Revisiting the Seasonal Trend Decomposition for Enhanced Time Series Forecasting

A team of researchers has published a paper proposing a significant enhancement to time series forecasting by revisiting a core technique: seasonal-trend decomposition. The work, accepted at ICASSP 2026, addresses a key challenge in multivariate forecasting across domains like finance, supply chain, and hydrology. The core innovation lies in a differentiated treatment of the decomposed components. Recognizing that reversible instance normalization is effective only for the trend component, the authors take a distinct approach for the seasonal component, applying backbone models directly without any normalization or scaling.

This strategic separation allows their method to reduce prediction errors more effectively. The paper reports an average reduction in Mean Squared Error (MSE) of around 10% across four established state-of-the-art baseline models on standard benchmark datasets. Furthermore, the team demonstrates the method's real-world efficacy by evaluating it on a hydrological dataset from United States Geological Survey (USGS) river stations, where it achieved significant improvements. A major practical advantage is the maintenance of linear time complexity, ensuring the approach remains scalable. The researchers also introduce dual-MLP models as more computationally efficient solutions derived from this framework.

The implications are substantial for any industry reliant on accurate forecasts. By providing a systematic way to boost the performance of existing models without a heavy computational penalty, this work offers a practical upgrade path for data science teams. The availability of the source code facilitates immediate experimentation and integration, potentially leading to more reliable predictions in critical areas like resource management, demand planning, and environmental monitoring.