Scattered Hypothesis Generation for Open-Ended Event Forecasting

A research team led by He Chang and Zhulin Tao has introduced SCATTER, a novel AI framework that fundamentally shifts how we forecast open-ended events. Current LLM-based forecasting methods typically predict only the single most probable outcome, which is inadequate for managing real-world uncertainty in domains like finance, geopolitics, or security. SCATTER reframes the task as 'scatter forecasting,' aiming to generate a broad, inclusive set of plausible future hypotheses that cover the space of possible events, moving from pinpoint predictions to scenario exploration.

The framework uses a reinforcement learning approach with a carefully designed hybrid reward function. This function has three key components: a validity reward to ensure hypotheses are semantically aligned with observed events, an intra-group diversity reward to encourage variation within response batches, and an inter-group diversity reward to promote exploration across distinct outcome modes. A 'validity-gated' score prevents the generation of wildly implausible futures, addressing the common AI problem of mode collapse where models output similar, safe answers.

In experiments on two real-world benchmark datasets—OpenForecast and OpenEP—SCATTER significantly outperformed existing strong baselines. The team has made their code publicly available, providing a practical tool for analysts and risk managers. This approach is particularly valuable for strategic planning, where understanding a range of potential futures, not just the most likely one, is critical for robust decision-making and contingency planning.