FORESIGHT: Frequency and Order-Aware REpresentation With Spectral-Domain and Inter-Channel Global Horizon Transformer

Abstrak

Forecasting multivariate time series remains challenging due to the coexistence of long-term dependencies, periodic structures, and complex inter-variable interactions. Existing methods often emphasize either frequency-aware or order-aware representations, without fully leveraging their complementary roles. To address this, we propose FORESIGHT (Frequency- and Order-aware Representation with Spectral-domain and Inter-channel Global Horizon Transformer), a dual-stream framework that models seasonal and trend components through coordinated representation learning. The model employs an Exponential Order-aware Adaptive (EOA) decomposition to separate input signals, followed by a spectral-channel encoding module (SCTEM) for capturing periodic patterns and inter-variable dependencies, and an order-aware module (ORM) for modeling temporal progression in trend components. An adaptive fusion mechanism integrates both streams in an end-to-end manner. Extensive experiments on five benchmark datasets (Weather, Exchange Rate, Traffic, ETTh1, and ETTm2) across multiple prediction horizons demonstrate competitive and consistent performance relative to strong baselines. Compared with FAITH, FORESIGHT achieves average improvements of 3.092% in MAE and 2.463% in RMSE while maintaining statistical parity under strict significance thresholds. These results suggest that integrating frequency- and order-aware representations leads to more stable and robust long-horizon forecasting.

Identifier

• DOI: https://doi.org/10.1109/access.2026.3700534
• OpenAlex: https://openalex.org/W7163696785