Paper: arXiv 2611.19903

Authors: Supriadi

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.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a highly complex mathematical framework integrating deep learning with portfolio optimization. It demonstrates strong empirical rigor through extensive out-of-sample backtesting across diverse market regimes. The novelty lies in the unified ST-GAT-NCO architecture for dynamic multi-asset allocation.

Research Flowchart

  flowchart TD
    A["Research Goal: Improve Multi-Asset Tactical Allocation by Addressing Covariance Ill-Conditioning and Dynamic Risk"] --> B{"Key Problem: Classical NCO Fails to Capture Dynamic, Asymmetric Downside Risk Contagion"}

    B --> C["Proposed Solution: ST-GAT-NCO Framework"]
    C --> D{"Methodology: Unified Deep Graph ML Architecture"}
    D --> D1["1. Causal Dilated 1D-Convolutions: Extract Temporal Momentum & Downside Risk Dynamics"]
    D1 --> D2["2. Multi-head Graph Attention Network (GAT): Learn Latent Spatial Node Representations from Dynamic Market Topologies"]
    D2 --> D3["3. Two-Stage Downside NCO Optimization: Guided by GAT Embeddings, under Institutional Constraints & Turnover Smoothing"]

    D --> E["Data/Inputs: 13 Global Asset Classes, 73-Month Out-of-Sample Window (Jan 2020 - Jan 2026)"]

    E --> F{"Computational Process: Non-Anticipative Walk-Forward Evaluation, Net of 10 bps Transaction Costs"}

    F --> G["Key Findings/Outcomes: Superior Performance & Risk Control"]
    G --> G1["CAGR: 5.32% (vs MV 3.68%, HRP 2.65%, Plain NCO 4.11%)"]
    G --> G2["Cumulative Net Return: 37.06%"]
    G --> G3["Excess Sharpe Ratio: 0.30; Sortino Ratio: 0.48"]
    G --> G4["Max Drawdown: -13.07% (vs EW -19.44%, Downside NCO -16.36%)"]
    G --> G5["Controlled Turnover: 15.31%/month"]
    G --> G6["Active Capital Preservation across diverse regimes (COVID, Monetary Tightening, Disinflation)"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Dynamic Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization (ST‑GAT‑NCO), a deep graph‑machine‑learning architecture that extracts temporal momentum and downside risk via causal dilated 1D‑convolutions and learns latent spatial node embeddings with a multi‑

Source sentence
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.

Assumes 6

  • A causal rolling lookback window of 24 months is used for training, covariance estimation and clustering, ensuring zero look‑ahead bias. model
    Source sentence
    B. Walk-Forward Backtesting Protocol To strictly ensure zero lookahead bias: • A causal rolling lookback window of 𝐿= 24 months is utilized for GNN training, covariance estimation, and clustering. • Out-of-sample evaluation covers 73 consecutive months from January 2020 through January 2026. • Execution friction is penalized at a realistic proportional flat fee of 𝑐= 10 bps (0.10%) on weight turnover [38]. •
  • Execution friction is modeled as a flat proportional fee of 10 bps on weight turnover. execution
    Source sentence
    B. Walk-Forward Backtesting Protocol To strictly ensure zero lookahead bias: • A causal rolling lookback window of 𝐿= 24 months is utilized for GNN training, covariance estimation, and clustering. • Out-of-sample evaluation covers 73 consecutive months from January 2020 through January 2026. • Execution friction is penalized at a realistic proportional flat fee of 𝑐= 10 bps (0.10%) on weight turnover [38]. •
  • Asset returns, volatilities and other features are observed monthly, implying a discrete monthly rebalancing schedule. data
    Source sentence
    III. Experimental Setup A. Multi-Asset Universe The empirical testbed spans 13 liquid global asset classes observed monthly from January 2018 to January 2026 (𝑇= 97 months), sourced from Yahoo Finance: • Equities: S&P 500 (SPY), Nasdaq 100 (QQQ), MSCI Emerging Markets (EEM). • Fixed Income: 7–10 Year US Treasuries (IEF), 20+ Year Treasuries (TLT), High Yield Corporate Bonds (HYG),
  • Market dependencies are non‑linear, time‑varying and asymmetric, motivating the use of dynamic graph topologies instead of static Pearson correlations. market
    Source sentence
    In practical multi-asset systems spanning equities, fixed income, commodities, arXiv:2611.19903v1 [q-fin.PM] 15 Nov 2026 2 currencies, and cryptocurrencies, market dependencies are non-linear, time-varying, and asymmetric [20]– [22].
Show 2 more
  • Turnover is smoothed via a convex turnover‑smoothing regularizer within the NCO optimization. model
    Source sentence
    These learned embeddings define an angular metric distance space that guides two-stage downside NCO optimization under institutional constraints and turnover smoothing.
  • Institutional risk constraints (e.g., weight limits) are imposed during the two‑stage downside NCO allocation. other
    Source sentence
    These learned embeddings define an angular metric distance space that guides two-stage downside NCO optimization under institutional constraints and turnover smoothing.

Data

Backtest sample: 2020-01 – 2026-01 · global multi-asset · multi-asset · monthly

  • Yahoo Finance market price data · Yahoo Finance market-prices public
    13 global asset classes (equities, treasuries, corporate bonds, commodities, real estate,
    Source sentence
    Evaluated over a 73-month strictly non-anticipative walk-forward out-of-sample window (January 2020 to January 2026) across 13 global asset classes (equities, treasuries, corporate bonds, commodities, real estate, cash, and digital assets) sourced from Yahoo Finance market data, net of 10 bps transaction costs, ST-GAT-NCO achieves a 5.32% annualized return (CAGR) and a 37.06% cumulative net return, III. Experimental

Recognised sources: Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

Flat fee of 10 bps per month on weight turnover (0.10% proportional cost).

Source sentence
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. B. Walk-Forward Backtesting Protocol To strictly ensure zero lookahead bias: • A causal rolling lookback

Finds 2

Reported performance (ST-GAT-NCO): Sharpe 0.30 (benchmark 0.72) · annual return 5.3% · total return 37.1% · max drawdown 13.1%

  • ST‑GAT‑NCO achieves a 5.32 % CAGR, 37.06 % cumulative net return, excess Sharpe ratio of 0.30 and Sortino ratio of 0.48 over a 73‑month walk‑forward out‑of‑sample period (Jan 2020–Jan 2026), with maximum drawdown limited to –13.07 % and one‑way turnover of 15. performance figure from elsewhere in the paper
    Source sentence
    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
  • The strategy 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 delivering lower drawdowns. performance figure from elsewhere in the paper
    Source sentence
    As demonstrated in Figure 2, the proposed ST-GAT-NCO delivers superior risk-adjusted alpha over all clustered and Markowitz baselines: • CAGR Alpha Enhancement: ST-GAT-NCO generates a CAGR of 5.32% (total return of 37.06%), outper-forming Plain NCO (4.11% CAGR, 27.75% return) by +121 bps per annum, Downside NCO (4.26% CAGR) by +106 bps, and classical Minimum Variance (3.68% CAGR) by +164 bps. • Sharpe and Sortino

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 10 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

Extracted 2026-11-17 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (38)

  1. Portfolio Selection (1952)
  2. Portfolio Selection: Efficient Diversification of Investments (1959)
  3. The Markowitz Optimization Enigma: Is ‘Optimized’ Optimal? (1989)
  4. Optimal Versus Naive Diversification: How Inefficient is the 1/N Portfolio Strategy? (2009)
  5. Global Portfolio Optimization (1992)
  6. A Well-Conditioned Estimator for Large-Dimensional Covariance Matrices (2004)
  7. Hierarchical Structure in Financial Markets (1999)
  8. Dynamics of Market Correlations: Taxonomy and Portfolio Analysis (2003)
and 30 more
  1. A Tool for Filtering Information in Complex Systems (2005)
  2. Correlation, Hierarchies, and Networks in Financial Markets (2010)
  3. Building Diversified Portfolios that Outperform Out of Sample (2016)
  4. Advances in Financial Machine Learning (2018)
  5. A Robust Estimator of the Efficient Frontier (2019)
  6. Machine Learning for Asset Managers (2020)
  7. Hierarchical Clustering-Based Asset Allocation: The Case of Equal Risk Contribution and Minimum Variance (2017)
  8. The Hierarchical Equal Risk Contribution Portfolio (2018)
  9. Hierarchical Clustering for Asset Allocation in Large Portfolios (2025)
  10. Semi-Supervised Classification with Graph Convolutional Networks (2017)
  11. Graph Attention Networks (2018)
  12. International Asset Allocation with Regime Shifts (2002)
  13. Regime Changes and Financial Markets (2012)
  14. Capital Market Equilibrium in a Mean-Lower Partial Moment Framework (1977)
  15. Downside Risk (1991)
  16. Multi-Period Trading via Convex Optimization (2017)
  17. Market Regime-Based Multi-Asset Portfolio Optimization Using Nested Clustered Optimization for Tactical Asset Allocation — → 2610.27751
  18. Mitigating Return Drag and Regime Whipsaws in Sector Rotation: An Asymmetric Downside-Aware Nested Clustered Optimization Framework — → 2611.05310
  19. Temporal Relational Ranking for Stock Prediction (2019)
  20. Knowledge Graph-Based Event Embedding Framework for Financial Quantitative Investments (2022)
  21. Graph-Based Dynamic Covariance Estimation for Risk Parity Portfolios (2024)
  22. Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting (2018)
  23. Graph WaveNet for Deep Spatial-Temporal Graph Modeling (2019)
  24. Deep Attentive Learning for Stock Movement Prediction from Social Media Text and Company Correlations (2020)
  25. Stock Selection via Spatiotemporal Hypergraph Attention Network: A Multi-Task Approach (2021)
  26. Graph Neural Networks in Financial Market Forecasting: A Comprehensive Survey (2023)
  27. A Review of Two Decades of Correlations, Hierarchies, Networks and Clustering in Financial Markets (2021)
  28. Mean-Semivariance Behavior: Downside Risk and Investment Decisions (2007)
  29. Introduction to Risk Parity and Budgeting (2013)
  30. Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance (2014)

2 of 38 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.