GLinear: A Data-Efficient Gaussian-Activated Linear Model for Time Series Forecasting
A single-layer linear architecture for multivariate time-series forecasting that augments a linear projection with a Gaussian Error Linear Unit (GELU) nonlinearity and Reversible Instance Normalization (RevIN), leveraging periodic patterns to forecast accurately from less historical data than prior linear predictors. Without self-attention or positional encoding, GLinear outperforms the NLinear, DLinear, and RLinear linear predictors and the Autoformer transformer in most multivariate settings across the ETTh1, Electricity, Traffic, and Weather benchmarks (MSE/MAE), showing that a simple, data-efficient linear model can rival complex architectures.
Bridging Simplicity and Sophistication using GLinear: A Novel Architecture for Enhanced Time Series
Recent work argues that simple linear models can match or beat complex Transformer architectures for time-series forecasting (TSF). This paper proposes GLinear, a data-efficient single-layer linear m…