A Unified Multi-Horizon Learning Framework for Time Series Forecasting
📌 Status — This repository accompanies a paper currently under review. The source code will be released here upon acceptance.
Long-term forecasting benchmarks are evaluated at several horizons (96, 192, 336, 720), and the usual pipeline trains a separate model for each one. AnyHorizonTS trains a single model that serves any horizon within a predefined range at inference time — no retraining and no horizon-specific checkpoints.
@article{termritthikun2026anyhorizonts,
title = {AnyHorizonTS: A Unified Multi-Horizon Learning Framework
for Time Series Forecasting},
author = {Termritthikun, Chakkrit and Dorji, Karma and
Umer, Ayaz and Lee, Ivan},
journal = {Information},
year = {2026},
note = {Under review}
}Please open an issue, or contact the corresponding author at chakkritt@nu.ac.th.