Nonlinear Factor Decomposition via Kolmogorov-Arnold Networks: A Spectral Approach to Asset Return Analysis
Researchers have developed a new method called KAN-PCA, which uses Kolmogorov-Arnold Networks to generalize classical Principal Component Analysis (PCA). This approach replaces linear projections with learned B-spline functions, aiming to capture more variance, especially during market crises when linear assumptions falter. Experiments on S&P 500 stocks demonstrated that KAN-PCA achieved a higher reconstruction R^2 than classical PCA with the same number of factors. AI
IMPACT Introduces a novel neural network approach to enhance traditional financial modeling techniques.