Bridging Simplicity and Sophistication using GLinear: A Novel Architecture for Enhanced Time Series Prediction

TL;DR

The paper addresses the challenge of achieving accurate multivariate time series forecasting with data efficiency, by bridging simplicity and sophistication through a novel GLinear architecture. GLinear combines Reversible Instance Normalization (RevIN) with a Gaussian-activated nonlinear mapping built from two small fully connected layers, operating per channel on normalized inputs to predict future steps with a lightweight, linear-like design. Empirical results on ETTh1, Electricity, Weather, and Traffic show that GLinear often outperforms state-of-the-art linear predictors (NLinear, DLinear, RLinear) and remains competitive with transformer-based Autoformer, all while requiring less historical data and offering faster training. The work demonstrates that a carefully integrated GELU-induced nonlinearity within a simple framework can capture nonlinear seasonality and local patterns, suggesting practical benefits for real-time TSF and resource-constrained settings. The authors provide the code at a public GitHub repository to facilitate reproducibility and further research.

Abstract

Time Series Forecasting (TSF) is an important application across many fields. There is a debate about whether Transformers, despite being good at understanding long sequences, struggle with preserving temporal relationships in time series data. Recent research suggests that simpler linear models might outperform or at least provide competitive performance compared to complex Transformer-based models for TSF tasks. In this paper, we propose a novel data-efficient architecture, \textit{Gaussian-activated Linear model (GLinear)}, for multivariate TSF that exploits periodic patterns to provide better accuracy. It achieves higher prediction accuracy while requiring less historical data than other state-of-the-art linear predictors. Four different datasets (ETTh1, Electricity, Traffic, and Weather) are used to evaluate the performance of the proposed predictor. A performance comparison with state-of-the-art linear architectures (such as NLinear, DLinear, and RLinear) and transformer-based time series predictors (Autoformer) shows that the GLinear, despite being data efficient, outperforms the existing architectures in most cases of multivariate TSF while being competitive in others. We hope that the proposed GLinear model opens new fronts of research and development of simpler and more sophisticated architectures for data and computationally efficient time-series analysis. The source code is publicly available on GitHub.

Figures (6)

• Figure 1: The architecture of LTSF-Linear.
• Figure 2: (a) The architecture of the NLinear predictor, (b) The architecture of the DLinear predictor. The illustration of the architecture is created based on the description in the original paper zeng2023transformers.
• Figure 3: The schematic of the architecture of RLinear. The illustration of the architecture is created based on the description in the original paper li2023revisiting.
• Figure 4: The architecture of the proposed GLinear predictor.
• Figure 5: Multivariate Forecasting performance comparison across the datasets and the models. Input sequence length is 336, and values of prediction lengths are {12, 24, 48, 96, 192, 336, 720}. The datasets are (a) Electricity, (b) ETTh1, (c) Traffic, and (d) Weather.