A1167
Title: ResNets and random matrix denoising on complex covariance structures: A cryptocurrency portfolio application
Authors: Andres Garcia - Autonomous University of Baja California (Mexico) [presenting]
Abstract: Covariance matrices estimated from short, noisy, non-Gaussian financial time series (particularly cryptocurrencies) are notoriously unstable. Sampling noise, heavy tails, asynchronous trading, and regime shifts inflate eigenvalues, distort principal directions, and degrade downstream portfolio optimization (e.g., Markowitz, risk parity, min-variance). Random matrix theory (RMT) offers theoretical eigenvalue estimation on high-dimensional settings, while residual neural networks (ResNets) can learn flexible, data-driven shrinkage and structure. The aim is to develop a hybrid RMT-informed ResNet denoiser for complex covariance structures and evaluate it on live-like crypto portfolio tasks.
