FRiskGPT: A Generative Foundation Model for Financial Risk Detection

The rapid growth of online payment services has intensified financial risks, while existing models typically detect each instance of risk in isolation, which neglects inter-risk and intra-risk dependencies across the entire user lifecycle, leading to suboptimal risk detection and fragmented intelligence. To overcome this limitation, we rethink financial risk modeling from a unification perspective, then propose FRiskGPT, the first financial risk-oriented generative foundation model that redefines user data and label data as a General Behavior Language, and unifies diverse tasks into next-behavior prediction. Specifically, FRiskGPT decomposes user features and labels into behaviors of users and systems, reactivating their temporal nature. Then FRiskGPT is trained by general behavior auto-regressive modeling, i.e., predicting the next behavior conditioned on all previous ones. Finally, for each downstream risk detection task, FRiskGPT can directly predict the presence of risks within the task-constrained behavior space, without fine-tuning. Extensive evaluations demonstrate FRiskGPT's superiority with average performance improvements of 12.8% across 6 real-world risk tasks compared to state-of-the-art baselines, with promising scaling potential.