ScenarioFlow: Multimodal Conditioning for Multivariate Time-Series Simulation

Risk teams need joint distributions of multi-day paths across multiple variables, conditioned on different contexts, from numerical market history to natural language. We introduce ScenarioFlow, a compact, self-supervised temporal foundation model for multivariate time-series simulation using autoregressive flow matching. ScenarioFlow learns directly from data rather than prescribing separate dynamics variable by variable. It jointly generates scenarios across variables such as implied-volatility surfaces, rates, credit, and equities. We compare ScenarioFlow against a wide range of traditional quantitative risk models, deep generative models, and fine-tuned temporal foundation models. ScenarioFlow achieves the state-of-the-art performance and is the only model to pass every criterion in our practitioner-facing evaluation. ScenarioFlow demonstrates native multimodality. Context can be supplied through time-series history or a natural-language description, while the frozen generator produces the paths. A risk manager can describe a market condition and receive conditional paths that exhibit macroeconomically coherent relationships. ScenarioFlow therefore unifies multimodal temporal context, stochastic trajectory simulation, and decision-making risk analysis in a reusable temporal foundation model.