Paper: arXiv 2408.03320

Authors: Siqiao Zhao, Zhikang Dong, Zeyu Cao, Raphael Douady

Abstract

When constructing portfolios, a key problem is that a lot of financial time series data are sparse, making it challenging to apply machine learning methods. Polymodel theory can solve this issue and demonstrate superiority in portfolio construction from various aspects. To implement the PolyModel theory for constructing a hedge fund portfolio, we begin by identifying an asset pool, utilizing over 10,000 hedge funds for the past 29 years’ data. PolyModel theory also involves choosing a wide-ranging set of risk factors, which includes various financial indices, currencies, and commodity prices. This comprehensive selection mirrors the complexities of the real-world environment. Leveraging on the PolyModel theory, we create quantitative measures such as Long-term Alpha, Long-term Ratio, and SVaR. We also use more classical measures like the Sharpe ratio or Morningstar’s MRAR. To enhance the performance of the constructed portfolio, we also employ the latest deep learning techniques (iTransformer) to capture the upward trend, while efficiently controlling the downside, using all the features. The iTransformer model is specifically designed to address the challenges in high-dimensional time series forecasting and could largely improve our strategies. More precisely, our strategies achieve better Sharpe ratio and annualized return. The above process enables us to create multiple portfolio strategies aiming for high returns and low risks when compared to various benchmarks.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced mathematical frameworks like PolyModel theory and polynomial regression (Eq 2, 3), coupled with deep learning (iTransformer), but its empirical rigor is grounded in substantial data (10k+ hedge funds over 29 years) and reported performance metrics like Sharpe ratio, though lacking visible implementation details.

Research Flowchart

  flowchart TD
  A["Research Goal: Construct Hedge Fund Portfolio<br>Using PolyModel Theory & iTransformer"] --> B["Data Collection & Input"]
  
  subgraph B ["Data Sources"]
      B1["10,000+ Hedge Funds<br>29 Years Data"]
      B2["Risk Factors:<br>Indices, Currencies, Commodities"]
  end

  B --> C{"PolyModel Theory<br>Implementation"}
  
  subgraph C ["PolyModel Processes"]
      C1["Address Sparse Data<br>& High-Dimensionality"]
      C2["Calculate Quantitative Measures<br>Long-term Alpha, SVaR"]
      C3["Calculate Classical Measures<br>Sharpe Ratio, MRAR"]
  end

  C --> D["iTransformer<br>Deep Learning Enhancement"]
  
  subgraph D ["Computational Analysis"]
      D1["High-dimensional<br>Time Series Forecasting"]
      D2["Capture Upward Trends<br>Control Downside Risk"]
  end

  D --> E["Portfolio Strategy<br>Construction"]
  
  subgraph E ["Strategy Outcomes"]
      E1["High-Return Portfolio"]
      E2["Low-Risk Portfolio"]
  end

  E --> F["Key Findings & Outcomes"]
  
  subgraph F ["Performance Results"]
      F1["Better Sharpe Ratio"]
      F2["Higher Annualized Return"]
      F3["Outperforms Benchmarks"]
  end

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 combining PolyModel theory with the iTransformer deep learning model to select hedge funds and construct high‑return, low‑risk portfolios.

Source sentence
we proposed the combination of PolyModel theory and iTransformer for hedge funds selection; the proposed strategies achieved much higher returns than the standard fund of fund benchmarks.

Assumes 6

  • The noise term ϵj has zero mean and is usually assumed to follow a normal distribution. statistical
    Source sentence
    ϵj is the noise term in the regression model with zero mean; usually it is assumed to be normal distribution but does not have to be.
  • Each regression function Φij(x) is approximated by a polynomial of degree up to 4 using Hermite polynomials. model
    Source sentence
    where Hk(x) is the Hermitian polynomial of degree k.
  • A polynomial of degree 4 is sufficiently flexible to capture the essential nonlinear relationship while keeping overfitting bearable. model
    Source sentence
    Based on authors’ practical experience, a polynomial of degree of 4 is flexible enough to capture nonlinear but essential relation between target and risk factor while usually suffer bearable overfitting.
  • At each time stamp the analysis uses only the past three years of return data to compute a dynamic p‑value score. data
    Source sentence
    at each time stamp, we only look at the past 3 years’ return data, and thus, we can have a dynamic P-Value Score series for each target asset Yi and risk factor Xj pair.
Show 2 more
  • Features such as SVaR and LTS are computed using the past 36 months of data for the next month’s prediction. data
    Source sentence
    we use the past 36 months data to compute features such as SVaR and LTS for the next month’s pred…
  • The asset pool consists of over 10,000 hedge funds spanning roughly 29 years. data
    Source sentence
    utilizing over 10,000 hedge funds for the past 29 years’ data.

Data

Backtest sample: 1994-04 – ? · multi-asset · monthly

  • Hedge fund return dataset macro unknown
    April 1994 to May 2023 · over 10,000 hedge funds
    Source sentence
    The data sets cover a long period from April 1994 to May 2023.
  • Risk factor dataset macro unknown
    hundreds of risk factors from equities, coupons, bonds, industrial indexes, etc.
    Source sentence
    our study incorporates an extensive universe comprising hundreds of risk factors from different domains, including equities, coupons, bonds, industrial indexes, and more.

Costs

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

Finds 3

  • The constructed strategies achieve better Sharpe ratio and annualized return than benchmarks. performance
    Source sentence
    our strategies achieve better Sharpe ratio and annualized return.
  • Strategy SA delivers the highest cumulative return, while WA is more stable with less drawdown; both outperform benchmark indices significantly. performance
    Source sentence
    SA has the best performance regarding the cumulative return; WA is more stable and suffers much less drawdown than SA. Both strategies outperform the benchmarks significantly.
  • The proposed combination of PolyModel features and iTransformer yields substantially higher returns than standard fund‑of‑fund benchmarks. performance
    Source sentence
    the proposed strategies achieved much higher returns than the standard fund of fund benchmarks.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (1 by independent authors)

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 (55)

  1. The C omprehensive T ex A rchive N etwork ( CTAN ) (1993)
  2. On measuring nonlinear risk with scarce observations (2010)
  3. The stressvar: A new risk concept for extreme risk and fund allocation (2010)
  4. Managing the downside of active and passive strategies: Convexity and fragilities (2019)
  5. Artificial Intelligence for Financial Markets: The Polymodel Approach (2022)
  6. PolyModel: Portfolio Construction and Financial Network Analysis (2023)
  7. itransformer: Inverted transformers are effective for time series forecasting (2023) — link
  8. Attention is all you need (2017)
and 47 more
  1. Layer normalization (2016) — link
  2. Reversible instance normalization for accurate time-series forecasting against distribution shift (2021)
  3. Non-stationary transformers: Exploring the stationarity in time series forecasting (2022)
  4. The Elements of Statistical Learning: Data Mining, Inference, and Prediction (2009) — link
  5. MRAR Illustrated — link
  6. Morningstar Risk-Adjusted Return — link
  7. The Morningstar Rating for Funds — link
  8. The StressVaR: A New Risk Concept for Superior Fund Allocation (2009) — link
  9. Transformers in time series: A survey (2022) — link
  10. Using Machine Learning Technique to Enhance the Portfolio Construction Based on PolyModel Theory (2023)
  11. Fin-gan: Forecasting and classifying financial time series via generative adversarial networks (2024)
  12. Hedge Fund Research, https://www.hfr.com/hfri-indices-index-descriptions — link
  13. The capital asset pricing model: Theory and evidence (2004)
  14. A systematic review of stock market prediction using machine learning and statistical techniques (2022)
  15. Deep hedging (2019)
  16. CP-PINNs: Changepoints Detection in PDEs using Physics Informed Neural Networks with Total-Variation Penalty (2023)
  17. Tackling data bias in music-avqa: Crafting a balanced dataset for unbiased question-answering (2024)
  18. Musechat: A conversational music recommendation system for videos (2024)
  19. Mapping the Invisible: Face-GPS for Facial Muscle Dynamics in Videos (2024)
  20. Every Image Listens, Every Image Dances: Music-Driven Image Animation (2025) — link
  21. Accurate structure prediction of biomolecular interactions with AlphaFold 3 (2024)
  22. Ast: Audio spectrogram transformer (2021) — link
  23. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models (2023)
  24. Face-GPS: A Comprehensive Technique for Quantifying Facial Muscle Dynamics in Videos (2023)
  25. Learning transferable visual models from natural language supervision (2021)
  26. Manifold-constrained Gaussian process inference for time-varying parameters in dynamic systems (2023)
  27. CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables (2024) — link
  28. SRDA: Mobile Sensing based Fluid Overload Detection for End Stage Kidney Disease Patients using Sensor Relation Dual Autoencoder (2023)
  29. Leveraging mobile sensing and bayesian change point analysis to monitor community-scale behavioral interventions: A case study on covid-19 (2022)
  30. Detection and analysis of interrupted behaviors by public policy interventions during COVID-19 (2021)
  31. Multiple distresses detection for Asphalt Pavement using improved you Only Look Once Algorithm based on convolutional neural network (2024)
  32. Evaluation of asphalt pavement texture using multiview stereo reconstruction based on deep learning (2024)
  33. Let community rules be reflected in online content moderation (2024) — link
  34. An Open and Large-Scale Dataset for Multi-Modal Climate Change-aware Crop Yield Predictions (2024)
  35. CP-PINNs: Data-Driven Changepoints Detection in PDEs Using Online Optimized Physics-Informed Neural Networks (2024)
  36. MMST-ViT: Climate Change-aware Crop Yield Prediction via Multi-Modal Spatial-Temporal Vision Transformer (2023)
  37. Mm-bsn: Self-supervised image denoising for real-world with multi-mask based on blind-spot network (2023)
  38. Self-supervised image denoising for real-world images with context-aware transformer (2023)
  39. An optimal control strategy for execution of large stock orders using long short-term memory networks (2023)
  40. A Deep Neural Network Algorithm for Linear-Quadratic Portfolio Optimization With MGARCH and Small Transaction Costs — → 2301.10869
  41. Research and implementation of cancer gene data classification based on deep learning (2023)
  42. Calibrate and Debias Layer-wise Sampling for Graph Convolutional Networks
  43. On variance estimation of random forests with Infinite-order U-statistics (2024)
  44. Attention-Enhancing Backdoor Attacks Against BERT-based Models
  45. Reinforcement subgraph reasoning for fake news detection (2022)
  46. Mamba Fusion: Learning Actions Through Questioning (2024) — link
  47. CLIPScope: Enhancing Zero-Shot OOD Detection with Bayesian Scoring (2024) — link

1 of 55 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.