The credit risk assessment remains a critical challenge for financial institutions as manual and semi-automated

loan evaluation processes often produce inconsistent decisions, high default rates, and operational inefficiencies.

This paper discusses an Integrated Data-Driven Loan Management Framework that comprises an Artificial

Neural Network (ANN) for credit default prediction, Structured Query Language (SQL) to systematically extract

and transform data, and Interactive Visual Analytics Dashboards to aid in providing transparency within the

decision-making process. For the study, a public loan dataset was used to pre-process the target dataset containing

a total of 38,577 records and 24 attributes with pre-processing techniques such as feature engineering, one-hot

encoding and class re-balancing with the use of Synthetic Minority Oversampling Technique (SMOTE) yielding

43 model-ready inputs.

The final ANN architecture was developed with 2 hidden layers (43 and 21 neurons with ReLU activation) and

better decision-making and stakeholder transparency than using a model in isolation; therefore, it serves as a

proof-of-concept prototype for intelligent loan management in banks and other financial institutions.

Keywords: Credit risk assessment, artificial neural network, loan default prediction, SMOTE, explainable AI.

The assessment of credit risk is an essential component of establishing sustainable lending practices. Financial

institutions must assess the likelihood that a borrower will meet their obligations to repay loans, as opposed to

assessing those borrowers who will default on their loans. Non-performing loans (NPLs) continue to represent

a substantial financial drain for banks and a major threat to the stability of the international financial system

(Altman, 1968). Many developing countries continue to rely on the manual assessment process performed by

loan officers when establishing a borrower’s creditworthiness. The result is that decisions are made based on

failing to account for the fact that there are complex and nonlinear relationships between the various

characteristics of the borrower that affect their likelihood of repaying a loan (West, 2000). Recently, machine

learning (ML) techniques, specifically Artificial Neural Networks (ANNs), have demonstrated a significantly

improved level of predictive accuracy over traditional statistical modeling methods for the classification of credit

risk (Bakpo & Kabari, 2009; Chuang & Huang, 2011; El Khair Ghoujdam et al., 2024; Hsieh, 2005; Peng & Tu,

2005). ANNs can learn hierarchical feature representations and nonlinear decision boundaries from data.

Despite these advances, several gaps remain in the literature and practice. The most notable gap is that a large

majority of studies have evaluated machine learning (ML) models in an isolated manner, simply presenting

classification metrics without providing any contextual information (e.g., complete decision support workflows)

within which the model resides (Rudin, 2019).

Assessment using ML, with the following contributions: (i) Data pipeline integration: SQL-based extraction is

combined with Python-based preprocessing, including encoding, scaling, and SMOTE rebalancing, to create a

reproducible and auditable data pipeline; (ii) Using an Artificial Neural Network (ANN) to create, train, and

evaluate a feed-forward neural network model for the purpose of predicting loan defaults, yielding a prediction

accuracy of 87% on the held-out test set; (iii) Providing explanatory support to end-users through the generation

of per-prediction SHAP explanations, thereby alleviating the interpretability problem associated with using

neural networks to create a predictive credit model; (iv) Providing aggregate insight into a portfolio through the

use of Tableau dashboards (including approval rate, funded amount, default distribution by loan grade, purpose,

and geography); these dashboards complement the SHAP-generated per-instance explanations; (v) Delivering a

complete decision-support system by packaging the framework as a Flask web application with a user-friendly

The Z-score model designed by Altman (1968) is known as one of the earliest mathematical methods to assess

corporate credit risk and bankruptcy likelihood. This early attempt was based on the use of several financial

ratios through a type of statistical technique known as multiple discriminant analysis (MDA).

With continued research, logistic regression has become the most commonly used method for creating credit

scorecards because of its ease of interpretation and because it is widely accepted by regulatory agencies (Chawla

et al., 2002). However, due to the linear nature of MDA and logistic regression models, they are not always able

to appropriately explain nonlinear interactions among the different features in a borrower's behavior data (West,

2000).

B. Neural Networks for Credit Risk

Another popular method of evaluating creditworthiness is the use of neural networks. A study performed by West

Hsieh (2005) further explored this area and developed a combined method of using self-organizing maps (SOMs)

and a second method by using K-means clustering for preprocessing the data before that data were fed into an

ANN, reporting improved accuracy after outlier reduction. Peng and Tu (2005) integrated a self-organizing map

to assess credit worthiness in Nigeria's banking sector and determined that using neural networks resulted in

improved credit evaluations compared with traditional, statistical models, but without incorporating

explainability or visualization components.

C. Addressing Class Imbalance

A typical loan data set is an imbalanced data set where the default representation is only a small number of

instances. Chawla et al. . (2002) proposed the synthetic minority oversampling technique (SMOTE) as an

interpolation method for generating additional synthetic data points in the minority class. The use of SMOTE

and its variants has become a standard preprocessing technique in credit risk modeling (García et al., 2015;

Rudin, 2019).

D. Explainable AI in Finance

The lack of clarity in neural networks has been highlighted as one of the major obstacles limiting their adoption

within the finance sector (James et al., 2013).

Lundberg and Lee (2017) introduced SHAP, which provides theoretically grounded local feature attributions

backpropagation neural network for early credit risk prediction in commercial banks and reported better

performance than logistic regression, but the study did not address full end-to-end system integration.

The literature reviewed suggests that credit risk research has predominantly focused on the accuracy of models,

with little regard for a) the entire data collection/processing/decision cycle, b) the accessibility of per-prediction

explanation to users who are not familiar with the technical aspects, and c) visualisation tools for reviewing

credit portfolio performance. This research addresses all three of these gaps in a single framework.

Figure 1 illustrates the overall architecture of the proposed explainable credit risk decision-support framework.

A publicly available dataset from LendingClub, obtained via Kaggle

(https://www.kaggle.com/datasets/datawitharyan/financial-loan-dataset), is used. There are 38,577 records in the

raw loan dataset, and these records contain 24 attributes, including the demographic information on borrowers,

Loan default prediction suffers from severe class imbalance because defaulted loans constitute a relatively small

proportion of all observations. To address this problem, the Synthetic Minority Oversampling Technique

(SMOTE) was applied exclusively to the training dataset using the equation 1 (Chawla et al. 2002). Given a

minority instance x

x

= x

+ λ

x

− x

narrowing scheme, as the first hidden layer is equal to the input layer (43 dimensions) so that when the data is

transformed through this layer there is sufficient rank (full rank), which provides maximum possible

transformation of the feature space. The second hidden layer contains 21 neurons (21 ≈ 43/2), approximately

half of the input dimension, allowing the network to compress the learned representation before classification.

• The sum of loan applications and comparison to previous month and month to date.

6: Apply one-hot encoding to categorical variables

15: For each applicant do

Predict default probability

Compute SHAP values

Identify Top-3 influential features

Generate explanation

End For

16: Store trained model (.keras)

17: Store feature metadata (.json)

18: Deploy model using Flask

19: Accept applicant information

21: Predict loan status

reached 87% by epoch 50. Convergence is consistent and shows no evidence of catastrophic overfitting, though

the absence of a separate validation curve indicates that future work should incorporate early stopping.

The model was evaluated on a held-out test set of 12,858 samples, and the classification metrics are reported in

Table 2. The overall accuracy was 87%, with a near-symmetric precision-recall trade-off (0.86–0.88) across both

classes, indicating that the SMOTE technique successfully mitigated the original class imbalance.