Paper: arXiv 2512.21798

Authors: Christophe D. Hounwanou, Yae Ulrich Gaba

Abstract

Synthetic financial data provides a practical solution to the privacy, accessibility, and reproducibility challenges that often constrain empirical research in quantitative finance. This paper investigates the use of deep generative models, specifically Time-series Generative Adversarial Networks (TimeGAN) and Variational Autoencoders (VAEs) to generate realistic synthetic financial return series for portfolio construction and risk modeling applications. Using historical daily returns from the S and P 500 as a benchmark, we generate synthetic datasets under comparable market conditions and evaluate them using statistical similarity metrics, temporal structure tests, and downstream financial tasks. The study shows that TimeGAN produces synthetic data with distributional shapes, volatility patterns, and autocorrelation behaviour that are close to those observed in real returns. When applied to mean–variance portfolio optimization, the resulting synthetic datasets lead to portfolio weights, Sharpe ratios, and risk levels that remain close to those obtained from real data. The VAE provides more stable training but tends to smooth extreme market movements, which affects risk estimation. Finally, the analysis supports the use of synthetic datasets as substitutes for real financial data in portfolio analysis and risk simulation, particularly when models are able to capture temporal dynamics. Synthetic data therefore provides a privacy-preserving, cost-effective, and reproducible tool for financial experimentation and model development.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces formal mathematical frameworks for portfolio optimization and generative models, but focuses on empirical evaluation using real financial data (S&P 500) with statistical tests and downstream tasks, resulting in moderate scores in both dimensions.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Generate realistic synthetic financial data<br>for portfolio and risk modeling"] --> B["Data Used:<br>Historical S&P 500 Daily Returns"]
  B --> C{"Computational Process:<br>Deep Generative Models"}
  C --> D["Time-series GAN<br>TimeGAN"]
  C --> E["Variational Autoencoder<br>VAE"]
  D --> F["Downstream Evaluation"]
  E --> F
  F --> G["Key Findings:<br>TimeGAN captures volatility & correlations<br>VAE smooths extreme movements<br>Synthetic data viable for financial tasks"]

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 using deep generative models, specifically TimeGAN and VAEs, to create realistic synthetic financial return series for portfolio optimization and risk modeling.

Source sentence
This paper investigates the use of deep generative models, specifically Time-series Generative Adversarial Networks (TimeGAN) and Variational Autoencoders (VAEs) to generate realistic synthetic financial return series for portfolio construction and risk modeling applications.

Assumes 4

  • Log‑returns of the S&P 500 are stationary and can be standardized to zero mean and unit variance before modeling. statistical
    Source sentence
    Stationarity is verified using the Augmented Dickey-Fuller (ADF) test. The series is then standardized to zero mean and unit variance before being input into the generative models.
  • The study focuses on a single market index, assuming results generalize to similar equity markets. market
    Source sentence
    The empirical study uses daily closing prices of the S&P 500 index from January 2000 to June 2024.
  • Baseline TimeGAN and VAE architectures are sufficient to capture the relevant temporal dynamics. model
    Source sentence
    The deep models considered were limited to baseline architectures; richer sequential variants could potentially capture more nuanced dependencies.
  • Synthetic datasets can be generated under comparable market conditions to the real data. data
    Source sentence
    Using historical daily returns from the S&P 500 as a benchmark, we generate synthetic datasets under comparable market conditions

Data

Backtest sample: S&P 500 stocks · equities · daily

  • S&P 500 daily closing prices · S&P market-prices public
    January 2000 to June 2024 · S&P 500 index
    Source sentence
    The empirical study uses daily closing prices of the S&P 500 index from January 2000 to June 2024.
  • Synthetic TimeGAN S&P 500 returns synthetic unknown
    S&P 500 index
    Source sentence
    When applied to mean–variance portfolio optimization, the resulting synthetic datasets lead to portfolio weights, Sharpe ratios, and risk levels that remain close to those obtained from real data.
  • Synthetic VAE S&P 500 returns synthetic unknown
    S&P 500 index
    Source sentence
    The VAE provides more stable training but tends to smooth extreme market movements, which affects risk estimation.

Costs

Unclear The classifier was not confident.

Finds 4

Reported performance (TimeGAN synthetic data trained portfolio): Sharpe 0.84 (benchmark 0.89)

  • TimeGAN synthetic data closely match real data in distributional shape, volatility, autocorrelation, and yield portfolio weights, Sharpe ratios, and risk metrics similar to those from real data. statistical
    Source sentence
    TimeGAN produces synthetic data with distributional shapes, volatility patterns, and autocorrelation behaviour that are close to those observed in real returns.
  • Synthetic portfolios trained on TimeGAN achieve Sharpe Ratio 0.84, Sortino Ratio 1.26, and Max Drawdown 25.1%, comparable to real‑data trained portfolios (Sharpe 0.89, Sortino 1.31, Drawdown 23.4%). performance
    Source sentence
    Table 4: Portfolio Performance Metrics (Synthetic vs. Real Training Data)
  • VAE synthetic data are more stable but tend to under‑represent extreme market movements, leading to slightly lower volatility and risk metric values. statistical
    Source sentence
    The VAE provides more stable training but tends to smooth extreme market movements, which affects risk estimation.
  • Synthetic datasets enable robust, privacy‑preserving experimentation, with portfolio weights remaining stable across multiple random seeds. other
    Source sentence
    Across repetitions, portfolio weights remained stable, confirming that synthetic datasets can reliably support downstream quantitative finance

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.

Cited by 1 paper 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 (19)

  1. Deep learning in finance (2017) — link
  2. Empirical asset pricing via machine learning (2020)
  3. A survey of synthetic data generation methods for privacy preservation and data augmentation (2022)
  4. Analysis of financial time series (2010)
  5. Generative adversarial nets (2014)
  6. Auto-encoding variational Bayes (2014) — link
  7. Time-series generative adversarial networks (2019)
  8. Data augmentation using generative adversarial networks for time series classification (2019)
and 11 more
  1. Evaluation of synthetic financial time series for deep learning models (2023)
  2. Portfolio selection (1952)
  3. Auto-encoding variational Bayes (2013) — link
  4. Time-series generative adversarial networks (2019)
  5. Modeling tabular data using conditional GAN (2019)
  6. Synthetic data generation for financial applications: A review (2021)
  7. Quant GANs: Deep generation of financial time series (2020)
  8. The synthetic data vault (2016)
  9. Deep neural networks and tabular data: A survey (2022)
  10. Fast Global Alignment Kernels (2011)
  11. Evaluating generative models for synthetic financial data — → 2512.21791

1 of 19 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.