The introduction of the Kolmogorov-Arnold Network (KAN) marked an important contribution to the field of deep learning, as it represented an alternative to the multilayer perceptron (MLP).
The MLP is of course the building block of many deep learning models, including state-of-the-art forecasting methods like N-BEATS, NHiTS and TSMixer.
However, in a forecasting benchmark using KAN, MLP, NHiTS and NBEATS, we discovered that KAN was generally very slow and consistently performed worse on various forecasting tasks. Note that the benchmark was done on the M3 and M4 datasets, which contain more than 99 000 unique time series with frequencies ranging from hourly to yearly.
Ultimately, at that time, applying KANs for time series forecasting was disappointing and not a recommended approach.
This has changed now with Reversible Mixture of KAN (RMoK) as introduced in the paper: KAN4TSF: Are KAN and KAN-based Models Effective for Time Series Forecasting?
In this article, we first explore the architecture and inner workings of the Reversible Mixture of KAN model…