TimeKAN: an adaptive frequency-decomposed Kolmogorov-Arnold network for long-term stock forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Real-world time series data often contain complex, intertwined frequency components.
• Traditional forecasting methods struggle with these complexities, leading to suboptimal accuracy.
• Existing approaches often apply uniform modeling, neglecting distinct frequency characteristics.

Purpose of the Study:

• To introduce TimeKAN, a novel Kolmogorov-Arnold Network (KAN)-based architecture for long-term time series forecasting.
• To address the limitations of unified modeling strategies in handling multi-frequency time series.
• To leverage KAN's function approximation capabilities for improved time series prediction.

Main Methods:

• TimeKAN employs a Cascading Frequency Decomposition (CFD) block for adaptive signal decomposition.
• A Multi-order KAN (M-KAN) representation learning block uses Chebyshev polynomials for specialized frequency band modeling.
• A frequency mixing block integrates information across bands using multi-head attention.

Main Results:

• TimeKAN demonstrated an average 21.5% RMSE improvement over state-of-the-art methods on four stock datasets.
• [Formula: see text] scores consistently exceeded 91%, indicating high prediction accuracy.
• Experiments showed significantly improved investment returns, validating practical utility.

Conclusions:

• TimeKAN offers a superior technical approach for complex long-term time series forecasting.
• The M-KAN block was identified as the most critical component for performance.
• This research highlights the potential of KAN architectures in advanced sequence modeling.