Dynamic Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization for Multi-Asset Portfolios

SupriadiarXiv 2026 · arXiv

TL;DR

This paper introduces ST-GAT-NCO, a hybrid framework that merges Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization to revolutionize multi-asset portfolio management. By replacing static correlation metrics with dynamic, non-linear graph embeddings, the model achieves superior risk-adjusted returns and significantly better tail-risk protection during extreme market regimes like the COVID-19 crash and the 2022 inflationary shock.

What Problem Does This Paper Address?

Modern Portfolio Theory (MPT), while foundational, is notoriously fragile. Classical Mean-Variance Optimization (MVO) often acts as an "error maximizer," where small noises in estimated returns or covariances lead to extreme, unstable asset weights.

To fix this, researchers previously developed Nested Clustered Optimization (NCO), which clusters assets to insulate the portfolio from noise. However, traditional NCO has a major blind spot: it relies on static Pearson correlation. In the real world, asset relationships are non-linear, asymmetric (e.g., stocks and bonds might move together during a crash but not during a rally), and time-varying. The authors aim to bridge the gap between "geometric deep learning" and robust mathematical portfolio theory to capture these complex, shifting market dynamics.

Methods

The proposed ST-GAT-NCO architecture operates in four sequential phases:

  1. Temporal Feature Extraction: The model uses causal dilated 1D-convolutions to process historical data (returns, volatility, momentum). The "causal" and "dilated" nature ensures the model learns long-term dependencies without "cheating" by looking at future data.
  2. Spatial Graph Attention (GAT): Assets are treated as nodes in a dynamic graph. A multi-head GAT learns how shocks propagate between assets (e.g., how a liquidity crunch in bonds might infect digital assets). This allows the model to identify "flight-to-safety" channels that simple correlation matrices miss.
  3. Latent Metric Geometry: The learned high-dimensional embeddings are projected onto a unit hypersphere. The distance between assets is calculated using a non-Euclidean angular metric, which serves as the input for hierarchical Ward clustering.
  4. Two-Stage Downside NCO: Finally, the portfolio is optimized in two steps: first, minimizing downside variance within each cluster, and second, optimizing the allocation across clusters. This is combined with a "turnover smoothing" mechanism to prevent excessive trading costs.

Datasets

The model was evaluated using a multi-asset universe of 13 global asset classes sourced from Yahoo Finance, spanning January 2018 to January 2026 (including a 73-month walk-forward out-of-sample window). The assets include:

  • Equities: S&P 500 (SPY), Nasdaq 100 (QQQ), MSCI Emerging Markets (EEM).
  • Fixed Income: 7-10 Year Treasuries (IEF), 20+ Year Treasuries (TLT), High Yield (HYG), Investment Grade (LQD).
  • Real Assets: Gold (GLD), Commodities (DBC), Real Estate (VNQ).
  • Currencies/Cash: US Dollar Index (UUP), Short-Term Bills (SHV).
  • Alternative Assets: Bitcoin (BTC-USD).

Key Insights and Ideas

  • Beyond Pearson: The core "aha" moment is that graph embeddings can represent market structure more accurately than a correlation table. The GAT identifies "defensive bridges"—for instance, recognizing that Gold (GLD) and the US Dollar (UUP) act as anchors during equity sell-offs.
  • Metric Space Axioms: By projecting embeddings onto a hypersphere, the authors ensure the resulting "distance" between assets rigorously satisfies mathematical axioms (like the triangle inequality), making the subsequent clustering much more stable.
  • Regime Adaptability: The model isn't just picking "good" assets; it's picking "good structures." During the 2022 rate hikes, the model successfully rotated out of long-duration bonds and into commodities and cash, demonstrating tactical agility that static models lack.

Key Results

Evaluated net of 10 bps transaction costs, ST-GAT-NCO significantly outperformed traditional benchmarks:

  • Return: Achieved a 5.32% CAGR and 37.06% cumulative return, beating Plain NCO (4.11%) and MVO (3.68%).
  • Risk-Adjusted Performance: Delivered a Sharpe ratio of 0.30 and a Sortino ratio of 0.48, effectively doubling the risk-adjusted efficiency of standard NCO methods.
  • Tail Risk: The Maximum Drawdown was restricted to -13.07%, compared to -19.44% for a simple Equal-Weight portfolio and -16.36% for Downside NCO.
  • Diversification: The model maintained a high "Effective Number of Assets" ($N_{eff} \approx 10.35$ in late cycles), proving it achieves performance through smart diversification rather than lucky concentration.

Limitations and Future Work

The authors acknowledge several areas for improvement:

  • Rebalancing Frequency: The simulation uses a discrete monthly schedule, which might miss opportunities or risks that emerge at an intraday level.
  • Execution Modeling: While 10 bps is a standard estimate, the model assumes a linear cost regime and doesn't account for market impact (slippage) that large institutional orders might cause.
  • Granularity: The study uses ETF proxies. Future research could apply this to a more granular universe of individual corporate bonds or equities.
  • Reinforcement Learning: The authors suggest that integrating Reinforcement Learning (RL) could further optimize the execution layer under non-linear market impacts.

Summary updated

Abstract

Nested Clustered Optimization (NCO) insulates asset allocation from Markowitz covariance ill-conditioning. However, classical tree clustering relies on static pairwise Pearson correlation, failing to capture time-varying, asymmetric downside risk contagion across multi-asset markets. We propose a unified deep graph machine learning architecture: Dynamic Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization (ST-GAT-NCO) for multi-asset tactical allocation. The framework extracts temporal momentum and downside risk dynamics via causal dilated 1D-convolutions, then deploys a multi-head Graph Attention Network (GAT) over dynamic market topologies to learn latent spatial node representations. These learned embeddings define an angular metric distance space that guides two-stage downside NCO optimization under institutional constraints and turnover smoothing. Evaluated over a 73-month non-anticipative walk-forward out-of-sample window (January 2020 to January 2026) across 13 global asset classes net of 10 bps transaction costs, ST-GAT-NCO achieves a 5.32% CAGR and 37.06% cumulative net return, delivering an excess Sharpe ratio of 0.30 and a Sortino ratio of 0.48. This substantially outperforms classical Minimum Variance (3.68% CAGR, 0.08 Sharpe), Hierarchical Risk Parity (2.65% CAGR), and Plain Symmetric NCO (4.11% CAGR, 0.15 Sharpe), while restricting maximum drawdown to -13.07% (versus -19.44% for Equal-Weight and -16.36% for Downside NCO) under a controlled one-way turnover of 15.31%/month. Regime analysis confirms active capital preservation across the 2020 COVID-19 crash, 2022 monetary tightening, and 2023-2026 disinflation rallies.
BibTeX
@misc{supriadi2026dynamic,
  author = {Supriadi},
  title = {{Dynamic Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization for Multi-Asset Portfolios}},
  year = {2026},
  eprint = {2611.19903},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19903},
}
Figure