<ul><li>SWIFT is a lightweight model proposed for Long-term Time Series Forecasting (LTSF) that is powerful and efficient in deployment and inference.</li><li>The model utilizes wavelet transform for lossless downsampling of time series, achieves cross-band information fusion with a learnable filter, and uses only one shared linear layer or one shallow MLP for sub-series' mapping.</li><li>Experiments demonstrate that SWIFT outperforms other models on multiple datasets, showcasing state-of-the-art performance, particularly beneficial for edge computing and deployment.</li><li>SWIFT-Linear, a variant of SWIFT, significantly reduces the number of parameters required for time-domain prediction, offering efficient resource utilization.</li></ul>

SWIFT: Mapping Sub-series with Wavelet Decomposition Improves Time Series Forecasting

Discover more