Regulation-Aware Generative Intelligence for Finance

Abstract

Trust and explainability are foundational to generative financial intelligence, shaping its use in fraud detection, risk forecasting, and regulatory decision-making. Our approach to fraud analysis applies generative classification methods to rich, noisy feature spaces, with fraud indicators framed along three dimensions: modeling objectives, feature spaces, and label structures. We achieve transparency, interpretability, and causal reasoning through a combination of fuzzy masking, a two-stage training process, and concept activation vectors. For dynamic risk forecasting, we use generative probabilistic time series models to capture state transitions within a responsible AI framework. The resulting risk-signal pipeline preserves data provenance and upholds ethical AI principles, and we demonstrate its transparency and regulatory alignment through application to both risk-averse and risk-seeking portfolios under a narrow banking regime.

Building on this foundation, generative financial intelligence also enables scenario generation and what-if analysis for financial decision-making. We model decision-maker responses to scenarios using a neural proxy, yielding a context-conditional decision-support framework that is model-agnostic and compatible with any financial scenario generator. This framework exposes predictive uncertainty via Monte Carlo sampling; while sampling efficacy declines under disturbance-prone models, the approach remains robust to a degree of random-label noise and shows partial robustness to adversarial perturbation. Together, our three research directions — crime prevention, risk governance, and decision-making regulation — draw on generative modeling, trustworthiness, and regulatory awareness, applied either jointly or in sequence.