A compact Kolmogorov-Arnold network mixer for long-term time series forecasting

Long-term time series forecasting (LTSF) underpins critical applications from energy management to weather prediction, yet achieving reliable multi-step-ahead accuracy remains challenging. Existing LTSF approaches, dominated by MLP- and Transformer-based architectures, either rely on simple linear mappings or introduce increasingly complex hand-crafted inductive biases, raising the question of whether a more expressive nonlinear modeling core could offer a useful alternative. In this work, we investigate whether Kolmogorov-Arnold Networks (KANs), which use learnable basis functions on network edges to model nonlinear relationships, can serve as effective modeling components for LTSF, and under which design choices they are most useful. Motivated by this question, we propose KANMixer, a compact KAN-centered architecture consisting of a multi-scale pooling frontend, KAN-based temporal mixing blocks, and KAN-based prediction heads. Unlike KAN-based forecasting models that combine KAN with decomposition-heavy or mixture-based pipelines, KANMixer is designed as a simple and controlled architecture for examining the role of KAN components in LTSF. Under a unified five-run reproduction protocol on seven standard benchmarks, KANMixer achieves competitive performance against representative LTSF baselines, especially on ETT-style datasets, while showing dataset-dependent limitations. Additional statistical tests, ablations, efficiency profiling, Gaussian-noise evaluation, and hyperparameter sensitivity analysis show that the practical value of KAN depends on basis-function choice, architectural placement, and computational constraints. These results suggest that KANs are promising but not plug-and-play components for LTSF, and that their benefits should be evaluated together with robustness and efficiency trade-offs.