Decoupling Scale and Pattern: A Unified Framework for Time Series Forecasting

Instance normalization methods like RevIN have become standard practice for handling non-stationary time series, yet their theoretical foundations remain poorly understood. Through comprehensive analysis, we reveal a fundamental flaw: these methods implicitly couple pattern and scale prediction, forcing models to simultaneously learn temporal patterns and predict future statistics. This coupling leads to unstable variance relationships and suboptimal performance, particularly when input variance fluctuates. We propose Decoupled Scale and Pattern Forecasting (DSPF), a novel unified framework that explicitly separates these tasks using specialized modules—a pattern forecaster operating on standardized data and a dedicated statistics forecaster that predicts future mean and variance. DSPF incorporates advanced feature engineering, including 2D temporal representations and multi-source inputs, to achieve true decoupling. Extensive experiments on Weather and Solar datasets demonstrate that DSPF establishes new state-of-the-art performance, with particularly strong improvements for long-term forecasting (14.3% average improvement on Weather dataset for H=720). Our theoretical validation confirms the coupling problem in existing methods and validates the effectiveness of explicit decoupling.