Paper: arXiv 2407.17645

Authors: Carlo Nicolini, Monisha Gopalan, Jacopo Staiano, Bruno Lepri

Abstract

We present the first application of modern Hopfield networks to the problem of portfolio optimization. We performed an extensive study based on combinatorial purged cross-validation over several datasets and compared our results to both traditional and deep-learning-based methods for portfolio selection. Compared to state-of-the-art deep-learning methods such as Long-Short Term Memory networks and Transformers, we find that the proposed approach performs on par or better, while providing faster training times and better stability. Our results show that Modern Hopfield Networks represent a promising approach to portfolio optimization, allowing for an efficient, scalable, and robust solution for asset allocation, risk management, and dynamic rebalancing.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: The paper introduces advanced deep learning architectures like Modern Hopfield Networks, which involve sophisticated energy-based models and attention mechanisms, indicating high math complexity. It also demonstrates strong empirical rigor through extensive combinatorial purged cross-validation, comparison against state-of-the-art baselines, and reporting of multiple financial metrics like Sharpe and Sortino ratios.

Research Flowchart

  flowchart TD
  A["Research Goal: Apply Modern Hopfield Networks<br>to Portfolio Optimization"] --> B["Data Input: Multiple Datasets<br>Equities Asset Allocation"]
  B --> C["Methodology: Combinatorial<br>Purged Cross-Validation"]
  C --> D["Computational Process:<br>Modern Hopfield Networks"]
  D --> E["Comparison: vs. LSTM & Transformers<br>Traditional Methods"]
  E --> F["Key Findings:<br>On Par or Better Performance"]
  E --> G["Key Findings:<br>Faster Training & Better Stability"]
  F --> H["Outcomes: Efficient, Scalable<br>Robust Asset Allocation Solution"]
  G --> H

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 applying modern Hopfield networks to portfolio optimization, introducing two architectures – Hopfield Pooling and Hopfield Encoder – as alternatives to LSTM and Transformer models.

Source sentence
We present the first application of modern Hopfield networks to the problem of portfolio optimization.

Assumes 6

  • Asset returns are computed as log‑returns, assumed to be more Gaussian‑like. data
    Source sentence
    We compute asset returns as log-returns rti = log Pt,i −log Pt−1,i instead of simple returns, as they are leaner to work with and tend to be better distributed towards a Gaussian distribution (Campbell et al., 1998).
  • Only long‑only portfolios are considered; weights are non‑negative and sum to one. market
    Source sentence
    For long-only portfolios asset weights are all positive and sum to 1.
  • Transaction costs and management fees are ignored in the reported results. execution
    Source sentence
    Importantly, in order to focus on specific behaviours of our methods, in the following Results Section we have ignored transaction costs and management fees.
  • Cross‑validation uses a purging and embargo period of 21 days (≈ one business month). statistical
    Source sentence
    he purging and embargoing periods we have set 21 days, corresponding approximately to one business month.
Show 2 more
  • Within each CV fold the training set is split 80%/20% for training and validation. statistical
    Source sentence
    Within each cross-validation training fold we further subdivided the training set into two distinct consecutive sets in proportions of 80% training and 20% validation, in order to keep track of the loss function and trigger early stopping.
  • Asset weights sum to one, allowing negative weights only in short‑portfolio settings (not used here). model
    Source sentence
    One could also deal with short portfolios, having positions with the only constraint of summing to 1, thus allowing negative weights.

Data

Backtest sample: multi-asset · daily

  • ETF · Yahoo Finance market-prices public
    2006-02-06 to 2023-12-29 · 4 ETFs (AGG, VIX, DBC, VTI)
    Source sentence
    The first dataset (named ETF in the text) consists of the four ETFs used by Zhang et al. (2020), although over a different temporal interval. The period analyzed spans from February 6th 2006 to December 29th 2023.
  • STOCKS · Yahoo Finance market-prices public
    2003-01-02 to 2022-12-30 · 66 Nasdaq securities
    Source sentence
    As the second dataset (named STOCKS in the text) we have obtained the spot prices of 66 securities from the Nasdaq index, from January 2nd 2003 to December 30th 2022.
  • FAMAFRENCH · Fama‑French data library market-prices public
    1973-01-01 to 2023-11-30 · 10 industry portfolios
    Source sentence
    As the intermediate size dataset (named FAMAFRENCH in the text), we considered the daily returns from 10 industry portfolios starting from January 1st 1973 until November 30th 2023. Data were obtained from the Fama‑French data library.
  • COMMODITIES · Yahoo Finance market-prices public
    2003-01-02 to 2023-12-29 · 19 commodity futures (gold to gasoline and soybean)
    Source sentence
    from 2003 to 2023. The dataset named COMMODITIES included 19 commodities from gold to gasoline and soybean.

Recognised sources: Yahoo Finance Kenneth French Data Library

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Importantly, in order to focus on specific behaviours of our methods, in the following Results Section we have ignored transaction costs and management fees.

Finds 5

Reported performance (Hopfield Pooling (HOP-POOL)): Sharpe 1.25 (benchmark 0.83) · annual return 13.0%

  • Modern Hopfield networks achieve performance on par or better than state‑of‑the‑art LSTM and Transformer models, with faster training times and better stability. performance
    Source sentence
    Compared to state-of-the-art deep-learning methods such as Long-Short Term Memory networks and Transformers, we find that the proposed approach performs on par or better, while providing faster training times and better stability.
  • Hopfield Encoder outperforms both LSTM and Hopfield Pooling on the COMMODITIES dataset according to Sharpe‑ratio statistical tests. statistical
    Source sentence
    the Hopfield Encoder performs better than both LSTM and Hopfield Pooling in the COMMODITIES dataset.
  • On the ETF and STOCKS datasets, deep‑learning methods (LSTM, HOP‑POOL, HOP‑TRA) achieve statistically significantly higher Sharpe ratios than traditional MVO and HRP methods. performance
    Source sentence
    The superiority of the end-to-end deep learning approaches (LSTM, HOP-POOL, HOP-TRA) are also clear in the STOCKS dataset, again with statistically significant higher Sharpe ratios between the deep learning methods against the traditional ones.
  • For the COMMODITIES and FAMAFRENCH datasets, all deep‑learning methods underperform compared to traditional approaches, likely due to poor generalization across multiple regimes. negative
    Source sentence
    Here all the deep learning methods fail in comparison to traditional approaches. We believe that this result stems from a poor generalization effect in return series where multiple regimes are present.
  • A simple MHN model often surpasses more intricate architectures such as LSTMs, achieving superior performance with significantly shorter training times. performance
    Source sentence
    Our results reveal that a simple MHN often surpasses more intricate architectures such as LSTMs, achieving superior performance within significantly shorter training times.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 41 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 (47)

  1. Post-modern portfolio theory comes of age (1994)
  2. Portfolio selection (1952)
  3. Mean-absolute deviation portfolio optimization model and its applications to Tokyo stock market (1991)
  4. Deep learning for portfolio optimization (2020)
  5. Robust Bayesian Allocation (2011)
  6. The intuition behind Black-Litterman model portfolios (2002)
  7. Estimation error and portfolio optimization: a resampling solution (2007)
  8. Learning to forget: Continual prediction with LSTM (2000)
and 39 more
  1. Long short-term memory (1997)
  2. Portfolio transformer for attention-based asset allocation (2022)
  3. Attention is all you need (2017)
  4. Time2vec: Learning a vector representation of time (2019) — link
  5. Hopfield networks is all you need (2020) — link
  6. On a model of associative memory with huge storage capacity (2017)
  7. Dense associative memory for pattern recognition (2016)
  8. A Universal End-to-End Approach to Portfolio Optimization via Deep Learning (2021) — link
  9. Advances in financial machine learning (2018)
  10. Graph neural networks for deep portfolio optimization (2023)
  11. Sortino: a ‘sharper’ratio (2013)
  12. Performance measurement in a downside risk framework (1994)
  13. Modern portfolio theory (2016)
  14. Perplexity from plm is unreliable for evaluating text quality (2022) — link
  15. Portfolio optimization with drawdown constraints (2004)
  16. Fixing weight decay regularization in adam (2018)
  17. Adam: A method for stochastic optimization (2014) — link
  18. Optimization of trading systems and portfolios (1997)
  19. An automated portfolio trading system with feature preprocessing and recurrent reinforcement learning (2021)
  20. An entropy-based approach to portfolio optimization (2020)
  21. Layer normalization (2016) — link
  22. Gaussian error linear units (gelus) (2016) — link
  23. Deep residual learning for image recognition (2016)
  24. Comparing individual means in the analysis of variance (1949)
  25. An algorithm for a letter-based representation of all-pairwise comparisons (2004)
  26. The Markowitz optimization enigma: Is ‘optimized’optimal? (1989)
  27. Neural networks and physical systems with emergent collective computational abilities. (1982)
  28. Mutual fund performance (1966)
  29. The econometrics of financial markets (1998)
  30. Pooling methods in deep neural networks, a review (2020) — link
  31. Training recurrent neural networks (2013)
  32. Large associative memory problem in neurobiology and machine learning (2020) — link
  33. A new frontier for Hopfield networks (2023)
  34. Fully flexible views: Theory and practice (2010) — link
  35. On sparse modern hopfield model (2024)
  36. Markowitz Portfolio Construction at Seventy — → 2401.05080
  37. Harry Markowitz and the philosopher’s stone (2024)
  38. Sparse and Structured Hopfield Networks (2024) — link
  39. Hopfield network (2007)

1 of 47 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.