Paper: arXiv 2311.16004

Authors: Szymon Kubiak, Tillman Weyde, Oleksandr Galkin, Dan Philps, Ram Gopal

Abstract

We present a novel process for generating synthetic datasets tailored to assess asset allocation methods and construct portfolios within the fixed income universe. Our approach begins by enhancing the CorrGAN model to generate synthetic correlation matrices. Subsequently, we propose an Encoder-Decoder model that samples additional data conditioned on a given correlation matrix. The resulting synthetic dataset facilitates in-depth analyses of asset allocation methods across diverse asset universes. Additionally, we provide a case study that exemplifies the use of the synthetic dataset to improve portfolios constructed within a simulation-based asset allocation process.

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 employs advanced generative neural networks (GANs, Encoder-Decoder architectures) and sophisticated data conditioning, indicating high mathematical complexity. It is highly data- and implementation-heavy, using real-world fixed income datasets, time-series transformations, and a detailed case study with portfolio performance evaluation, showing strong empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Improve Asset Allocation in Fixed Income"] --> B["Methodology: Synthetic Data Generation<br>(CorrGAN + Encoder-Decoder)"]
  B --> C["Input Data:<br>Real Historical Market Data"]
  C --> D["Step 1: Generate Synthetic Correlation Matrices<br>(Enhanced CorrGAN)"]
  D --> E["Step 2: Generate Synthetic Feature Data<br>(Encoder-Decoder conditioned on Correlations)"]
  E --> F["Outcome: Comprehensive Synthetic Dataset"]
  F --> G["Final Result: Enhanced Portfolio Construction<br>via Simulation-Based Allocation"]

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 enhancing CorrGAN to generate synthetic correlation matrices and an Encoder-Decoder model to sample additional asset data conditioned on those matrices, creating a synthetic dataset for fixed income asset allocation.

Source sentence
Our approach begins by enhancing the CorrGAN model to generate synthetic correlation matrices. Subsequently, we propose an Encoder-Decoder model that samples additional data conditioned on a given correlation matrix.

Assumes 4

  • Synthetic datasets can reduce reliance on historical data and improve out-of-sample performance of asset allocation methods. data
    Source sentence
    We hypothesize that the synthetic dataset, with its reduced reliance on historical data, can address the limitations of empirical datasets and improve the performance of asset allocation methods during out-of-sample periods.
  • Portfolios are constructed once and not re‑balanced during the out‑of‑sample period. execution
    Source sentence
    Once the initial investment is made, the portfolio allocations are not adjusted.
  • Tracking error volatility (TEV) measured as the standard deviation of portfolio returns relative to the benchmark is an appropriate risk metric. statistical
    Source sentence
    tracking error volatility (TEV; measured as the standard deviation of the portfolio’s returns relative to the benchmark).
  • The optimizer aims to minimize 1‑month deviations from the benchmark while meeting an excess return target. model
    Source sentence
    the optimizer aims to minimize the 1-month deviations of the portfolio from the benchmark, reducing the tracking error volatility, while ensuring that the expected return is on average higher than or equal to a specified target above benchmark’s expected return.

Data

  • Empirical market-prices unknown
    40 fixed income indices
    Source sentence
    Simulated returns (Rassets) and expected returns are derived from three different datasets: the empirical dataset ("Empirical" variant), the synthetic dataset ("Synthetic" variant), or a combination of both ("Combined" variant).
  • Synthetic synthetic proprietary
    40 fixed income indices
    Source sentence
    For the further parts of the analysis, we created a “synthetic” dataset by sampling 64,000 correlation matrices from the GAN
  • Combined synthetic proprietary
    40 fixed income indices
    Source sentence
    Simulated returns (Rassets) and expected returns are derived from three different datasets: the empirical dataset ("Empirical" variant), the synthetic dataset ("Synthetic" variant), or a combination of both ("Combined" variant).

Recognised sources: Bloomberg

Costs

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

Finds 3

  • Synthetic and combined datasets yield lower tracking error volatility and higher average excess expected returns than the empirical dataset, especially under forward‑return simulations where TEV is about 5 bps lower. performance
    Source sentence
    The variants utilizing the synthetic dataset ("Synthetic" and "Combined") demonstrated lower TEVs and higher average excess expected returns compared to the "Empirical" variant. This difference was particularly notable in the "FR Sims" options, where the TEVs of the "Synthetic" and "Combined" variants were 5 bps lower
  • Excess return to TEV ratios improve significantly with synthetic data, confirmed by one‑sided t‑tests with p‑values below 1%. statistical
    Source sentence
    The improvements in the excess expected returns to TEV ratios were confirmed by one-sided t-tests, which yielded p-values lower than 1%.
  • Combined synthetic‑empirical datasets often outperform purely synthetic datasets, showing higher t‑statistics in both simulation modes. performance
    Source sentence
    Notably, the "Combined" variants exhibited higher t-stats than the "Synthetic" variants in both the "FR Sims" and "MV Sims" options, suggesting that augmenting the empirical dataset may offer more benefits to practitioners than solely replacing it with a synthetic dataset.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 24 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 (28)

  1. Corrgan: Sampling realistic financial correlation matrices using generative adversarial networks (2020)
  2. cCorrGAN: Conditional correlation GAN for learning empirical conditional distributions in the elliptope (2021)
  3. Checkerboard artifact free sub-pixel convolution: A note on sub-pixel convolution, resize convolution and convolution resize (2017) — link
  4. Variational Autoencoders: A Hands-Off Approach to Volatility (2022) — link
  5. Auto-encoding variational bayes (2013) — link
  6. Generative adversarial networks (2020)
  7. Nips 2016 tutorial: Generative adversarial networks (2016) — link
  8. Tactical investment algorithms (2019)
and 20 more
  1. Building diversified portfolios that outperform out-of-sample (2016)
  2. Generative adversarial networks for financial trading strategies fine-tuning and combination (2021)
  3. Alleviating class imbalance in actuarial applications using generative adversarial networks (2021)
  4. Tackling the Exponential Scaling of Signature-Based Generative Adversarial Networks for High-Dimensional Financial Time-Series Generation (2022)
  5. Matrix evolutions: synthetic correlations and explainable machine learning for constructing robust investment portfolios (2021)
  6. Unsupervised representation learning with deep convolutional generative adversarial networks (2015) — link
  7. Autoencoding Conditional GAN for Portfolio Allocation Diversification (2022) — link
  8. Scenario generation for market risk models using generative neural networks (2022)
  9. Validation of machine learning based scenario generators — → 2301.12719
  10. Tail-gan: Nonparametric scenario generation for tail risk estimation (2022) — link
  11. Improved training of wasserstein gans (2017)
  12. Deconvolution and checkerboard artifacts (2016)
  13. Machine learning for asset managers (2020)
  14. A mean/variance analysis of tracking error (1992)
  15. Generative Adversarial Networks in finance: an overview (2021) — link
  16. On variational learning of controllable representations for text without supervision (2020)
  17. A robust estimator of the efficient frontier (2016)
  18. A unified approach to interpreting model predictions (2017)
  19. Pagan: Portfolio analysis with generative adversarial networks (2019) — link
  20. Deep hedging, generative adversarial networks, and beyond (2021) — link

1 of 28 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.