An Intelligent Blockchain-GAN Framework for Risk Management in International Trade Finance

Theoretical and Conceptual Framework

This study’s theoretical foundation is derived from three interconnected fields: supply chain finance (SCF) frameworks, financial risk management theory, and technical trust mechanisms found in generative modeling and blockchain. According to financial theory, risk management is the ongoing evaluation and reduction of uncertainties that impact credit exposure, transaction authenticity, and liquidity in international trade networks. This is furthered by SCF theory, which places a strong emphasis on the cooperation of customers, suppliers, and financial institutions in order to maximize working capital and lower financing friction through open data sharing. According to this theoretical framework, blockchain technology is represented by Hyperledger Fabric embodies the concepts of immutability, decentralization, and trustless consensus, which are consistent with the theories of information transparency and transaction-cost economics. Concurrently, dual neural networks (generator and discriminator) progressively improve the precision and robustness of synthetic data modeling in Generative Adversarial Networks (GANs), which are based on adversarial learning theory. Rare, dishonest, or fraudulent financial behaviors that are usually underrepresented in real datasets can be simulated with the use of this dual architecture. The conceptual justification for creating a blockchain–GAN-based SCF system that improves dynamic risk assessment, transparency, and forecasting accuracy in international trade finance contexts is provided by integrating these theories.

Data Pre-processing and Encoding

To prepare the finance market information for research and instruction in the deep learning system, the first steps are data cleaning and encoding. In order to ensure consistency and increase model resilience, the two datasets were first equalized by removing outliers and missing values. To guarantee the quality, consistency, and suitability of the data provided by the supply chain finance (SCF) system for input into the GAN-based risk enhancement system, the first phase of the intended structure involves preprocessing, such as cleaning and encoding.

Standardization of SCF Datasets

SCF databases, which are frequently found in a variety of formats, may include invoice numbers, payment plans, interest amounts, and delays in shipment. Using normalization of z-scores and diagonal bias during training of models.

x ′ = x − μ σ

(1)

here x is the feature's starting value, μ is its mean, and σ is its standard deviation. This conversion ensures that all features maintain a unit variation, a zero mean, and a degree of indifference to variations in model scale.

Encoding Categorical and Temporal Attributes

Categorical features like country, currency, and transaction type are processed through one-hot encoding. Let C be a categorical variable with n distinct values. It will be represented as a binary vector v ∈ { 0 , 1 } n , where one component is 1 and all the rest are 0.

OneHot ( C i ) = v i = [ 0 , 0 , … . , 1 , … . , 0 ]

(2)

Temporal features such as transaction timestamps are mapped to cyclic phases to exploit periodicity (e.g., month, quarter). For a temporal feature t∈[0,T), we use:

sin enc ( t ) = sin ( 2 π t T ) , cos enc ( t ) = cos ( 2 π t T )

(3)

This enables the model to capture periodic financial phenomena like emerging patterns tied to recurring fiscal cycles or seasonal risks.

Cleaning and Normalizing Financial Records for GAN Input

Data inconsistencies like missing values x_i,j = NaN, outliers, or non-numeric entries are addressed through imputation and outlier removal strategies. Mean imputation performs:

x i , j = 1 N ∑ k = 1 N x k , j if x i , j = NaN

(4)

For skewed numeric data, min-max scaling can be applied:

x ′ = x − min ( x ) max ( x ) − min ( x )

(5)

Thus, the preprocessed dataset X∈R^n×d (n being the number of trade finance records and d the number of encoded features) is ready to be utilized as an input to the GAN for synthetic data generation.

Risk Augmentation via GAN

In the second phase, Generative Adversarial Networks (GANs) enhance risk modeling by creating synthetic scenarios in trade finance, especially those with infrequent or fraudulent activities that do not have adequate representation in reality. Each GAN consists of two components, a Generator G and a Discriminator D, both of which are neural networks engaged in a minimax game:

min G max D V ( D , G ) = E x ~ p data ( x ) [ log D ( x ) ] + E z ~ p z ( z ) [ log ( 1 − D ( G ( z ) ) ) ]

(6)

The Generator and Discriminator operate in a feedback loop where: G(z;θ_G) learns to create realistic-looking trade finance records from random noise inputs z∼N(0, I).D(x;θ_D) learns to separate x∼p_data from G(z)∼p_G.

Through this iterative training process, G becomes adept at producing x_syn = G(z) that captures the statistical patterns present in genuine financial transactions. In this case, fraudulent transactions or edge-case financial risks are simulated by adding to the training set with augmented minority class synthetics (e.g., payment delays, over-invoicing, double financing) using label-agnostic patterns set by control grids (cGANs) such as Conditional GANs.Where y for the cGANs represents class labels assigned to risk scenarios such as “fraud,”“default,”“delay” for specific labeled risks.

min G max D V ( D , G ) = E x ~ p data ( x | y ) [ log D ( x | y ) ] + E z ~ p z ( z ) [ log ( 1 − D ( G ( z | y ) ) ) ]

(7)

Following the training, the GAN generates a new set of samples as { x 1 syn , x 2 syn , … . . , x m syn }. These samples are combined with the real samples { x 1 real , x 2 real , … . , x n real } in order to create a stronger and balanced dataset:

D aug = D real ∪ D syn

(8)

Where this combined dataset is D aug ∈R^(n+m)×d.Then used this dataset to train downstream machine learning models for predicting risk in financial systems, to improve generalization to unseen or rare patterns.

Algorithm 1: Risk Augmentation Using GANs

Input: SCF real dataset Dreal and random noise prior z∼N(0, I)

Step 1: initialize Generator G(z; θ_G) and Discriminator D(x; θ_D)

Step 2: While not converged, repeat:

Step 3:Draw from trade records x ∼ D real

Step 4: From noise distribution, generate z∼p_z(z)

Step 5: Generate synthetic sample as x_syn=G(z)

Step 6: Update D to maximize: use Equation 6

log D ( x ) + log ( 1 − D ( G ( z ) ) )

Step 7: Update G to minimize: use Equation 7

log ( 1 − D ( G ( z ) ) )

Step 8: Model Output: G trained for deriving D_syn.

Step 9: Use Draw Update Augmented Dataset as:

D aug = D real ∪ D syn use Equation 8

Step 10: Model Final Classifier on D aug for risk-robust prediction datasets.

This approach to GAN augmentation of the dataset builds resilience to the data gap problem in fraud and anomaly detection, thus enhancing overall detection capabilities of the models in cross-border and high-risk international trade finance transactions.

Risk Augmentation via GANs process functions within a sophisticated step-by-step algorithm 1 that improves the ability of predictive models to detect risks in international trade finance. As the initial step, a real-world SCF (Supply Chain Finance) dataset, D_real, is available as a starting point for the training process. Simultaneously, a random noise vector z is drawn from a normal Gaussian distribution N(0, I) to work as an input for the generator of the GAN. The structure of a GAN has two major parts: A Generator G and A Discriminator D. The Generator attempts to produce realistic trade finance samples targeting certain statistical properties of authentic trade transactions, including, but not limited to, iterative or fraudulent patterns, such that the synthetic samples generated trade finance samples are expressed as x_syn = G(z). The Discriminator combines both real data and the generated models and characterizes them as levels of reality. It then alters its parameters to maximize the number of real data and the number of fake data correctly classified as real and synthetic.

The generator and discriminator are trained in an adversarial loop, where G learns to fool D, and D learns to become more accurate in detection. This is governed by a minimax optimization process. Over successive iterations, G becomes proficient in producing realistic synthetic data that replicates the intricate sophistication of real trade finance transactions, including edge cases such as payment defaults or over-invoicing. After the training converges, the generator is then used to produce a synthetic dataset D syn , which is then combined with the original dataset D aug = D real ∪ D syn to form an augmented dataset.

Now, this fused dataset is richer in range and diversity of scenarios, including infrequent yet critical risk scenarios. This augmented dataset is utilized in training subsequent neural network ensembles or even decision tree ensembles to enhance the prediction accuracy of financial risk using a broad-based and flexible model. The main propositional advantage offered is the possibility of simulating and training on risk scenarios that would otherwise be underrepresented or not present in actual data, thus strengthening international trade finance risk management.

Blockchain Integration with Hyperledger Fabric

Integrating Blockchain with Hyperledger Fabric adds a new decentralized, secure, and verifiable layer for transacting business in Supply Chain Finance (SCF). Hyperledger Fabric is a permissioned blockchain framework aimed at enterprise-level applications where privacy, detailed access control, and trust are vital. The primary goal is to record all the SCF transactions on a distributed ledger and enforce smart contracts, or what Fabric calls chain code, to automate critical finance processes like invoice approval, payment release, and risk scoring.

Let’s denote each SCF transaction as a tuple SCF which covers Payer, Receiver, Amount, and Status. The Executed Transaction will look like this for n Transactions:

T i = ( P i , R i , A i , S i )

(9)

Where: P_i: Payer, R_i: Receiver, A_i: Amount, S_i: Status of transaction (pending, approved)

Every T_i transaction gets executed and logged through a smart contract C running on Hyperledger Fabric. Smart contracts enforce that an invoice must have business rules evidenced, which state as follows:

C ( T i ) : if InvoiceValidated ( T i ) and RiskScore ( T i ) ≤ δ ⇒ Approve ( T i )

(10)

Here, InvoiceValidated ( T i ) - performed by invoice recognition via OCR or matching on tokenized evidence. RiskScore ( T i ) - performed externally via a classifier augmented with Generative Adversarial Network (GAN) or other means.δ—predefined risk tolerance threshold.

LedgerStat e t + 1 = LedgerStat e t ∪ { Hash ( T i ) }

(11)

Upon validation, the smart contract issues the event log transaction on the blockchain.

Immutability ensures a decentralized trust system since any modification would require breaking cryptography, changing the hash, and the transaction record no longer equals the block state. For SCF transactions:

Hash ( T i ) = SHA 256 ( P i | | R i | | A i | | S i )

(12)

This ensures immutability, since any tampering would change the hash, and the transaction would no longer match the recorded block state.

SCF Compliance Modelling and Real-Time Risk Evaluation

SCF Compliance Modeling and Real-Time Risk Evaluation involves the trade finance events’ dynamic linking with risk analytics to enable continuous compliance, as well as proactive financial exposure mitigation. This process starts with associating each incoming SCF transaction T_i and a risk score r_i, which is computed based on real-time factors such as payment delays, invoice mismatches, and contract fulfillment deviations. This may be expressed as:

r i = ϕ ( T i , C i , H i )

(13)

Where: T_i is the transaction data, C_i the execution outcome of the corresponding smart contract, H_i the historical behavior features or heuristics, and ϕ the risk evaluation function which normalizes score r_i∈[0,1].

Once r_i is calculated, it is checked against a moving threshold δ(t), which in some cases changes due to market turbulence or company risk appetite volatility. If the condition r_i>δ(t) is met, the system goes into alerting, stops further payments, or physically routes the transaction for manual processing. In addition, real-time risk assessment can be performed by the integration of smart contract logs with an analytics engine that ingests streaming data from the blockchain. This facilitates the perpetual containment of exposure monitoring concerning all participants. The financial exposure E_j of an entity j at time t is estimated as:

E j ( t ) = ∑ i ∈ T j w i . r i

(14)

Where: T_j (t) is the set of active transactions for entity j, w_i is the transaction value or weight, r_i is the risk score of transaction. This bounded flexibility metric ensures SCF operations undergo full compliance in governed layer frameworks while being operationally unilateral, resilient, and transparent against operational deception, fraud, or manipulation. Trust is enhanced amongst stakeholders while vulnerabilities to financial stocks are further reduced.

The integrated architecture illustrated in Figure 2 of blockchain and GAN technologies for implementing Supply Chain Finance (SCF) systems is aimed at the automation, security, and risk management of trade transactions. The workflow initiates at the point where a buyer or supplier requests a trade by submitting an invoice, which forms the SCF pipeline. An invoice must pass validation checks and SCF protocols to ensure compliance and authenticity. An audit trail and compliance log are generated in parallel, capturing process transparency. Validation permits immutably recording the transaction in a blockchain ledger on Hyperledger Fabric. Precise timestamping renders the data unchangeable while the ledger remains accessible. Concurrently, triggering of smart contracts facilitates receipt automation and credit decision on autonomous rule-based systems.

OPEN IN VIEWER

Figure 2.

Workflow of Blockchain and GAN integration in SCF processes, showing smart contract execution, risk scoring, and synthetic data augmentation on Hyperledger Fabric.

Smart credit approval shifts the process to an execution phase of smart contracts, which is followed by payment settlement, completing the financial transaction. Concurrently, Generative Adversarial Networks (GANs) receive the artificial but realistic transaction records as inputs to generate synthetic trade data. The GAN's capacity to replicate complex trade scenarios is enhanced by ongoing instruction revisions that include synthetic data and feedback samples. Both polished synthetic and actual data are processed by a Risk Detection Module, which has a discriminator and a classifier to detect fraud or high-risk transactions. These types of transactions are examined in greater detail by a risk assessment module, which analyzes the risk, measures its levels, and offers additional recommendations. In the end, all transaction data—both synthetic and real—is encoded and merged into a single dataset that is ideal for machine learning, reporting, and analysis. This process enables data-driven, intelligent, and SCF settings.

Data Set Description

The proposed work is based on the dataset concept and structure from the base paper which centers on the cryptocurrency transaction data for financial risk analysis. Specifically, the base paper utilized cryptocurrency market data from 2017 to 2020 which was retrieved from CoinMarketCap (Zhang et al., 2025) and consisted of 61 cryptocurrencies with 10,000 records in total. This dataset contained transactional attributes such as the name of the cryptocurrency, daily price, risk indicators, trading volumes, and temporal market activity. This work aims to extend the base paper by incorporating its dataset of cryptocurrencies, which initially included 10,000 records from 61 digital assets collected from CoinMarketCap between 2017 and 2020. This dataset containing pricing, trading volumes, and risk metrics is the construction and behavior basis for the synthetic datasets in the proposed study. In particular, this framework of cryptocurrency data is employed to train Generative Adversarial Networks (GANs) to simulate synthetic scenarios of trade finance, including extreme financial events that are rare and high-risk. With the proposed method, two main datasets are employed: real-world SCF datasets containing transactional data and risk attributes, and cryptocurrency-transaction-behavior GANs-simulated synthetic data. This amalgamation improves the reliability and comprehensiveness of the international trade finance risk prediction system. The dataset variables description is given in Table 2.

OPEN IN VIEWER

Table 2.

Supply Chain Finance (SCF) Dataset Variables and Their Categorical or Numerical Attributes.

Field Name	Description	Data Type
Transaction_ID	Unique identifier for trade transaction	Categorical
Buyer_ID / Seller_ID	Participants in the SCF transaction	Categorical
Invoice_Amount	Amount of invoice	Numerical
Currency	Currency of the transaction	Categorical
Transaction_Type	Type: Export, Import, Invoice Factoring, etc.	Categorical
Risk_Label	Binary/Multiclass: Low, Medium, High risk	Categorical
Risk_Score	Score from GAN Discriminator	Float (0–1)
Payment_Status	Paid / Overdue / Pending	Categorical
Country_Origin	Origin country	Categorical
Country_Destination	Destination country	Categorical
Transaction_Date	Date of transaction	Date
Blockchain_Hash	Immutable hash of the transaction	Categorical

Because it simulates actual transactional behavior like invoice amounts, payment delays, trade regions, and risk labels—key factors influencing credit and operational risk in global trade ecosystems—the dataset used in this study has a direct bearing on the fields of international trade finance and supply chain finance (SCF). These characteristics are indicative of real-world supply chain interactions, where liquidity risk and knowledge asymmetry are most common. The design of the dataset is consistent with earlier SCF studies (e.g., Li et al., 2023; Zhang et al., 2025) that simulate financial exposure and fraud detection using transactional and risk-indicator properties.

We calculated several similarity and diversity measures between generated and real samples in order to statistically assess the quality of the synthetic data created by the GAN. In particular, close statistical alignment was validated by the Fréchet Distance (FD = 6.41) and Maximum Mean Discrepancy (MMD = 0.013 ± 0.004), suggesting that the synthetic data maintained diversity while preserving the distributional features of the original SCF data. The validity of data fusion for downstream model training was further supported by pairwise feature correlations (average Pearson r = .93 between real and synthetic sets) and entropy-based diversity indices (H = 0.89), which confirmed that GAN augmentation did not overfit or collapse into limited modes.

Experimental Setup

The dataset included Supply Chain Finance (SCF) transaction records with 12 categorical and 34 numerical characteristics, including risk label, trade region, invoice value, and payment delay. Data pretreatment included removing duplicates and outliers that were more than three standard deviations (>3σ), as well as one-hot encoding of categorical variables and Min–Max scaling of numerical features. A Generative Adversarial Network (GAN) was used to create the synthetic data, and only samples that obtained a discriminator confidence score of at least 0.8 were kept and combined with the original dataset. To ensure robustness, the pooled data was divided into 70% training, 15% validation, and 15% testing sets. Five-fold stratified cross-validation was then performed across three randomly selected seeds (42, 99, and 123). The model’s performance was assessed using ROC-AUC, accuracy, precision, recall, F1 score, and false positive rate. Bootstrap resampling was used to calculate the 95% confidence intervals (n = 1,000). Lastly, improvements above baseline machine learning models were statistically confirmed using McNemar’s tests and paired t-tests, with significance established at p < .05. The Simulation and Execution was shown in Table 3.

Hardware and Software Configurations

OPEN IN VIEWER

Table 3.

Simulation and Execution Environment for the Proposed GAN–Blockchain Framework.

Parameters	Values
Programming language	Python
Operating system	Ubuntu 22.04 LTS
Hardware	Intel Xeon 32-Core CPU, NVIDIA RTX 4090 GPU (24GB), 128GB RAM
Frameworks and libraries	PyTorch 2.3.0, scikit-learn 1.5, Hyperledger Fabric SDK (Node.js v2.5), Docker 24.0
Random seed control	NumPy, Torch, OS, and CUDA random states fixed to ensure determinism
Data source	Cryptocurrency APIs and CSV files
Modelling algorithms	GAN
Cross-validation strategy	70% training, 30% testing with time-series split
Evaluation tools	RMSE, MAE calculators, and confusion matrix tools for classification tasks

Performance Metrics

A set of quantitative evaluation methods is used to assess and classify the financial risks of an international trade finance system. These methods include:

This measure attempts to calculate the maximum loss that may occur to a certain portfolio within a specific time frame and confidence level.

Va R α = inf { l ∈ R : P ( L > l ) ≤ 1 − α }

(15)

Here, L is—loss random variable, α is—confidence level.

It is also referred to as Expected Shortfall, and it measures the loss that exceeds the value of the stated level of the threshold.

CVa R α = E [ L | L > Va R α ]

(16)

Useful primarily for trade failure tail risk analysis.

In a GAN framework, the discriminator’s confidence score D(x) can be seen as a risk score for the application of a synthetic versus real transaction filter.

Risk Score ( x ) = 1 − D ( x )

(17)

D(X) = 0, wherein x is a transaction vector input, and the closer D(x) is to 0, the more anomalous, hence more risk.

To assess the effectiveness of smart contracts in Hyperledger Fabric SCF systems

BLI = Block Commit Time − Transaction Submit Time Total Transactions Processed

(18)

It demonstrates the flair of SCF protocol to tickle risk quickly and appropriately receive communication for action.

The comparison of GAN-based models with other classifiers employs the standard metrics of machine learning evaluation:

Accuracy = TP + TN TP + FP + TN + FN

(19)

GAN-based models are compared against other classifiers based on the standard machine learning metrics for evaluation. These include:

F 1 − score = 2 × Precision × Recall Precision + Recall

(20)

Where TP, FP, TN, and FN are true positives, false positives, true negatives, and false negatives, respectively.

Performance Evaluation of Proposed Model

To calculate Value at Risk (VaR), a financial or synthetic trade dataset's daily value distribution is shown in Fig 3. As can be seen, the distribution appears to be roughly normal because most of the data are grouped around the mean. The red dashed line at −0.0295 represents the 95% VaR, which indicates that there is a 5% chance that daily losses will exceed 2.95% of the capital. The orange dashed line at -0.0406 represents the 99% VaR, which indicates a 1% chance of losses higher than 4.06%. These cutoff points are essential estimates that aid traders and investors in planning and evaluating scenarios where losses are anticipated to be greatest under normal market conditions. The use of financial swaps, exposure to credit risk, and capital needs are all influenced by this Figure 3.

OPEN IN VIEWER

Figure 3.

Value at Risk (VaR) distribution showing 95% and 99% thresholds for portfolio loss estimation.

The suggested approach employs Generative Adversarial Networks (GANs) to fabricate realistic but extreme financial trading scenarios, while providing auditable and tamper-proof records of transactions with SCF protocols based on Hyperledger Fabric. The augmentation provided by GANs improves measures of risk such as CVaR, as illustrated in the Figure 4. Edged-case risk data is enriched through the aforementioned GANs, which identify complex tail risks that traditional models fail to capture. On the other hand, the unchangeable and transparent blockchain guarantees data integrity, which strengthens the estimation of risk. The proposed system demonstrates a reduction of CVaR value, indicating better resilience to extreme losses compared to traditional statistical models (e.g., −0.13 becomes 0.1063). This illustrates improving management of downside risk in international trade finance using AI combined with distributed ledger technologies, thus showcasing AI and DLT synergy. The synthesis model data utilized for trade finance returns CVaR is illustrated in Figure 4 with ranges of approximately −0.16 to 0.18 as lower and upper bounds, respectively. VaR (Value at Risk) at 95% confidence and equal to −0.0863 is the red dashed line. This means in the five worst cases, returns are expected at or below −8.63%.CVaR (Conditional Value at Risk), average loss in the worst five scenarios, as shown with an orange dashed line at −0.1063. This is the average loss for tail risk defined beyond the Value at Risk threshold.The kernel density estimator accurately portrays the probability distribution of returns in question; as such, it is overlayed on the histogram.

OPEN IN VIEWER

Figure 4.

Conditional Value at Risk (CVaR) results highlighting tail-risk reduction through GAN augmentation.

In Figure 5 shows a risk scores produced by the discriminator of a GAN for simulated financial transaction data from four cryptocurrencies: Bitcoin (BTC), Ethereum (ETH), Litecoin (LTC), and Ripple (XRP). The risk scores capture the discriminator's assessment of whether a financial transaction is anomalous or possibly fraudulent. The scores indicate the level of confidence assigned by the machine to the identification of risk, wherein 0 represents low risk and 1 denotes high risk. Each colored bar shows the frequency distribution of unique transactions within one dataset and demonstrates the distribution of transactions for each risk category.

OPEN IN VIEWER

Figure 5.

GAN discriminator risk score distribution across asset classes (BTC, ETH, LTC, XRP).

The distribution appears to be around normal (bell-shaped) and clustered around 0.5, indicating that the GAN discriminator seems to be set to assume that most transactions are of medium risk, with far fewer assigned as extremely low or high. Identical characterizations in the dataset’s shape and spread such as BTC having more medium-risk scores, XRP a slightly wider spread, illustrate variations in the digital asset ecosystems and how risk patterns manifest across different assets. This model’s ability to delineate transaction types demonstrates its discriminatory power, indicating that different asset classes could possess different risk paradigms. This is noteworthy when considering the expansion of the model into international trade finance risk through Distributed Ledger Technology (DLT) systems, such as Hyperledger Fabric. Understanding the effectiveness of visualization techniques about risk scoring models based on Generative Adversarial Networks (GAN) would serve as a critical component in financial system architectures.

In Figure 6 shows the blockchain latency index (BLI) for the years 2017 to 2020, using simulated transaction data. BLI tracks the average transaction latency (the time between submission and confirmation) within a set normalization of one hour. The graph’s individual data points correspond to the daily average BLI for the respective date which indicates the speed or slowness of transaction confirmations on blockchain's simulated network. The Y axis displays the BLI values, which oscillate between 0.03 and 0.04. This indicates that in the aggregate, most transactions recorded in this period experienced a 3 to 4% delay of an hour (1.8–2.4 min). The X axis represents the years January 2017 and December 2020. It suggests high daily variability and displays sporadic spikes possibly indicating bursts of increased congestion or processing delays, but, over time, there was no strong upward or downward trend. Therefore, it suggests constant network performance under the given simulated conditions.

OPEN IN VIEWER

Figure 6.

Blockchain Latency Index (BLI) trends from 2017 to 2020 showing transaction confirmation performance.

Table 4 shows the “Performance Comparison of ML Methods” illustrates the comparison in the performance of five machine learning models: GAN-F-SCF, LSTM, RNN, SVM, and NB on four metrics—Accuracy: Evaluation, Precision, Recall, and F1 Score with the confidence interval of 95%. The GAN-F-SCF model (Generative Adversarial Networks and Supply Chain Finance) outperforms the other methods in all metrics, achieving close to 98% to 99% in each. This particularly emphasizes its capability to identify and manage financial risks in international trade finance. The LSTM model performs second best, especially in Recall, but remains several points lower than the GAN-F-SCF. RNN performed moderately across all metrics, while SVM and Naive Bayes(NB) significantly lagged, especially in Recall and F1 Score, which are essential for risk prediction. With tight confidence intervals and the best mean performance across all measures, the suggested GAN-F-SCF model demonstrated robust generalization across random seeds and folds. These improvements over baseline models were statistically significant (p < .05), as proven by McNemar’s tests and paired t-tests.

OPEN IN VIEWER

Table 4.

Performance Comparison of ML Models with 95% Confidence Intervals (p < .05).

Method	Accuracy (%)	Precision (%)	Recall (%)	F1 Score (%)
GAN-F-SCF	98.42 ± 0.31	98.88 ± 0.27	97.91 ± 0.34	98.39 ± 0.29
LSTM [3]	92.34 ± 0.48	91.56 ± 0.52	93.24 ± 0.46	92.39 ± 0.44
RNN [3]	89.21 ± 0.57	90.78 ± 0.61	88.53 ± ± 0.59	89.64 ± 0.55
SVM [3]	85.67 ± 0.68	86.45 ± 0.65	84.76 ± 0.70	85.60 ± 0.63
NB [38]	80.45 ± 0.74	81.23 ± 0.77	79.34 ± 0.82	80.27 ± 0.76

Note. [3] refers to deep learning–based baseline models (LSTM and RNN) used for performance comparison, [38] denotes a classical probabilistic baseline model (Naive Bayes) included for comparative evaluation.

The existing work (Zhang et al., 2025) addresses the market risk of a cryptocurrency while applying Hierarchical Risk Parity (HRP) and Reinforcement Learning (RL) for managing portfolios of digital assets. It deals with the general financial market risk, including fraudulent activities like money laundering, and the risks associated with the collapse of centralized cryptocurrencies. While it utilizes classical machine learning approaches for active risk management and rebalancing, its integration with blockchain is shallow, and its implications for trade finance or supply-chain contexts are minimal.

Discussions

As for the proposed approach, it builds on the prior art by adding an AI and blockchain-based risk management system for international trade finance and supply chain finance (SCF), which is multi-layered. The present work proposes the use of Generative Adversarial Networks (GANs) not only for modeling but for generating synthetic data to simulate rare but critical financial risks (such as a fraudulent default). This fills the gap in the main paper that only included the most extreme values taken from real-world datasets.The proposed methodology uses blockchain technology and smart contracts to log SCF transactions. This ensures vital immutability, transparency, and confidence in international trade, where parties usually lack trust. The platform's automated compliance checks and dynamic risk-based credit rating enable immediate decision-making. The architecture of the suggested system is primarily driven by data preprocessing, blockchain transaction tracking, origin block enhancement, and risk-based SCF orchestration. These connections, which also boost its ability to grow and flexibility, make it much more relevant to organizations than investment assessments. The proposed study expanded on the initial premise of using crypto asset risk management to trade finance by incorporating technologies like Hyperledger Fabric and GANs for real-time transactional simulation, verification, and security. As a result, risk estimates were 12% to 17% accurate. For computerized commerce systems, this enhances control, uniformity, and confidence.