2002; Orgler 1971; Steenackers and Goovaerts 1989). Financial and employment factors, such as employment duration, loan amount, loan term, homeownership, monthly income, bank account status, car ownership, existing mortgages, loan purpose, and collateral, are also frequently incorporated (Ong et al. 2005; Lee and Chen 2005; Greene 1997; Sarlija et al. 2004; Orgler 1971; Steenackers and Goovaerts 1989). Some models extend to spousal information (Orgler 1971), while less common criteria might include area codes or length of banking relationship (Bellotti and Crook 2009; Banasik and Crook 2007; Andreeva 2006; Banasik et al. 2003).

The number of criteria employed varies significantly, with no universally ideal number; selection often depends on data availability, context, and country-specific factors (Abdou and Pointon 2011). Many models build upon the five core FICO factors: payment history, amounts owed, length of credit history, new credit, and credit mix (Zhang et al. 2020).

Credit scoring data generally originate from two primary sources: traditional data (loan specifics, repayment behavior, and credit activity) collected by financial institutions or credit bureaus, and alternative data obtained from third parties (Knutson 2020). Alternative data sources are diverse, including utility payments, mobile device usage, social media activity, online transactions, and behavioral metrics, offering supplementary insights into lifestyle and potential creditworthiness (Knutson 2020). While traditional scoring predicts future financial behavior based on past actions, alternative scoring leverages non-financial data to infer creditworthiness. Utility bills and rental payment history are often considered to be “credit-like” alternative data due to their long-term nature (Knutson 2020). Public records and indicators of residential stability have also been utilized.

2.3.2. Credit Scoring Criteria in the Vietnamese Context

In Vietnam, as an emerging market with a rapidly expanding consumer finance sector, research specifically detailing the criteria used in personal credit scoring often centers on the framework established by the National Credit Information Center of Vietnam (CIC), which is sometimes augmented by the criteria used by commercial banks. CIC’s scoring process relies on classical statistical methods and utilizes nine indicators across three main groups: (i) Debt Balance and Status (total debt, number of lenders, highest debt group, and repayment term), (ii) Repayment History (months with substandard debt, years with bad debt, and lenders with bad debt), and (iii) Credit Relationship History (years of relationship and new loans taken) (Le and Dang 2016).

Vietnamese commercial banks typically adapt this CIC foundation. For instance, Dong A Bank adds factors like gender, education, employer type, and income (Mai 2014). Vietinbank incorporates job tenure, housing status, family structure, and bank relationship details (V. T. Le 2010). Asia Commercial Joint Stock Bank (ACB) uses personal data (age, education, residence, and income source) alongside loan specifics (V. T. Le 2010), while Agribank employs an extensive system for individual business customers (Huynh 2017). This demonstrates an adaptation of the standard model to specific institutional risk assessment needs.

2.3.3. Unique Considerations and Criteria for Student Credit Scoring

Assessing student creditworthiness poses unique challenges, largely due to their typically limited or non-existent formal credit history, often termed as a “thin file” (Baker and Montalto 2019; Baum and Steele 2010). This data scarcity diminishes the effectiveness of traditional models, necessitating the inclusion of criteria specific to the student experience.

Academic performance represents a critical category. Metrics such as Grade Point Average (GPA), field of study, educational institution prestige, and year of study serve as valuable proxies for discipline, future earning potential, and the likelihood of degree completion and subsequent loan repayment (Jackson and Reynolds 2013; Adams and Moore 2007). The research indicates a link between lower academic performance and higher-risk credit behaviors (Adams and Moore 2007), potentially making the academic data potent predictors for this segment.

Given students’ common reliance on external support, their socio-economic background, including familial aspects, is highly relevant. Parental income, assets, education level, and the number of household dependents can signal available financial support and baseline stability, influencing the student’s debt needs and management capacity (Addo et al. 2016; Houle 2014). Basic demographics, such as age and residency status, provide further context (Baker and Montalto 2019).

Alternative data and non-traditional financial indicators gain particular importance. Information on part-time employment (status, income, hours, and stability) offers a direct measure of current repayment ability (Robb and Pinto 2010; Mendes-Da-Silva et al. 2012). Data on living expenses, housing situation (dormitory, family home, and renting), tuition fees, and potentially utility or mobile phone payment patterns act as proxies for financial management skills, financial pressures, and payment discipline in the absence of formal credit records (Robb and Pinto 2010).

Finally, any existing borrowing history, however limited (including prior loan details like purpose, amount, term, and repayment), remains a vital input if available. Consequently, a robust student credit scoring model necessitates a multi-dimensional approach, integrating academic achievements, socio-economic context, employment details, alternative financial behaviors, and personal characteristics alongside any available traditional credit data. This holistic perspective is essential for accurately evaluating the unique risk profile of student borrowers.

2.4. Machine Learning Models and Deep Learning Models in Credit Scoring

Machine Learning has emerged as a powerful tool in credit scoring, offering enhanced predictive performance and the ability to model complex, nonlinear relationships in borrower data. A range of algorithms has been applied in credit risk assessment, each with distinct characteristics influencing their suitability for specific data contexts and regulatory requirements. Commonly used models include Logistic Regression, Decision Trees, Random Forest, Support Vector Machine, and advanced boosting techniques such as Extreme Gradient Boosting and Light Gradient Boosting Machine.

Logistic regression remains a widely used baseline model due to its simplicity and interpretability; however, its performance declines in the presence of nonlinear interactions (Meng et al. 2025). Random Forest offers improved accuracy and is robust against overfitting, while also providing useful insights into feature importance (Han 2024). Gradient boosting methods, particularly Extreme Gradient Boosting and Light Gradient Boosting Machine, have gained recognition for their strong performance in large-scale datasets and competitive results in real-world applications (Meng et al. 2025; Mukhanova et al. 2024). Neural networks, though capable of capturing intricate patterns, often face limitations in interpretability, which can hinder their deployment in highly regulated environments (Chen 2025).

Similarly, Lessmann et al. (2015) concluded that Support Vector Machine is competitive with Artificial Neural Networks in terms of predictive performance after evaluating 41 classification methods on multiple credit scoring datasets. In contrast, Teles et al. (2021), through a comparative analysis of Random Forest and Support Vector Machine, found Random Forest to be a promising option for personal credit risk management due to its speed and simplicity, whereas Support Vector Machine outperformed Random Forest in classification accuracy for credit recovery prediction tasks.

Overall, in the context of personal credit risk assessment, ML algorithms such as K-Nearest Neighbors, Support Vector Machine, Decision Trees, Random Forest, Adaptive Boosting, and Gradient Boosting have consistently demonstrated superior performance compared to traditional logistic regression models.

In terms of strengths, ML models generally surpass traditional statistical methods in predictive accuracy (Shukla and Gupta 2024). Their scalability allows for efficient processing of large and heterogeneous datasets, and their flexibility enables integration of various data types, including alternative and behavioral data sources (Shukla and Gupta 2024). However, these advantages are accompanied by notable challenges. A key concern is the lack of transparency in model decision-making, particularly in DL systems, which complicates model validation and regulatory compliance (Chen 2025). Additionally, many ML models are sensitive to data imbalance, which may distort risk predictions and reduce the reliability of outcomes in minority classes (Han 2024). Finally, computational complexity remains a barrier to deployment, especially in institutions with limited technical infrastructure (Shukla and Gupta 2024).

To address the lack of transparency, particularly in complex ensemble or deep learning models, Explainable AI (XAI) techniques have gained traction. Methods like Shapley values are increasingly applied post-hoc to quantify the contribution of each input feature to a model’s prediction for specific instances or globally (Babaei et al. 2023). Such techniques are vital for understanding model behavior, debugging, and meeting regulatory expectations for high-risk applications like credit scoring.

Deep learning has increasingly been adopted in credit scoring due to its capacity to model complex, nonlinear relationships and effectively utilize diverse data types (Bari et al. 2024; Mienye et al. 2024). Prominent architectures, such as Convolutional Neural Networks, Long Short-Term Memory Networks, and various hybrid models, have demonstrated strong performance in capturing hidden patterns and improving prediction outcomes, particularly for underrepresented borrower groups. Convolutional Neural Networks are particularly effective in detecting spatial correlations and have been applied to alternative data sources beyond traditional financial indicators (Mienye et al. 2024). Long Short-Term Memory Networks are designed to process sequential data and are well-suited to modeling temporal patterns in borrower behavior. Hybrid models, which integrate features from architectures such as Recurrent Neural Networks and Deep Neural Networks, have shown improved accuracy, especially in contexts involving sparse or incomplete credit histories (Kimani et al. 2024).

Key strengths of DL in credit risk assessment include superior predictive performance, driven by the ability to model intricate, nonlinear data relationships (Bari et al. 2024). These models are also highly flexible, capable of processing both structured and unstructured inputs, thereby enhancing the inclusivity of credit evaluations (Bari et al. 2024). Moreover, hybrid models often achieve higher evaluation metrics compared to single-architecture approaches, reflecting greater robustness in predictive tasks (Kimani et al. 2024).

Nonetheless, several challenges persist. DL models typically lack interpretability, which can hinder transparency and raise ethical concerns, particularly in high-stakes decision-making contexts like credit scoring (Bari et al. 2024). In addition, their computational demands are substantial, making large-scale deployment difficult for institutions with limited technological infrastructure (Kimani et al. 2024). Performance is also sensitive to data quality; inconsistencies or noise in the input data can significantly degrade model reliability (Mienye et al. 2024). Overall, while DL offers meaningful advancements in the accuracy and adaptability of credit scoring systems, its successful integration into financial services requires careful attention to interpretability, fairness, and operational feasibility.

In Vietnam, a representative developing economy, the adoption of ML and DL models in credit scoring has accelerated to meet the growing demand for inclusive and data-driven credit assessment. Limited credit history, evolving borrower behavior, and the rapid expansion of consumer finance have made traditional models less effective, prompting institutions to explore more adaptable and predictive approaches. ML models such as logistic regression continue to serve as baseline classifiers due to their simplicity and interpretability (Van Trung and Vuong 2024). However, more advanced methods like Random Forest have demonstrated higher accuracy by aggregating decision trees to reduce overfitting, making them suitable for diverse borrower profiles in the Vietnamese context (Tran et al. 2021). Gradient boosting frameworks—including Light Gradient Boosting Machine and CatBoost—have emerged as leading solutions in handling imbalanced datasets and capturing complex data structures, which are common in developing markets (Tran et al. 2021). By contrast, Support Vector Machines have shown relatively lower effectiveness compared to ensemble methods in local experiments (Tran et al. 2021). DL models have also gained traction. Sequential Deep Neural Networks offer enhanced representation learning and have outperformed traditional models in several local studies (Tai and Huyen 2019). Moreover, Convolutional Neural Networks, though originally designed for image analysis, have been adapted for structured credit data, showing potential in uncovering hidden patterns not easily captured by conventional models (Tai and Huyen 2019). Despite these advancements, concerns remain regarding transparency, fairness, and regulatory oversight. As Vietnam continues to digitalize its financial infrastructure, the adoption of ML and DL credit scoring must be accompanied by ethical and context-aware governance frameworks (Lainez and Gardner 2023).

Despite the growing application of ML and DL in credit scoring, especially in emerging markets like Vietnam, several important gaps remain. Research specifically focused on student credit scoring is limited, and direct comparisons between ML and DL models using real-world student loan data are scarce. While prior studies have explored ML and DL in both personal and corporate credit contexts (Lessmann et al. 2015; Teles et al. 2021; Golbayani et al. 2020), their application to student borrowers, with unique features such as academic performance and personal traits, remains underexamined.

Moreover, few studies address the practical feasibility of advanced models in educational finance, particularly regarding interpretability and deployment cost. This study seeks to fill these gaps by applying and comparing selected ML models (Random Forest, Gradient Boosting, and Support Vector Machine) and a DL model (PyTorch) to predict student loan eligibility in Vietnam, offering both performance insights and practical implications for data-driven lending in education.

3. Model Evaluation

3.1. Evaluation Metrics

To ensure a comprehensive and robust evaluation of model performance, this study adopts a multi-dimensional framework tailored to the specific nature of classification (Kuhn and Johnson 2018).

For binary classification models that predict whether a student is credit-eligible or not, the following metrics are applied: (i) Accuracy measures the overall proportion of correct predictions over total predictions. Though commonly used, it may be misleading with imbalanced data (James et al. 2013); (ii) Confusion Matrix provides a detailed breakdown of true positives, false positives, true negatives, and false negatives, helping to to identify specific misclassification patterns (Fawcett 2006); (iii) Precision evaluates the proportion of true positives among all predicted positives, making it particularly valuable in the contexts false positives incur high costs (Sokolova and Lapalme 2009); (iv) Recall (Sensitivity) captures the proportion of actual positives correctly identified, which is critical in credit scoring, when missing a high risk borrower (false negative) can have serious implications (Powers 2011); (v) F1-Score, the harmonic mean of Precision and Recall, offers a balanced metric particularly useful under class imbalance conditions (Géron 2019).

In the case of deep neural networks, performance monitoring during training is also conducted: Loss Function Tracking across training epochs provides insight into the learning process and potential overfitting. A significant gap between training and validation loss typically signals poor generalization performance (Bengio et al. 2013).

By integrating these evaluation metrics across different model types and training processes, this framework enables a well-rounded and data-driven comparison. It supports the identification of the most suitable algorithm for assessing student loan eligibility in the context of Vietnam’s higher education landscape.

3.2. Results and Discussion

Table 1. Comparison of models based on classification metrics.

Model	Train Accuracy	Test Accuracy	Confusion Matrix	Class 0 (Precision/Recall/F1)	Class 1 (Precision/Recall/F1)	Macro Avg F1
Random Forest	95.8%	82%	[[121, 18], [19, 47]]	86%/87%/0.87	72%/71%/0.72	0.79
Gradient Boosting	99.9%	75%	[[107, 32], [20, 46]]	~84%/~77%/0.80	59%/70%/0.64	0.72
Support Vector Machine	98.3%	80%	[[125, 14], [25, 41]]	83%/90%/0.87	75%/62%/0.68	0.77
Pytorch DNN	86.2%	85.5%	[[125, 14], [20, 46]]	86%/90%/0.88	77%/70%/0.73	0.81

summarizes the classification performance of all evaluated models. Among them, the Deep Neural Network (PyTorch) achieved the highest overall accuracy at 85.55%, indicating strong predictive capability compared to traditional ML models. Evaluating its performance in more detail reveals strong metrics for identifying non-eligible borrowers (Class 0 Precision: 86%, Recall: 90%, F1: 0.88). For predicting eligible borrowers (Class 1), the DNN demonstrated good Precision (77%) and Recall (70%), resulting in an F1-Score of 0.73. Its overall Macro Average F1-Score was 0.81, the highest among the models.

In comparison, the Random Forest model delivered robust and well-balanced results, with an overall accuracy of 82% and a Macro Average F1-Score of 0.79. A closer look at the class-specific metrics shows that it achieved a Precision of 86% and a Recall of 87% for class 0 (non-eligible borrowers), and a Precision of 72% and a Recall of 71% for Class 1 (eligible borrowers). These results suggest that Random Forest not only maintains high overall accuracy but also performs reliable discrimination between both target classes, making it a strong candidate for credit eligibility classification in practice.

The Support Vector Machine model achieved an overall accuracy of 80%, demonstrating relatively high Precision for Class 1 (75%) but a lower Recall (62%). This suggests that while Support Vector Machine is effective in minimizing false positives among creditworthy students, it tends to miss a considerable number of true positives. Therefore, it may be more appropriate in contexts where prediction confidence is prioritized over coverage, such as conservative credit approval processes.

The Gradient Boosting model yielded the lowest overall accuracy among the evaluated classifiers (75%). Nevertheless, it achieved 70% for Class 1, suggesting a stronger tendency to identify creditworthy students. However, its Precision for this class was only 59%, implying a higher rate of false positives. This trade-off may limit its practical applicability in contexts where misclassification costs are high.

Finally, the DL model (PyTorch) was trained over 20 epochs, during which the loss steadily decreased from 0.5446 to 0.2256 (see

Table 2. Training process metrics of the DL model.

Model	Initial Loss	Final Loss	Number of Epochs
DL (PyTorch)	0.5446	0.2256	20

). This consistent reduction in loss indicates a stable learning process with no signs of overfitting, supporting the model’s high classification accuracy. These results underscore the potential of DL techniques in handling complex student credit data.

The analysis indicates that each model exhibits distinct strengths and trade-offs in the context of student loan eligibility prediction. Among them, Random Forest stands out as a robust and well-balanced model that achieved solid results, with balanced F1-Scores across both target classes, confirming its reliability in handling both types of predictive tasks.

The DL model implemented in PyTorch recorded the highest overall classification accuracy (85.55%), demonstrating a strong capacity to learn and represent complex, nonlinear relationships within the dataset. This result affirms the potential of DL techniques in modeling high-dimensional educational and financial features. Nevertheless, the effectiveness of this approach depends on the availability of sufficient computational resources, making it most suitable when predictive accuracy is prioritized over interpretability.

In contrast, Support Vector Machine showed reliable precision for Class 1 predictions (75%), indicating its suitability in contexts where minimizing false positives for creditworthy students is critical.

Gradient Boosting exhibited the lowest overall classification accuracy, yet it attained relatively high recall for Class 1 (70%). This result highlights its ability to identify students with loan eligibility but also reveals a trade-off in the form of lower precision (59%), suggesting a higher incidence of false positives. As such, the model may be better suited for cases where false negatives are more costly than false positives.

In summary, the selection of an optimal model should be guided by the specific objectives of the loan eligibility prediction task. When overall accuracy is the primary concern, DL emerges as the most effective solution. If a balance between predictive performance and model interpretability is desired, Random Forest represents an appropriate choice. In scenarios where misclassification costs are asymmetric, Gradient Boosting may be preferred for its recall-oriented strengths, while Support Vector Machine may be selected for its higher precision.

Vietnamese students, like many in developing economies, often lack formal credit histories due to limited or no prior engagement with banking institutions. Consequently, traditional credit scoring mechanisms, which are typically reliant on documented financial behavior, are inadequate for predicting loan eligibility in this population. In this context, non-traditional variables such as living costs, tuition fees, and part-time work status serve as proxy indicators of financial stability and repayment capacity. These features help bridge the informational gap left by the absence of formal credit records, rendering them particularly valuable in algorithmic models for identifying latent patterns of loan eligibility.

Vietnam’s formal lending system remains ill-equipped to serve individual, low-income borrowers such as students, as noted by Iwase (2011) and Malesky and Taussig (2009). Bureaucratic inefficiencies, relationship-based lending practices, and the lack of standardized credit evaluation frameworks contribute to the systematic exclusion of students from formal credit channels. As a result, data-driven models that incorporate broader socioeconomic indicators, particularly non-traditional features, are more effective in capturing the financial realities of students operating within a fragmented and unequal credit environment.

The inclusion of variables such as ‘TUITION’, ‘PHONECOST’, and ‘LIVINGCOST’ in the model underscores the considerable financial pressure faced by students amid rising education expenses and limited institutional support. These features indirectly reflect both the financial burden and the likelihood of students seeking loans to cover essential needs, which is consistent with Chapman and Liu’s (2013) analysis of the growing repayment burden among student borrowers. In contrast, variables such as ‘BICYCLE’ or ‘DORM’, while potentially relevant to lifestyle or living arrangements, may be too commonplace or culturally normalized to serve as meaningful predictors of credit behavior.

Given the lack of transparency and weak regulatory oversight in Vietnam’s credit system (Vu 2024), many students resort to informal lenders. These lenders often do not rely on formal documentation but instead evaluate borrowers based on observable indicators such as employment status or monthly expenditures. This practice mirrors the model’s reliance on non-traditional features, reinforcing the importance of such variables not only from a predictive standpoint but also in reflecting real-world financial decision-making in an underregulated credit market.

The findings from this study offer important implications for the design of student credit policies in developing economies, particularly in the context of Vietnam, where the formal lending infrastructure remains underdeveloped and poorly aligned with the financial realities of university students.

First, the demonstrated predictive power of both traditional and non-traditional variables underscores the need to develop alternative credit scoring frameworks tailored specifically to student borrowers. Most students lack formal credit histories, so conventional scoring methods based on prior financial behavior are insufficient. Policymakers and financial institutions should consider incorporating alternative data sources such as tuition fees, living costs, part-time employment status, and mobile phone expenses into risk assessment models. These indicators serve as practical proxies for students’ financial capacity and repayment behavior in the absence of conventional credit records.

Second, the high prevalence of informal lending practices among students calls for the development of targeted microcredit programs with flexible repayment terms. Such programs could offer small loan amounts, low interest rates, and deferred repayment schedules aligned with students’ academic calendars or post-graduation income streams. These features would reduce reliance on informal or exploitative lending channels, which often impose high interest rates and lack borrower protections.

Third, there is a strong case for institutional collaboration between universities, banks, and state agencies to improve credit access for students. Universities can play an important role in verifying student enrollment, academic progress, and financial need, thereby lowering the information asymmetry faced by lenders. In turn, banks can develop tailored financial products with simplified application processes and greater transparency.

Fourth, the study highlights the value of integrating behavioral and educational data into credit evaluations. Incorporating indicators such as academic performance, housing arrangements, and financial self-reliance may enhance the precision of credit assessments and provide a more holistic understanding of student borrowers.

Finally, to ensure sustainability and borrower protection, there is a pressing need to improve the regulatory framework governing student lending. Policies should establish clear criteria for student loan eligibility, cap interest rates for education-related loans, and offer legal mechanisms for debt restructuring or forgiveness in cases of post-graduation financial hardship. Transparent oversight mechanisms are also necessary to monitor both formal and informal credit activities targeting students. In sum, a comprehensive, data-informed, and student-centered approach to credit policy is essential for promoting financial inclusion in higher education. The integration of alternative data, institutional coordination, and regulatory reform can collectively enhance the accessibility, equity, and sustainability of student credit systems in Vietnam and similar developing contexts.

4. Methodology

4.1. Research Design

This study aims to evaluate and compare the predictive performance of traditional ML models, including Random Forest, Gradient Boosting, and Support Vector Machine, with a DL model built using the PyTorch framework, in the context of student loan eligibility prediction.

This research adopts a quantitative design comprising four main stages as visualized in Figure 2. First, primary data were collected through online surveys targeting university students. Second, the dataset was preprocessed through cleaning, encoding of categorical variables, and normalization to ensure consistency across features. Third, predictive models were developed using selected ML algorithms—Random Forest, Gradient Boosting, and Support Vector Machine—as well as a DL model implemented with PyTorch. Finally, model performance was evaluated using standard metrics, including Accuracy, Precision, Recall and F1-Score.

PyTorch is selected for the DL component due to its flexibility, ease of model customization, and alignment with current academic and industry practices in ML and DL development.

4.2. Data Collection and Processing

4.2.1. Data Collection

The study data were collected through a random survey of students using a questionnaire. The survey instrument included questions designed to capture key demographic, academic, financial, and behavioral attributes. For instance, students were asked about their age, gender, and cohort (year of admission). Financial information included categorical questions about household income (using ranges like ‘under 100 million VND’, ‘100–200 million VND’, etc.) and expenditure using predefined ranges (e.g., ‘5–10 million VND’, ‘10–15 million VND’, etc.), the number of working family members, and whether the student currently held a part-time job. Living situation details covered aspects like housing type (e.g., Dorm, Living with Family, Acquaintance, or Alone), duration at the current residence, and estimated monthly living (ranges such as ‘<VND 2 million, ‘VND 2–4 million’, etc.), etc. Major universities in Vietnam were randomly selected from large cities such as Hanoi, Da Nang, and Ho Chi Minh City.

Table 3. Information Attributes.

Category	Information Attributes	Relevant Literature
Personal Information	Age, gender, hometown, current residence, and duration of stay	Baker and Montalto (2019)
Academic Background	University, program, major, enrollment year (admission year), Grade Point Average, academic conduct scores, and Tuition	Jackson and Reynolds (2013); Adams and Moore (2007)
Family Financial Status	Household income, monthly expenses, number of dependents, and number of working-age family members	Addo et al. (2016); Houle (2014)
Part-time Employment	Job type, working hours, income, and job stability	Robb and Pinto (2010); Mendes-Da-Silva et al. (2012)
Credit and Loan History	Most recent loan, lending institution, loan amount, term, interest rate, and loan purpose	Baker and Montalto (2019); Baum and Steele (2010)
Living Conditions and Transportation	Housing type, living expenses, main means of transportation, and travel costs	Robb and Pinto (2010)

presents the collected information.

4.2.2. Data Processing

The research sample consists of university students from institutions across Vietnam, representing a diverse range of academic disciplines, years of study, and household economic backgrounds. This diversity enhances the representativeness of the sample and increases the generalizability of the study’s findings to the broader student population.

A total of over 1200 responses were collected through the initial survey. After removing incomplete or invalid entries, the final dataset included 1024 valid observations. Each record contained 21 input attributes and one target variable representing student loan eligibility. Categorical variables were encoded into a numerical format to ensure compatibility with ML algorithms, and all features were normalized to maintain consistency in scale. The processed dataset was then used to train and evaluate the selected ML and DL models.

4.2.3. Research Variables

The dependent variable in this study represents whether a student has prior borrowing experience. Based on the survey question asking if the student had ever received a loan (‘Have you ever borrowed?’), this target variable (Loan) employs a binary classification: it is coded as 1 for students who reported having received any loan in the past, and 0 for students who reported no prior borrowing experience. The models developed aim to classify students into these two groups based on the collected independent variables.

The independent variables were extracted from the structured survey data and correspond to the major categories outlined in the questionnaire, including personal information, academic background, family financial status, part-time employment, credit and loan history, and living conditions. A detailed summary of these variables is provided in

Table 4. Variable descriptions used in the model.

Category	Variables	Description
Personal Information	ID	Identifier
	Age	Age
	Gender	Male, female
Academic	Cohort	Cohort group (e.g., year of admission)
	Tuition	Tuition fee
Families Finance	Hhsize	Number of family members
	Labours	Number of working members in the family
	HHIncome	Household income
	HHExpenditure	Household expenditure
Employment and housing	Work	Having a part-time job or not
	Dorm	Living in a dormitory
	Family	Living with family
	Acquaintance	Living with acquaintances
	Alone	Living alone
	LivingDuration	Duration of residence at current place
Personal expenditure	Livingcost	Total living expenses
	Phonecost	Mobile phone expenses
Transportation	Distance	Distance from residence to school
	Bicycle	Traveling by bicycle
	Motorbike	Traveling by motorbike
	PublicVehicle	Using public transportation
	Othervehicle	Other means of transportation

.

It is important to note that several variables initially collected were excluded from the final models due to practical data considerations encountered during preprocessing. For instance, variables like university, program, and major exhibited excessive unique values, making one-hot encoding infeasible and other encoding methods potentially uninformative without significant grouping efforts beyond the scope of this study. Meanwhile, enrollment year (Cohort) showed a high correlation with age and was retained as the primary temporal indicator related to academic progression. Finally, GPA and academic conduct scores suffered from significant missing values and inconsistencies in reporting scales across different institutions, preventing their reliable use as predictors for the entire dataset.

4.3. Model Development and Training

For the purposes of model development and evaluation, the dataset was split into a training set (80%) and a test set (20%). This split ensures that the models are not overfitted to the training data and are capable of making accurate predictions on previously unseen data. The same approach was applied consistently across all four models examined in this study. These models can be categorized into two main groups. The first group consists of traditional ML algorithms, including Random Forest, Gradient Boosting, and Support Vector Machines. The second group includes a DL model, specifically, a Deep Neural Network implemented using the PyTorch framework. (A detailed Exploratory Data Analysis (EDA) of the dataset, including variable distributions and relationships, is presented in Appendix A, while detailed model specifications are presented in Appendix B).

Random Forest (Ho 1995; Breiman 2001) is implemented as an ensemble of decision trees trained on bootstrap samples of the data with random feature selection at each split. In this study, the Random Forest model was trained using 80% of the dataset (training set), with the remaining 20% reserved for testing. The number of trees (estimators) was set to 100, and the maximum depth was tuned through cross-validation to prevent overfitting. The Gini index was used as the splitting criterion. Feature importance scores were extracted to assess the contribution of each variable to the credit prediction task. Performance was evaluated using Accuracy, Precision, Recall, and F1-Score on the test set.

To provide further insights into the relative influence of different input variables on the prediction outcomes, feature importance analysis from the trained Random Forest model was conducted, as detailed in Figure 3 below.

The analysis reveals a significant reliance on both traditional and non-traditional variables for model prediction, reflecting the importance of alternative data in the context of limited credit history (e.g., Knutson 2020; Robb and Pinto 2010). Among the most significant features are traditional demographic factors like AGE and HHLNCOME (Household Income), alongside several other non-traditional features. These non-traditional features include WORK (Student’s Work Status), TUITION, PHONECOST, and LIVINGCOST. However, not all these non-traditional features are significant, such as BICYCLE, DORM, ALONE, and PUBLIC_VEHICLE.

This mixture of traditional and non-traditional variables, on the one hand, highlights the necessity of incorporating factors beyond normal credit factors when assessing students’ credit, but on the other hand, it also notes that not all such variables contribute significantly to the predictive model. Regardless, overall, the prominence of non-traditional variables like specific educational and living costs may play a critical role in filling the information gap left by students' lack of formal credit history.

Gradient Boosting, introduced by Friedman (2001), was implemented using a decision tree-based ensemble with sequential learning. The model was trained on the same training set (80%) using a learning rate of 0.1, a maximum depth of 3, and 100 boosting iterations, following guidelines from Malohlava and Candel (2020). Loss minimization was performed using binary cross-entropy for the classification version and MSE for the score-based version. Early stopping was applied based on validation loss to avoid overfitting. Evaluation metrics included Accuracy, F1-Score, and MSE. Support Vector Machine (Vapnik 1998) was applied using a radial basis function kernel, selected for its ability to capture nonlinear boundaries. The data were scaled using standard normalization before training. A grid search was used to tune the regularization parameter (C) and kernel coefficient (gamma). The model was trained on 80% of the data and tested on the remaining 20%. Support Vector Machine was evaluated primarily for binary classification using Accuracy, Precision, Recall, and F1-Score. Although probabilistic outputs are not native to Support Vector Machine, probability calibration was applied post hoc using Platt scaling to enhance interpretability. Figure 4 presents the feature importance derived from the trained Gradient Boosting model. Consistent with the findings of the Random Forest, ‘age’ emerges as a primary predictor, alongside variables related to living situation and expenses, such as ‘LivingDuration’ and ‘HHExpenditure’. This provides further insight into the factors driving the Gradient Boosting model’s predictions.

A feedforward Deep Neural Network was developed using PyTorch to model the complex relationships among the 21 features. The architecture consisted of an input layer, two hidden layers with 64 and 32 neurons, respectively, Rectified Linear Unit activation functions, and a final sigmoid output layer for binary classification. The model was trained using the Adam optimizer, binary cross-entropy loss, a learning rate of 0.001, and mini-batch gradient descent (batch size = 32). Training ran for 100 epochs with early stopping based on validation loss. To evaluate generalization performance, the test set was used to compute Accuracy and F1-Score for classification.

5. Conclusions

This study has systematically analyzed and compared the performance of four distinct ML models—Random Forest, Gradient Boosting, Support Vector Machine, and DL using PyTorch—in addressing the problem of predicting student loan eligibility. The results reveal that no single model demonstrates consistent superiority across all evaluation metrics; instead, each model possesses unique strengths and limitations. Consequently, model selection should be guided by the specific objectives of the application, the nature of the available data, and the trade-offs between evaluation criteria such as accuracy, interpretability, and computational efficiency.

The study contributes to a deeper understanding of how ML techniques can be applied to student loan eligibility prediction, particularly for a population segment that typically lacks traditional credit history. These insights provide a practical foundation for financial institutions and educational organizations in selecting appropriate models for credit assessment. The study confirms the recurring observation that ensemble methods, such as Random Forest and Gradient Boosting, tend to outperform individual classifiers in handling the complexities of credit data, even with limited information, similarly to studies in diverse developing contexts like Pakistan (Abbas and Hussein 2024) and Indonesia (Malik and Hermawan 2018).

Based on the results, several recommendations can be made. First, the implementation of DL models shows strong potential in improving credit evaluation processes for students, a group that often lacks traditional credit histories. DL’s ability to capture complex, non-linear relationships makes it particularly well-suited for this purpose. Second, the integration of predictive modeling into financial advisory services at universities should be considered. Applications may include the use of loan scenario simulations, personalized financial planning tools, and the development of student-specific credit scoring systems, all of which can enhance financial decision-making and early risk identification.

Despite the promising results, this study is not without limitations. These include the limited representativeness of the sample, the absence of longitudinal (time-series) data, and the exclusion of potentially relevant socio-psychological variables. Addressing these limitations is essential for improving model generalizability and real-world applicability.

Future research should focus on extending the current models to incorporate temporal dynamics, integrating data from multiple sources (such as academic performance, behavioral indicators, and financial transactions), and evaluating the long-term impacts of predictive interventions. Such advancements will not only refine the precision and robustness of loan eligibility models but also inform the design of more inclusive and sustainable student loan policies.

Ultimately, by pursuing these directions, we aim to contribute to the development of intelligent, equitable, and resilient financial systems that support educational access and long-term financial well-being for students.