GWS-STNet: A Contractive Spatio-Temporal Architecture with Metabolic Saliency for Systemic Stress Detection in JSE Equity Markets
Abstract
Predicting systemic stress in high-dimensional equity markets remains challenging, owing to non-stationarity, heavy tails, and regime shifts. This paper introduces GWS-STNet (Gaussian-Weighted Swin Spatio-Temporal Network), a deep learning architecture grounded in functional analysis and thermodynamic analogy. A Gaussian-Weighted Swin Operator (Gσ) replaces the standard shifted-window attention mechanism with a kernel-regularised counterpart on the Hilbert space (H) of financial spatio-temporal voxels. The principal theoretical contribution is a proof, via the Banach Fixed-Point Theorem, that the window-level attention operator (Gσ) is a strict contraction (Lipschitz constant of κ
Fields of Study & Subjects
Stock Market Forecasting Methods
Complex Systems and Time Series Analysis
Generative Adversarial Networks and Image Synthesis
Publication details
- Journal / source
- Mathematical and Computational Applications
- Publisher
- MDPI AG
- Volume
- 31
- Issue
- 4
- Pages
- 133–133
- Publication date
- 11 July 2026
- Language
- EN