Autocorrelation Reintroduces Spectral Bias in KANs for Time Series Forecasting
Kolmogorov-Arnold Networks were theorized to sidestep spectral bias that plagues standard neural networks, but new research reveals this advantage collapses under temporal autocorrelation common in time series forecasting. The finding exposes a critical gap between KAN theory (which assumes statistical independence) and real-world deployment constraints. Researchers propose Discrete Cosine Transform preprocessing as a mitigation, signaling that architectural innovations require domain-specific validation before claiming broad superiority. This matters for practitioners evaluating KANs as alternatives to transformers and RNNs in forecasting pipelines.
Modelwire context
ExplainerThe deeper issue here is not just that KANs underperform under autocorrelation, but that the theoretical guarantees researchers cited when positioning KANs as superior were derived under an independence assumption that almost no real-world time series satisfies. The DCT fix is a workaround, not a resolution of the underlying mismatch.
This is largely disconnected from recent activity in our archive, as Modelwire has not yet covered the KAN architecture wave that followed the original 2024 paper or the subsequent benchmarking work comparing KANs to transformer-based forecasters like PatchTST and iTransformer. This story belongs to a broader pattern in ML research where architectural claims made on controlled benchmarks erode when tested against the messy statistical properties of production data, a pattern worth tracking as KANs accumulate applied use cases.
Watch whether forecasting benchmark suites like Monash or the M-series competitions begin incorporating KAN baselines with and without DCT preprocessing in the next two quarters. If KANs with DCT consistently match or beat linear baselines on autocorrelated series, the preprocessing step earns its place in standard pipelines; if not, the architecture's practical case for time series narrows considerably.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsKolmogorov-Arnold Networks · KANs · Discrete Cosine Transform · DCT
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.