Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective
Event sequence forecasting is framed as a conditional generative modeling problem where no single method achieves both high prediction accuracy and distributional realism simultaneously, highlighting a fundamental trade-off. The work introduces Wasserstein BaryBooster, an algorithm that combines samples from a diffusion model to produce the best possible predictor, proving that distributional realism can be successfully converted into accurate predictions.