An Efficient Decomposition-Driven Linear Framework for Long-Term Time-Series Forecasting

Recent advances in long-term time-series forecasting have been predominantly driven by Transformer-based architectures, which demonstrate strong performance through powerful attention mechanisms and high model capacity. However, these models often suffer from substantial computational overhead, slow training and inference speeds, and complex feature extraction procedures, making them less suitable for real-time or resource-constrained scenarios. In this work, we revisit the potential of simple yet effective models and propose a novel forecasting framework (iLinear) based on multilayer perceptions, which we refer to as a lightweight linear MLP model. Our model captures long-term temporal dependencies without relying on attention mechanisms by leveraging a hierarchical structure with linear projections and non-linear transformations. This design not only ensures fast iteration and efficient training but also maintains strong representational capacity for complex temporal dynamics. Comprehensive experiments conducted on several widely used time-series benchmark datasets demonstrate that our model consistently outperforms state-of-the-art Transformer-based methods regarding forecasting accuracy, model size, and computational efficiency. These results highlight the feasibility of using simple neural architectures for long-term forecasting and suggest promising directions for future research in efficient time-series modeling.