Existence Precedes Value: Joint Modeling of Observational Existence and Evolving States in Time Series Forecasting
PyTorch Implementation of Timeflies.
🚩 2026-06-11: Initial upload to arXiv (PDF).
Real-world time series are often incomplete and irregular, where future observations may not always exist. As shown in the figure below, Timeflies moves beyond the traditional impute-then-forecast paradigm and the irregular-aware paradigm that assumes future timestamps are known. Instead, it introduces an observation-state joint paradigm: first infer whether a future observation will occur, and then predict its value only when the observation is likely to be valid.
Timeflies follows a dual-stream pipeline to jointly model observational existence and evolving states. The observation stream captures missingness patterns from masks and missing intervals, while the value stream models temporal dynamics. Through reliability-aware patch embedding, observation-guided value attention, and a dual prediction head, Timeflies produces both future observation probabilities and value forecasts, making predictions more suitable for real-world irregular time series.
If you find this repo useful, please consider citing our paper as follows:
@article{hu2026existence,
title={Existence Precedes Value: Joint Modeling of Observational Existence and Evolving States in Time Series Forecasting},
author={Hu, Yifan and Chen, Hongzhou and Liu, Peiyuan and Liu, Yiding and Dong, Zewei and Yang, Jiang-Ming},
journal={arXiv preprint arXiv:2606.13571},
year={2026}
}Special thanks to the following repositories for their invaluable code and datasets:
If you have any questions, please contact huyf0122@gmail.com or submit an issue.