SALT: Structure-Aligned Learning for Time-Series Forecasting

Sheo yon Jhin, Noseong Park

Keywords: Time series forecasting

Abstract: Time-series forecasting is vital in domains such as economics, energy, and traffic. Although many models exploit the decomposable nature of time-series, they are typically trained with a single objective (e.g., MSE), which imposes structural limits on their performance. We empirically demonstrate that this paradigm gives rise to two characteristic challenges: gradient interference, where heterogeneous components conflict, and spectral bias, where dominant low-frequency structures overshadow informative high-frequency ones. To move beyond these limitations, we introduce SALT (\textbf{S}tructure-\textbf{A}ligned \textbf{L}earning for \textbf{T}ime-Series), which combines Iterative Dominant Extraction (IDE) with Separable Training to optimize components independently. Our theoretical and empirical analyses show that this regime reduces cross-term errors, balances convergence across frequencies, and consistently surpasses the conventional methods across six backbones and nine benchmarks.

Primary Area: learning on time series and dynamical systems

Submission Number: 10377