FastTF: 4 Parameters are All You Need for Long-term Time Series Forecasting
Keywords: Time series forcasting, Machine learning, Model lightweighting
TL;DR: We achieve the best long-term time series forecasting results with an astonishingly small number of parameters through a simple method, which has significantly propelled the deployment of models on devices with limited computational power.
Abstract: Time series forecasting is essential across various sectors, including finance, transportation, and industry. In this paper, we propose FastTF, a powerful yet lightweight model in Time-Frequency domain for long-term time series forecasting. Our aim is to push the boundary of model lightweighting and facilitate the deployment of lightweight model on resource-constrained devices. Leveraging the global nature and information compressibility of the time series in frequency domain, we introduce patch-wise downsampling, Sparse Frequency Mixer (SFM), and patch predictor to capture the temporal variations of frequency components across different patches. Experimental results on five public datasets demonstrate that FastTF with very few parameters outperforms several state-of-the-art models and demonstrates a strong generalization capability. Notably, on the ETTh1 dataset, FastTF with only 4 parameters achieves a performance that is close to the DLinear and FITS in the horizon-96 forecasting. Furthermore, we deployed our model on a FPGA development board (Zynq UltraScale+ RFSoC ZCU208 Evaluation Kit), where the corresponding resource usage statistics illustrate that our model has a very low computational overhead and latency, making it easily implemented on hardware devices.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6843
Loading