ICASSP 2026 - CCF B

Towards Reliable Time Series Forecasting under Future Uncertainty

Ambiguity and novelty rejection for forecasting

Reliable time-series forecasting overview

TL;DR

This work improves forecasting reliability by rejecting ambiguous or distribution-shifted predictions before deployment risk grows.

Authors: Ninghui Feng†, Songning Lai†, Xin Zhou, Jiayu Yang, Kunlong Feng, Zhenxiao Yin, Fobao Zhou, Zhangyi Hu, Yutao Yue, Yuxuan Liang, Boyu Wang, Hang Zhao

Venue: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026

Abstract

Forecasting systems often fail because future uncertainty is not visible at training time. This work combines ambiguity rejection and novelty rejection to identify low-confidence or out-of-distribution future states, allowing the model to abstain from unreliable predictions when future labels are unavailable.

Key Idea

  • Uses ambiguity rejection to abstain under high predictive uncertainty.
  • Uses novelty rejection to detect distribution-shifted future conditions.
  • Targets deployment settings where future ground truth is unavailable.

Poster Figure

Towards Reliable Time Series Forecasting under Future Uncertainty

Overview of ambiguity and novelty rejection for reliable time-series forecasting.