Dynamic Graph Attention for Regime-Conditional Convex Portfolio Allocation

Abstract

Traditional portfolio optimization assumes stationarity and treats assets independently, failing catastrophically during regime transitions. We propose a hierarchical Conditional Portfolio Optimization framework pairing a quadratic programming Worker with a learned Supervisor that dynamically modulates portfolio aggressiveness. An initial tabular Supervisor overfits to path-specific crisis signatures, collapsing on synthetic market histories, exposing the structural inability of flat models to capture inter-asset contagion. To address this, we develop a Dynamic Graph Neural Network Supervisor inspired by the CRISP framework [9]: an LSTM temporal encoder coupled with a multi-head Graph Attention Network over a fully connected asset graph that learns which relationships matter dynamically. Trained end-to-end on a differentiable Sharpe ratio objective, the GNN achieves a 1.052 Sharpe ratio with 53% drawdown reduction versus the unsupervised baseline. We validate robustness via synthetic market histories (Chan 2018), walk-forward cross-validation, and regime alignment analysis confirming autonomous defensive behavior during the 2020 COVID crash and 2022 inflation shock. Code: https://github.com/Western-Artificial-Intelligence/ -Portfolio-Optimizer-Hierarchical-Regime-Conditional-CPO