GANs and LLMs transform credit risk and fraud detection in finance

Generative artificial intelligence is poised to reshape how banks and other financial institutions predict risks, according to a new peer-reviewed study published in Information. Their research underscores that the strategic use of GAN-based data augmentation and LLM-driven text analytics can not only improve predictive accuracy in risk modeling but also address long-standing challenges in model interpretability and compliance.

The study, titled “Application of Generative AI in Financial Risk Prediction: Enhancing Model Accuracy and Interpretability”, evaluates how generative AI methods can outperform traditional econometric approaches by handling nonlinear, high-dimensional datasets that often undermine legacy credit-scoring and fraud-detection tools. The work highlights that while generative AI can close performance gaps in high-risk decision-making, its integration must be managed to meet data-security, cost, and ethical standards.

Generative AI reduces data acarcity and enhances model accuracy

The study primarily focuses on the application of Generative Adversarial Networks (GANs) to overcome data scarcity and class imbalance, two persistent problems in financial risk modeling. Traditional models often fail to capture the full range of risk factors because they are trained on datasets that under-represent rare but critical events such as defaults or fraudulent transactions. By generating synthetic credit and transaction records that mirror the statistical patterns of real data, GANs enriched the training base for risk models.

The researchers reported that incorporating GAN-generated samples into credit-risk classifiers raised the area under the curve (AUC) score from 0.82 to 0.88 and boosted the minority-class F1 score from 0.45 to 0.58. These gains translate into stronger sensitivity to default cases and more balanced risk predictions. Moreover, the generative augmentation improved probability calibration, yielding more reliable risk scores for financial decision-makers.

In fraud detection, the team demonstrated that combining traditional structured transaction data with Large Language Model (LLM)-based semantic features captured from transaction descriptions significantly improved the recall rate for identifying fraudulent activity. Fine-tuning a pretrained BERT model on unstructured text raised recall from 60 percent to 85 percent, highlighting how generative AI’s capacity to process textual data complements conventional numerical models.

Generative AI also proved effective in market-risk early warning. The hybrid use of GANs to simulate extreme market scenarios and LLMs to analyze sentiment in financial news improved stress testing and revealed a strong correlation between sentiment signals and market volatility. These results suggest that generative models can strengthen institutional preparedness for unexpected market events.

Addressing interpretability, data security, and deployment challenges

While the performance improvements are notable, the study emphasizes that interpretability and data protection remain crucial barriers to adoption. Financial institutions are subject to strict regulations under frameworks such as the European Union’s General Data Protection Regulation (GDPR), anti-money-laundering protocols, and know-your-customer requirements. Complex generative models risk becoming opaque “black boxes,” complicating regulatory audits and internal risk governance.

To address this, the researchers introduced a Generative Adversarial Explanation (GAX) framework, combining GAN-based counterfactual data generation with SHAP values to highlight the most influential features behind model predictions. This approach enhances transparency for both regulators and internal compliance teams without sacrificing performance.

The study also evaluated the use of differential privacy to protect sensitive financial information. By introducing controlled noise into the training process, institutions can limit exposure of individual data points while maintaining statistical utility. The research quantified a privacy–utility trade-off, observing that stronger privacy settings slightly reduced model accuracy but still delivered significant performance improvements over legacy methods.

Deployment complexity and training costs were also flagged as critical challenges. Large models require powerful computing infrastructure and specialized personnel, creating financial and operational hurdles for smaller institutions. The authors suggested cost-saving measures such as transfer learning, which fine-tunes only the final layers of pretrained models, and model compression techniques like knowledge distillation to reduce computational demands. They also highlighted the potential of federated learning, which enables collaborative model training across institutions without sharing raw data, thereby improving both privacy protection and cost-efficiency.

Future directions for responsible and scalable adoption

While generative AI has the potential to transform risk prediction, its benefits will only be realized through careful management of technical, regulatory, and ethical factors. The authors recommend prioritizing the development of more interpretable generative models, including attention-based visualization techniques and transparent scoring for synthetic data. Such advances can help address regulators’ demand for explainable decision-making in credit approvals and fraud alerts.

Integrating blockchain technology with generative models is identified as a promising next step. Blockchain’s immutable and decentralized structure can help safeguard model integrity and ensure transparent record-keeping in risk assessments. The researchers also encourage exploring generative AI applications beyond risk detection, such as portfolio management, where synthetic scenarios can test asset performance under diverse market conditions, and quantitative trading, where sentiment-driven signals from LLMs can support investment strategies.

Equally important is the establishment of comprehensive data governance frameworks. The authors advocate a three-tier system in which data owners, model developers, and independent auditors share responsibility for privacy-by-design practices and continuous monitoring. This approach can help institutions strike a balance between innovation and accountability, preventing misuse of synthetic data or generative tools for fraudulent reporting.

The research highlights that training-data quality remains the backbone of reliable predictive models. Even with sophisticated generative tools, flawed or biased input data can propagate errors and unfair outcomes. Effective preprocessing, including noise reduction, outlier handling, and equitable sampling, is therefore essential to sustaining the credibility of generative AI systems in finance.