Multiresolution cryptocurrency forecasting with audit-oriented KAN mappings

Yuan Gao

doi:10.1016/j.eswa.2026.134125

Return

Multiresolution cryptocurrency forecasting with audit-oriented KAN mappings

2026-08-21

PRE

DOI：10.1016/j.eswa.2026.134125

Original

Original request for help

Save

Abstract

En 中文

• DWT decomposes cryptocurrency signals into multiresolution components. • Resolution-specific KAN mixers model coefficient sequences before reconstruction. • Selected spline mappings support audit-oriented component inspection. Abstract Accurate cryptocurrency forecasting is challenging because market prices and related blockchain measurements exhibit pronounced non-stationarity and contain patterns at multiple temporal resolutions. Existing deep forecasting models often process these patterns within a single time-domain representation and provide limited access to the nonlinear transformations learned within individual components. To address these issues, we propose DecoKAN, a multiresolution forecasting framework that combines a multi-level Discrete Wavelet Transform (DWT), resolution-specific Kolmogorov–Arnold Network (KAN) mixers, and inverse-wavelet reconstruction. DWT separates each historical window into approximation and detail coefficient sequences, which are forecast channel-wise by KAN mixers operating across temporal patches and latent dimensions before reconstruction. The spline-parameterized mappings further support a bounded audit pathway through coefficient-importance analysis, inference-time spline-component zeroing, and local symbolic approximation. Experiments on four Electricity Transformer Temperature (ETT) benchmarks and three cryptocurrency datasets covering Bitcoin, Ethereum, and Monero demonstrate competitive general forecasting performance and the lowest horizon-macro standardized mean squared error (MSE) among the supervised models on all three cryptocurrency datasets. A checkpoint-level audit further shows that 44.93% ± 11.20% of low-importance spline components can be zeroed with an average MSE change of only +0.03%±0.13%. These findings indicate that multiresolution coefficient forecasting with spline-parameterized mappings can combine competitive predictive performance with bounded component-level auditability. Download: Download high-res image (221KB) Download: Download full-size image

Keywords:

Cryptocurrency

Time series forecasting

Kolmogorov–Arnold networks

Wavelet transform

Model auditing