Multi-asset portfolio optimization under non-stationary financial markets requires robust state representations across market phase transitions. This paper proposes distribution-robust graph representation learning for portfolio optimization (DR-GRL-PO), which learns asset-dependency graph representations as robust cross-asset structural priors for policy learning. DR-GRL-PO consists of a market-phase invariant graph contrastive encoding module (MPIGCE), a distribution-robust predictive coding module (DRPC), and a portfolio policy learning module (PPL). MPIGCE learns invariant structural priors from Spearman-based asset-dependency graphs, DRPC incorporates these priors into dual-scale predictive branches with invariant ranking consistency, and PPL integrates structural priors and predictive states for dynamic asset allocation. The model is evaluated on three separate datasets for portfolio construction, using daily data from CSI-300 (2011–2021), NASDAQ-100 (2011–2021), and Cryptocurrency (2017–2026) markets. The results show that DR-GRL-PO mainly improves wealth growth, annualized profitability, and risk-adjusted performance, while maintaining a certain degree of downside-risk control and a favorable upside–downside return balance. Its performance across separate market categories and market phases provides evidence of robustness under non-stationary market conditions. These findings indicate that robust cross-asset structural priors can support more reliable dynamic portfolio allocation.
Meng et al. (Wed,) studied this question.