Forecasting Functions, Not Points: Time Series Forecasting via Polynomial-Fourier Coefficient Prediction
Abstract
Environmental sensor signals, including meteorological, air-quality, and eco-hydrological measurements, are highly characterized by a slowly varying trend and a small number of low-frequency periodic components. In long-term forecasting benchmarks for these signals, the models that explicitly capture trend and seasonality have consistently achieved strong performance. However, in practical deployments, the dominant periods of environmental signals vary between regimes, which heavily forces a deployed forecaster to operate over a wide range of varying prediction horizons H. Standard point forecasting models, such as DLinear, PatchTST, and TimeMixer, tie the output projection weights directly to H. Consequently, any change in the horizon severely increases the decoder parameter count and requires complete, expensive retraining. To the best of our knowledge, no prior studies have proposed a flexible forecasting framework that eliminates this horizon-dependent structural rigidity. Therefore, we propose a novel functional forecasting framework called CoefNet (Coefficient-prediction Network), which approaches long-term time-series forecasting by predicting the coefficients of a continuous function ƒ(t) composed of polynomial and Fourier terms. The decomposition into trend and seasonality is explicitly encoded in the structure of ƒ(t) itself, allowing the same trained coefficients to be evaluated on any arbitrary target grid without additional parameters or retraining. Since a pure polynomial-Fourier basis captures global trend and periodicity but may not capture the detailed local variations of signals, we further propose a per-sub-window low-rank modulation scheme. This adaptable scheme captures localized fluctuations by perturbing the shared coefficients at a minimal 4% parameter overhead. As a result, CoefNet successfully achieves the coarse-to-fine prediction property without requiring post-hoc interpolation. Extensive experimental results on four datasets of real-world environmental sensors demonstrate that the proposed approach is robust against varying horizons, attaining state-of-the-art performance in 13 of 16 settings against nine baselines for standard point forecasting.