AI Model Flags Risky International Trade Transactions With Up to 98.9% Accuracy

Researchers have unveiled an artificial intelligence system that classifies international trade transactions as low, medium, or high risk with reported accuracies between 96 and 98.9 percent, combining an adaptive deep convolutional neural network with a swarm-inspired optimization algorithm modeled on the behavior of termite alates. The study, published in Discover Internet of Things on 16 October 2026, presents the model, called TAO-ADCNN, as a response to the growing complexity of cross-border settlement, where traditional rule-based verification struggles to keep pace with dynamic trading conditions.

The work was led by Renran Miao and addresses a persistent bottleneck in global commerce. As digital financial platforms and cross-border e-commerce expand, settlement processes have grown more intricate, and the authors argue that manual verification, paper-based records, and fixed rule-based applications are time-intensive, error-prone, and poorly suited to fast-moving transaction environments. Their proposed framework instead draws on Internet of Things data, including shipment tracking, GPS location, sensor readings, and delivery delays, alongside financial features such as transaction value, currency exchange rate fluctuations, document validation scores, and buyer-seller trust ratings.

At the core of the method is a supervised multi-class classification task. Each transaction in the dataset carries one of three predefined risk labels, and the adaptive deep convolutional neural network learns hierarchical representations from normalized transaction features while the termite alate optimization algorithm tunes the network’s parameters. The optimizer mimics the phototactic behavior of termite alates, in which individuals move toward the brightest locations and away from the darkest, with each candidate solution’s fitness determining its brightness. This attraction-repulsion scheme, combined with the periodic replacement of poorly performing candidates, drives the search through the high-dimensional space of hyperparameters, including learning rate, kernel size, number of convolutional layers, batch size, dropout, and weight initialization.

The neural architecture itself incorporates a class separability regularization term into its cost function, designed to reduce intra-class distance while increasing inter-class distance in the learned feature space. According to the paper, this encourages the network to produce features that cluster tightly around their class centroids and separate cleanly from other risk categories, which the authors link to more reliable discrimination between genuine and suspicious transactions. The network was trained with the Adam optimizer and categorical cross-entropy loss, with convolutional parameters adaptively optimized by the termite algorithm during training.

The evaluation relied on the IoT Trade Settlement Risk dataset, a synthetic collection of 9,990 samples obtained from Kaggle that simulates IoT-enabled trade settlement scenarios. It contains transaction-related, logistics, payment, and IoT monitoring attributes, including GPS variation, temperature fluctuations, shipment delays, and tampering signals, as well as payment behavior indicators such as credit score, payment delay, and dispute history. Before training, the researchers applied Min-Max normalization to place features on a common scale, Savitzky-Golay filtering to smooth transaction signals while preserving peaks and trends, and principal component analysis to reduce dimensionality and highlight the most significant trade indicators. Critically, all preprocessing transformations were fitted on the training data only and then applied to the test set to prevent data leakage.

Experiments were implemented in Python using a 70:30 training-testing split with five-fold cross-validation. Across the five folds, accuracy ranged from 97.8 to 98.3 percent, with a mean of 98.0 percent and a very low standard deviation of 0.0019, which the authors interpret as evidence of stable and consistent performance across different data partitions. On the full evaluation, the model achieved 98.45 percent accuracy, 98.90 percent precision, 98 percent recall, and a 98.41 percent F1-score on the ITS Risk dataset, with a reported ROC-AUC of 0.982.

The model was also benchmarked against a range of alternatives on a second dataset, the CoinMarketCap collection of 61 cryptocurrencies with 10,000 records. Competitors included support vector machines, recurrent neural networks, long short-term memory networks, a GAN-based approach, and ensemble baselines such as random forest, XGBoost, LightGBM, and CatBoost, all trained and evaluated on the same data. On the CoinMarketCap dataset, TAO-ADCNN reached 98.43 percent accuracy, 98.89 percent precision, 98.40 percent F1-score, and 97.92 percent recall. Paired t-tests and McNemar’s tests at the 95 percent confidence level, comparing the proposed model with the strongest baseline, GAN-F-SCF, produced p-values below 0.05 on both datasets, which the authors cite as evidence that the observed gains are statistically significant rather than the product of chance variation.

An ablation study supported the design choices, showing that each successive component, from Min-Max normalization through Savitzky-Golay filtering, PCA-based feature reduction, and TAO-based hyperparameter optimization, contributed incremental improvements in accuracy, precision, recall, and F1-score relative to a baseline version of the network. The full architecture achieved the best results, which the authors present as evidence that combining adaptive hyperparameter optimization with careful preprocessing benefits transaction risk categorization. Visualization analyses in the paper also examined buyer reliability against transaction frequency, temporal flows of IoT data events, and the relationship between transaction frequency and payment delay, describing how anomalous events appeared as localized clusters rather than random scatter over time.

The authors suggest that the predicted risk categories could support intelligent transaction verification and anomaly monitoring during settlement operations, and that risk classification outcomes could assist auditors and financial institutions in transaction verification and regulatory compliance, with prediction results reviewable to support traceability. However, the study is explicit about its boundaries. The model was validated exclusively through offline experiments on a benchmark synthetic Kaggle dataset, which the authors acknowledge may not fully capture the complexity and variability of real-world international trade settlement environments. Performance may be affected by noisy or previously unseen transaction patterns, and the system is sensitive to input feature quality and hyperparameter settings.

Notably, explainability analysis, runtime metrics such as training time, inference latency, throughput, and computational cost, scalability assessment, and deployment in operational trade settlement systems were all beyond the scope of the study. The authors state that future work will focus on validating the framework with real-world trade datasets, integrating explainable AI techniques to improve decision transparency, evaluating computational efficiency and scalability at large scale, and assessing deployment in real-time IoT-enabled settlement systems. Until such validation is carried out, the reported figures describe performance on simulated data rather than demonstrated benefits in live financial infrastructure, a distinction that matters for any institution considering the technology for actual settlement risk control.

Subject of Research: A deep learning model combining termite-inspired optimization with an adaptive convolutional neural network for classifying transaction risk in international trade settlement.

Article Title: AI and IoT driven intelligent risk control and verification system for international trade settlement

Article References: Miao, R. (2026). AI and IoT driven intelligent risk control and verification system for international trade settlement. Discover Internet of Things, 6(1), Article 214. https://doi.org/10.1007/s43926-026-00478-5

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00478-5

Keywords: artificial intelligence, Internet of Things, international trade, risk classification, deep learning, convolutional neural network, swarm intelligence, fraud detection, trade settlement, machine learning, financial technology, synthetic dataset