A new study gives LSTM a clear win over Kolmogorov-Arnold Networks for predicting messy, real-world financial data.
Researchers ran a head-to-head comparison of baseline KAN, implemented via PyKAN, against LSTM networks on stochastic, non-stationary financial time series. Using root mean square error in normalized feature space as the accuracy measure, LSTM outperformed KAN across every prediction horizon tested. KAN did converge faster during training, but that speed advantage did not translate into better predictions. The study used a direct multi-output forecasting protocol and was careful to note that its findings apply only to baseline KAN — not to specialized variants like Temporal KAN or Time-Frequency KAN.
KAN arrived with genuine theoretical appeal: the Kolmogorov-Arnold representation theorem promises a form of interpretability that black-box neural networks cannot easily offer. For compliance-heavy financial applications, that transparency has real value. But interpretability means little if the model's error rates are too high to trust in production. The study hands practitioners a data point they can cite when defending a less fashionable architecture choice.
LSTM has been the workhorse of sequential data modeling for decades, and researchers keep publishing its obituary every time a new architecture appears. This paper is a reminder that "newer" and "better" are not synonyms — and that the specialized KAN variants designed to fix exactly these sequential shortcomings still have a gap to close before the benchmark flips.