Paper: arXiv 2406.11886

Authors: Haoren Zhu, Pengfei Zhao, Wilfred Siu Hung NG, Dik Lun Lee

Abstract

Financial assets exhibit complex dependency structures, which are crucial for investors to create diversified portfolios to mitigate risk in volatile financial markets. To explore the financial asset dependencies dynamics, we propose a novel approach that models the dependencies of assets as an Asset Dependency Matrix (ADM) and treats the ADM sequences as image sequences. This allows us to leverage deep learning-based video prediction methods to capture the spatiotemporal dependencies among assets. However, unlike images where neighboring pixels exhibit explicit spatiotemporal dependencies due to the natural continuity of object movements, assets in ADM do not have a natural order. This poses challenges to organizing the relational assets to reveal better the spatiotemporal dependencies among neighboring assets for ADM forecasting. To tackle the challenges, we propose the Asset Dependency Neural Network (ADNN), which employs the Convolutional Long Short-Term Memory (ConvLSTM) network, a highly successful method for video prediction. ADNN can employ static and dynamic transformation functions to optimize the representations of the ADM. Through extensive experiments, we demonstrate that our proposed framework consistently outperforms the baselines in the ADM prediction and downstream application tasks. This research contributes to understanding and predicting asset dependencies, offering valuable insights for financial market participants.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • 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 concepts like ConvLSTM and MoE transformations, requiring knowledge of neural network architectures and spatiotemporal modeling, warranting high math complexity. It includes extensive experiments with baselines, application to portfolio risk reduction, and discusses implementation details, indicating strong empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Predict Financial Asset Dependencies"] --> B["Data<br>Time-series of asset returns"]
  B --> C["Methodology<br>Asset Dependency Matrix ADM"]
  C --> D["Core Innovation<br>Static/Dynamic Transformation<br>Optimizes Asset Ordering"]
  D --> E["Computational Model<br>ADNN with ConvLSTM<br>Video Prediction Approach"]
  E --> F["Outcomes<br>Superior ADM Prediction &<br>Downstream Portfolio Performance"]

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 the Asset Dependency Neural Network (ADNN) to predict future Asset Dependency Matrices using spatiotemporal deep learning.

Source sentence
To tackle the challenges, we propose the Asset Dependency Neural Network (ADNN), which employs the Convolutional Long Short-Term Memory (ConvLSTM) network

Assumes 3

  • No short selling is allowed, i.e., asset weights are constrained to [0,1]. execution
    Source sentence
    For simplicity, we assume that no short selling is allowed, i.e. asset weight ωi ∈[0, 1].
  • Prior knowledge of the future standard deviation vector is assumed to eliminate extra variability. model
    Source sentence
    we assume prior knowledge of the future standard deviation vector to eliminate the extra variability.
  • The ADM has no natural ordering, requiring rearrangement for spatiotemporal modeling. data
    Source sentence
    assets in ADM do not have a natural order. This poses challenges to organizing the relational assets to reveal better the spatiotemporal dependencies among neighboring assets

Data

Backtest sample: 2005-09 – 2020-08 · S&P 100 / NASDAQ-100 / DJI-30 stocks · equities · monthly

  • dataset 7 market-prices unknown
    32 stocks
    Source sentence
    We randomly choose one stock dataset (dataset 7) as the source for constructing the diversified portfolio based on the predicted ADMs. All of the methods under evaluation build up their portfolios from the same 32 stocks in dataset 7

Costs

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

Finds 2

Reported performance (Asset Dependency Neural Network (ADNN)): Sharpe 2.00 (benchmark 0.45) · annual return 12.1% · max drawdown 10.7%

  • ADNN achieves lower portfolio risk (1.67×10⁻⁴) than DCC, improving risk by 13.9% over DCC and 24.8% over the Previous ADM baseline. performance
    Source sentence
    ADNN outperforms DCC with an average risk of 1.67 × 10−4, which ranks only second to the optimal portfolio. In terms of average risk, ADNN achieves an improvement of 24.8% over Previous and 13.9% over DCC.
  • PT-ADNN, which uses both positional rearrangement and MoE transformation, yields the best prediction accuracy among ADNN variants. predictive-accuracy
    Source sentence
    After comparing multiple ADNN frameworks with different transformation functions, we can observe that PT-ADNN that utilizes both positional rearrangement and MoE transform…

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 (109)

  1. Numerically stable generation of correlation matrices and their factors (2000)
  2. Monte Carlo methods in financial engineering (2004)
  3. Polynomial representation of pictures (1986)
  4. Fractal image coding: A review (1993)
  5. A review on image feature extraction and representation techniques (2013)
  6. PAGAN: Portfolio Analysis with Generative Adversarial Networks (2019) — link
  7. Convolutional LSTM network: A machine learning approach for precipitation nowcasting (2015) — link
  8. Deep multi-scale video prediction beyond mean square error (2015) — link
and 101 more
  1. Learning to decompose and disentangle representations for video prediction (2018) — link
  2. Stochastic variational video prediction (2017) — link
  3. Eidetic 3d lstm: A model for video prediction and beyond (2018)
  4. The Factor–Spline–GARCH model for high and low frequency correlations (2012)
  5. Machine Learning in Asset Management—Part 1: Portfolio Construction—Trading Strategies (2020)
  6. Machine Learning in Asset Management—Part 2: Portfolio Construction—Weight Optimization (2020)
  7. A Review on Deep Learning Techniques for Video Prediction (2020) — link
  8. Are betas best? (1978)
  9. Simple rules for optimal portfolio selection: the multi group case (1977)
  10. Out-of-sample tests of forecasting accuracy: an analysis and review (2000)
  11. Dependence structure and risk measure (2003)
  12. Estimating the dependence structure of share prices–implications for portfolio selection (1973)
  13. Sector rotation and monetary conditions (2008)
  14. On the assessment of risk: Some further considerations (1974)
  15. Accurate, large minibatch sgd: Training imagenet in 1 hour (2017) — link
  16. Betas and their regression tendencies (1975)
  17. A note on using cross-sectional information in Bayesian estimation of security betas (1973)
  18. 60 Years of portfolio optimization: Practical challenges and current trends (2014)
  19. Model-based deep reinforcement learning for dynamic portfolio optimization (2019) — link
  20. Continuous-time mean–variance portfolio selection: A reinforcement learning framework (2020)
  21. Dissemination of health information through social networks: Twitter and antibiotics (2010)
  22. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation (1982)
  23. Generalized autoregressive conditional heteroskedasticity (1986) — link
  24. Deep learning for finance: deep portfolios (2017)
  25. A deep reinforcement learning framework for the financial portfolio management problem (2017) — link
  26. Adam: A method for stochastic optimization (2014) — link
  27. Building diversified portfolios that outperform out of sample (2016)
  28. Discovering Bayesian market views for intelligent asset allocation (2018)
  29. The Hitchhiker’s Guide to the Galaxy (1995) — link
  30. Portfolio Selection (1952)
  31. Anticipating Correlations: A New Paradigm for Risk Management (2011)
  32. Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models (2002)
  33. Large dynamic covariance matrices (2019)
  34. Building diversified portfolios that outperform out of sample (2016) — link
  35. Dynamics of market correlations: Taxonomy and portfolio analysis (2003) — link
  36. Clustering and information in correlation based financial networks (2004) — link
  37. Spread of risk across financial markets: Better to invest in the peripheries (2013) — link
  38. Interconnectedness risk and active portfolio management: The information-theoretic perspective (2019) — link
  39. Asset allocation: new evidence through network approaches (2019) — link
  40. Can machine learning-based portfolios outperform traditional risk-based portfolios? The need to account for covariance misspecification (2019) — link
  41. Ten things you should know about the dynamic conditional correlation representation (2013)
  42. Analysis of financial time series (2005)
  43. On estimating the expected return on the market: An exploratory investigation (1980) — link
  44. The Markowitz Optimization Enigma: Is ‘Optimized’ Optimal? (1989) — link
  45. The effect of errors in means, variances, and covariances on optimal portfolio choice (1993) — link
  46. Leverage aversion and risk parity (2012) — link
  47. Toward maximum diversification (2008) — link
  48. Modelling time-varying correlations of financial markets (2003)
  49. Sequence to Sequence Learning with Neural Networks (2014) — link
  50. Long Short-Term Memory (1997)
  51. Generating Sequences With Recurrent Neural Networks (2014) — link
  52. On the difficulty of training Recurrent Neural Networks (2013) — link
  53. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (2017) — link
  54. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts (2018)
  55. Deep Learning (2016)
  56. Hierarchical mixtures of experts and the EM algorithm (1993) — link
  57. Unsupervised Learning of Video Representations using LSTMs (2016) — link
  58. A Review on Deep Learning Techniques for Video Prediction (2020) — link
  59. Representation Learning: A Review and New Perspectives (2013) — link
  60. Beyond Short Snippets: Deep Networks for Video Classification (2015) — link
  61. Learning Spatiotemporal Features with 3D Convolutional Networks (2015) — link
  62. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms (2017)
  63. Options on the maximum or the minimum of several assets (1987)
  64. Bivariate option pricing using dynamic copula models (2005) — link
  65. A capital asset pricing model with time-varying covariances (1988)
  66. Multivariate simultaneous generalized ARCH (1995)
  67. Multivariate GARCH models: a survey (2006)
  68. Asset pricing with a factor-ARCH covariance structure: Empirical estimates for treasury bills (1990)
  69. A multi-dynamic-factor model for stock returns (1992)
  70. Long short-term memory recurrent neural network architectures for large scale acoustic modeling (2014)
  71. Exploiting Spatio-Temporal Correlations with Multiple 3D Convolutional Neural Networks for Citywide Vehicle Flow Prediction (2018) — link
  72. End-to-end Prediction of Driver Intention using 3D Convolutional Neural Networks (2019) — link
  73. On the best quadratic approximation of nonlinear systems (2001) — link
  74. QAS: Real-Time Quadratic Approximation of Subdivision Surfaces (2007) — link
  75. Representation Learning: A Review and New Perspectives (2014) — link
  76. Image super-resolution via deep recursive residual network (2017)
  77. Learning factored representations in a deep mixture of experts (2013) — link
  78. Adaptive mixtures of local experts (1991)
  79. Forecasting Asset Dependencies to Reduce Portfolio Risk (2022)
  80. Universal Approximation Using Feedforward Neural Networks: A Survey of Some Existing Methods, and Some New Results (1998) — link
  81. Comparing unconstrained parametrization methods for return covariance matrix prediction (2022)
  82. Adaptive estimation of covariance matrices via cholesky decomposition (2010)
  83. Estimating risk parameters (1999)
  84. Regime changes and financial markets (2012)
  85. A global averaging method for dynamic time warping, with applications to clustering (2011)
  86. Business groups as hierarchies of firms: determinants of vertical integration and performance (2013)
  87. S_I_LSTM: stock price prediction based on multiple data sources and sentiment analysis (2022)
  88. A graph-based CNN-LSTM stock price prediction algorithm with leading indicators (2023)
  89. Financial time series forecasting with multi-modality graph neural network (2022)
  90. Multi-graph convolutional network for relationship-driven stock movement prediction (2021)
  91. A novel graph convolutional feature based convolutional neural network for stock trend prediction (2021)
  92. St-trader: A spatial-temporal deep neural network for modeling stock market movement (2021)
  93. Mean–variance portfolio optimization using machine learning-based stock price prediction (2021)
  94. Portfolio optimization with return prediction using deep learning and machine learning (2021)
  95. End-to-end risk budgeting portfolio optimization with neural networks (2023)
  96. Deep portfolio optimization via distributional prediction of residual factors (2021)
  97. Recent advances in reinforcement learning in finance (2023)
  98. Reinforcement-learning based portfolio management with augmented asset movement prediction states (2020)
  99. DeepTrader: a deep reinforcement learning approach for risk-return balanced portfolio management with market conditions Embedding (2021)
  100. Learning constitutive relations using symmetric positive definite neural networks (2021) — link
  101. Deep hashing-based dynamic stock correlation estimation via normalizing flow (2023)

0 of 109 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.