Authors
Ziqiong Li, Heyu Chai, Xinru Liu, Shengjun Liu
Published in
IEEE transactions on neural networks and learning systems. Volume PP. Sep 18, 2026. Epub Sep 18, 2026.
Abstract
Currently, time series forecasting methods typically follow a "decompose-and-forecast-independently" strategy, where the original series is first decomposed into trend and seasonal components, and different components are then modeled separately. However, in real-world data, trends and seasonality are often dynamically coupled and mutually influential. Ignoring such intercomponent dependencies may weaken the benefits of decomposition-based modeling, thereby limiting forecasting performance. To address this issue, we propose a forecasting framework that explicitly models the coupling between trend and seasonality. In particular, by jointly modeling amplitude and phase, we align trend intensity with seasonal phase at the feature level, thereby capturing their intrinsic relationship. This design improves the interpretability of the coupled representation and yields consistent performance gains across different datasets and prediction horizons. Moreover, we design a heteroscedastic Laplace loss function that combines uncertainty weighting with heteroscedastic modeling, reducing the impact of error accumulation over long prediction horizons and enhancing robustness to outliers and heavy-tailed noise. Extensive experiments on 12 real-world datasets demonstrate that TSCNet achieves competitive or leading performance in terms of forecasting accuracy and long-horizon stability. The code and datasets are available at https://github.com/QiongLouLi/TSCNet.
PMID:
42758953
Bibliographic data and abstract were imported from PubMed on 19 Sep 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 49
- Comments 0