A Compact Kolmogorov–Arnold Network Mixer for Long-Term Time Series Forecasting

Kolmogorov–Arnold Networks are usually bolted onto an existing forecasting backbone, which makes it hard to tell what the KAN itself contributes. KANMixer is built the other way round: a compact, KAN-centered architecture — multi-scale pooling frontend, KAN-based temporal mixing blocks, KAN-based prediction heads — deliberately stripped of auxiliary machinery so the KAN components can be examined on their own terms.

Under a unified five-run reproduction protocol on seven standard benchmarks, KANMixer is competitive with representative LTSF baselines, particularly on ETT-style datasets, while showing dataset-dependent limitations. Statistical tests, ablations, efficiency profiling, Gaussian-noise evaluation and hyperparameter sensitivity analysis all point the same way: the practical value of a KAN depends on basis-function choice, on where in the architecture it sits, and on the compute budget you are willing to spend.

The conclusion is deliberately unglamorous. KANs are promising but not plug-and-play components for long-term forecasting, and their benefits are only meaningful when weighed against robustness and efficiency trade-offs.

Preprint version (earlier, differently titled): KANMixer: a minimal KAN-centered mixer for long-term time series forecasting, arXiv:2508.01575

Recommended citation: Jiang, L., Hou, D., Wang, Y., Su, Y., Xing, S., Chen, W., Zhang, X., Tu, Z., Zhang, Z., Lin, F., Zielewski, M., & Yamada, K. (2026). A compact Kolmogorov–Arnold network mixer for long-term time series forecasting. Scientific Reports. https://doi.org/10.1038/s41598-026-59667-5
Download Paper