Probabilistic time-series forecasting typically comes in two flavors. Quantile forecasts produce a compact summary of uncertainty at each future time step, while diffusion-based forecasters generate uncertainty through sampled future trajectories. Each approach has trade-offs: quantiles are easy to interpret, while trajectories preserve dependencies across time.
The new arXiv paper sits at the intersection. The authors study how predicted quantiles can serve as distributional controls for a diffusion forecaster. Rather than treating quantiles as a separate output or evaluation target, the idea is to feed them into the diffusion process to shape the distribution of generated trajectories.
The abstract, as provided, does not detail the method or report experimental results. At this stage, the contribution is primarily conceptual: reframing quantiles as controls for diffusion, rather than as competing representations of uncertainty.