Abstract
Abstract
Financial fraud and enterprise risk rarely emerge as independent node-level events: liquidity stress, suspicious payments, ownership dependencies, guarantee relations, and transaction chains create interacting pathways through which local shocks may spread across a financial network. Existing graph fraud detectors are effective at classifying suspicious entities, while recent temporal graph models and generative world models improve sequence prediction. However, a risk-management system also needs to answer a different class of questions: what future network state would arise under a specified intervention, which transmission channels would carry the shock, and how much fraud or systemic risk would have been avoided under a counterfactual policy? We formulate these questions within a unified interventional generative process and propose RIFT-WM, a Risk-Intervention Flow Transport World Model. Its central object is an action-indexed stochastic infinitesimal generator whose semigroup jointly governs continuous latent-risk drift, topology-aware risk transport, discrete fraud jumps, and event intensities. Counterfactuals are generated by abducting episode-specific structural disturbances and reusing the same disturbances in factual and intervened ''twin worlds,'' so that only the designated mechanism changes. A sub-stochastic transport constraint prevents unexplained creation of endogenous risk mass, while semigroup consistency makes predictions coherent across temporal resolutions and rollout horizons. We further derive intervention influence operators, enterprise-level counterfactual effects, and network-level systemic-risk decompositions from the same generator. The evaluation framework spans Elliptic, Elliptic++, T-Finance, DGraph-Fin, AMLSim, and PaySim and covers fraud detection, long-horizon forecasting, calibration, counterfactual recovery, robustness, and efficiency. Counterfactual claims are restricted to controlled simulator settings in which interventions can be replayed under matched exogenous disturbances. Code is available at https://anonymous.4open.science/r/RIFT_WM-4151/.