Paper: arXiv 2601.07792

Authors: Javier Mancilla, Theodoros D. Bouloumis, Frederic Goguikian

Abstract

Portfolio optimization under cardinality constraints transforms the classical Markowitz mean-variance problem from a convex quadratic problem into an NP-hard combinatorial optimization problem. This paper introduces a novel approach using THRML (Thermodynamic HypergRaphical Model Library), a JAX-based library for building and sampling probabilistic graphical models that reformulates index tracking as probabilistic inference on an Ising Hamiltonian. Unlike traditional methods that seek a single optimal solution, THRML samples from the Boltzmann distribution of high-quality portfolios using GPU-accelerated block Gibbs sampling, providing natural regularization against overfitting. We implement three key innovations: (1) dynamic coupling strength that scales inversely with market volatility (VIX), adapting diversification pressure to market regimes; (2) rebalanced bias weights prioritizing tracking quality over momentum for index replication; and (3) sector-aware post-processing ensuring institutional-grade diversification. Backtesting on a 100-stock S and P 500 universe from 2023 to 2025 demonstrates that THRML achieves 4.31 percent annualized tracking error versus 5.66 to 6.30 percent for baselines, while simultaneously generating 128.63 percent total return against the index total return of 79.61 percent. The Diebold-Mariano test confirms statistical significance with p less than 0.0001 across all comparisons. These results position energy-based models as a promising paradigm for portfolio construction, bridging statistical mechanics and quantitative finance.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper uses advanced statistical mechanics (Ising Hamiltonians, Boltzmann distributions) and NP-hard optimization, indicating high math complexity. It also provides concrete backtesting results with specific metrics, statistical significance tests, and implementation details (JAX library, GPU sampling), demonstrating strong empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve Index Tracking via Non-Convex Portfolio Optimization"] --> B["Methodology: Energy-Based Model using THRML"]
  B --> C["Inputs: S&P 500 100-Stock Universe 2023-2025"]
  C --> D["Computational Process: GPU-Accelerated Probabilistic Inference"]
  D --> E["Dynamic Coupling & Sector-Aware Post-Processing"]
  E --> F["Outcomes: 4.31% Tracking Error vs 5.66-6.30% Baseline"]
  F --> G["Outcomes: 128.63% Total Return vs 79.61% Index Return"]
  G --> H["Statistical Significance: Diebold-Mariano p < 0.0001"]

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 a novel THRML-based approach for cardinality-constrained index tracking that samples portfolios from a Boltzmann distribution using GPU-accelerated block Gibbs sampling and introduces three key innovations.

Source sentence
We implement three key innovations: (1) dynamic coupling strength that scales inversely with market volatility (VIX), adapting diversification pressure to market regimes; (2) rebalanced bias weights prioritizing tracking quality over momentum for index replication; and (3) sector-aware post-processing ensuring institut

Assumes 6

  • Cardinality constraint limits the number of holdings to a fixed K (30 stocks). model
    Source sentence
    Target cardinality (K) 30 stocks
  • All selected assets are equally weighted after selection. execution
    Source sentence
    After selection, we apply equal weights to selected assets.
  • Transaction costs are a flat 10 basis points per trade. execution
    Source sentence
    Transaction cost 10 bps per trade
  • Training and test periods are non‑overlapping with a walk‑forward methodology. data
    Source sentence
    Training Period: January 2010 – December 2022 (13 years) • Test Period: January 2023 – December 2025 (3 years)
Show 2 more
  • Liquidity is measured by normalized average trading volume. model
    Source sentence
    Liquidityi = AvgVolumei / maxj AvgVolumej
  • Dynamic coupling strength depends on VIX levels. model
    Source sentence
    dynamic coupling strength that scales inversely with market volatility (VIX)

Data

Backtest sample: 2023 – 2025 · S&P 500 stocks · equities · monthly

  • Historical price data · Yahoo Finance market-prices public
    January 2010 to December 2025 · 100 large‑cap U.S. equities (S&P 500 constituents)
    Source sentence
    Historical price data from January 2010 to December 2025 was obtained via Yahoo Finance. Prices were adjusted for splits and dividends, with missing values forward-filled.
  • VIX data macro unknown
    January 2010 to December 2025
    Source sentence
    VIX data for dynamic coupling was obtained for the same period.

Recognised sources: Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

Flat 10 basis points per trade applied in backtest results.

Source sentence
Transaction cost 10 bps per trade

Finds 5

Reported performance (THRML (Thermodynamic HypergRaphical Model Library)): Sharpe 1.87 · total return 128.6% · max drawdown 17.1%

  • THRML achieves a tracking error of 4.31%, 24% lower than the next‑best baseline (NSGA‑II at 5.66%). performance
    Source sentence
    Tracking Error: THRML’s 4.31% TE is 24% lower than the next-best baseline (NSGA‑II at 5.66%)
  • THRML generates a total return of 128.63% versus the S&P 500 index return of 79.61%, yielding an information ratio of 1.85. performance
    Source sentence
    THRML generates 49 percentage points of excess return (128.63% vs 79.61% index), yielding an Information Ratio of 1.85
  • The Sharpe ratio of THRML is 1.87, exceeding all baselines by more than 34%. performance
    Source sentence
    The Sharpe ratio of 1.87 exceeds all baselines by 34%+ margin.
  • Diebold‑Mariano tests show statistical significance at p < 0.0001 for all baseline comparisons. statistical
    Source sentence
    All comparisons are statistically significant at the p < 0.0001 level, rejecting the null hypothesis that THRML and baselines have equal tracking accuracy.
  • Maximum drawdown for THRML is -17.09%, the shallowest among all methods. performance
    Source sentence
    THRML achieves the shallowest maximum drawdown (-17.09%) while delivering the highest returns

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 83 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (23)

  1. Portfolio Selection (1952)
  2. Computational study of a family of mixed-integer quadratic programming problems (1996)
  3. Efficient Asset Management: A Practical Guide to Stock Portfolio Optimization and Asset Allocation (2008)
  4. The Sharpe Ratio Efficient Frontier (2012)
  5. Honey, I Shrunk the Sample Covariance Matrix (2004)
  6. Robust Portfolio Optimization and Management (2007)
  7. Building Diversified Portfolios that Outperform Out of Sample (2016)
  8. Algorithm for Cardinality Constrained Quadratic Optimization (2009)
and 15 more
  1. A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II (2002)
  2. Optimization by Simulated Annealing (1983)
  3. Quantum Annealing for Combinatorial Optimization: A Comparative Study (2022)
  4. Beitrag zur Theorie des Ferromagnetismus (1925)
  5. Reducing the Dimensionality of Data with Neural Networks (2006)
  6. THRML : Thermodynamic HypergRaphical Model Library (2024)
  7. Thermodynamic Computing: The Future of Probabilistic AI (2024)
  8. A Quantum Approximate Optimization Algorithm (2014) — link
  9. A Tutorial on Formulating and Using QUBO Models (2019) — link
  10. Quantum computing for finance: Overview and prospects (2019)
  11. A quantum-inspired classical algorithm for recommendation systems (2019)
  12. Comparing Predictive Accuracy (1995)
  13. Portfolio Construction Through Mixed-Integer Programming at Grantham, Mayo, Van Otterloo and Company (1999)
  14. Risk-Based and Factor Investing (2016)
  15. Blocking Gibbs sampling in very large probabilistic expert systems (1995)

0 of 23 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.