Improving Linear Models for Univariate Long-term Time Series Forecasting

Transformer models have achieved state-of-the-art performance in fields such as natural language processing and computer vision. However, they encounter difficulties when applied to time series forecasting due to the permutation invariance of their self-attention mechanism. This has encouraged adapting simpler models, including those based on CNNs, RNNs, and MLPs. While MLP-based models succeed at capturing the moving average and seasonal patterns in data, they struggle with trends and sudden changes. To address this issue, we propose a Univariate Multi-Scale Linear (UMS-Linear) model that leverages timestamps and multi-scale decomposition. UMS-Linear improves the forecasting ability of MLP-based models by capturing the intricate dynamics of data. Empirical results demonstrate that UMS-Linear surpasses existing methods across multiple benchmarks, indicating its potential to advance time series forecasting accuracy.