Abstract
Time Series Forecasting (TSF) is an important application across many fields.There is a debate about whether Transformers, despite being good atunderstanding long sequences, struggle with preserving temporal relationshipsin time series data. Recent research suggests that simpler linear models mightoutperform or at least provide competitive performance compared to complexTransformer-based models for TSF tasks. In this paper, we propose a noveldata-efficient architecture, GLinear, for multivariate TSF that exploitsperiodic patterns to provide better accuracy. It also provides betterprediction accuracy by using a smaller amount of historical data compared toother state-of-the-art linear predictors. Four different datasets (ETTh1,Electricity, Traffic, and Weather) are used to evaluate the performance of theproposed predictor. A performance comparison with state-of-the-art lineararchitectures (such as NLinear, DLinear, and RLinear) and transformer-basedtime series predictor (Autoformer) shows that the GLinear, despite beingparametrically efficient, significantly outperforms the existing architecturesin most cases of multivariate TSF. We hope that the proposed GLinear opens newfronts of research and development of simpler and more sophisticatedarchitectures for data and computationally efficient time-series analysis.