Paper: arXiv 2512.21791

Authors: Christophe D. Hounwanou, Yae Ulrich Gaba, Pierre Ntakirutimana

Abstract

Data scarcity and confidentiality in finance often impede model development and robust testing. This paper presents a unified multi-criteria evaluation framework for synthetic financial data and applies it to three representative generative paradigms: the statistical ARIMA-GARCH baseline, Variational Autoencoders (VAEs), and Time-series Generative Adversarial Networks (TimeGAN). Using historical S and P 500 daily data, we evaluate fidelity (Maximum Mean Discrepancy, MMD), temporal structure (autocorrelation and volatility clustering), and practical utility in downstream tasks, specifically mean-variance portfolio optimization and volatility forecasting. Empirical results indicate that ARIMA-GARCH captures linear trends and conditional volatility but fails to reproduce nonlinear dynamics; VAEs produce smooth trajectories that underestimate extreme events; and TimeGAN achieves the best trade-off between realism and temporal coherence (e.g., TimeGAN attained the lowest MMD: 1.84e-3, average over 5 seeds). Finally, we articulate practical guidelines for selecting generative models according to application needs and computational constraints. Our unified evaluation protocol and reproducible codebase aim to standardize benchmarking in synthetic financial data research.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 9.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematical frameworks including generative models (VAE, TimeGAN) with formal objectives (ELBO, adversarial losses) and statistical measures (MMD, autocorrelation), while also presenting rigorous empirical benchmarks with specific quantitative results on S&P 500 data and a reproducible codebase.

Research Flowchart

  flowchart TD
  Start["Research Goal: Evaluate Synthetic Financial Data Generators for Financial Modelling"]
  
  Data["Data: Historical S&P 500 Daily Data"]
  
  Models["Methodology: Three Generative Paradigms<br>1. ARIMA-GARCH (Statistical Baseline)<br>2. Variational Autoencoder (VAE)<br>3. Time-series GAN (TimeGAN)"]
  
  Process["Computational Evaluation<br>Fidelity: Maximum Mean Discrepancy (MMD)<br>Structure: Autocorrelation & Volatility Clustering<br>Utility: Portfolio Optimization & Volatility Forecasting"]
  
  Outcomes["Key Findings:<br>ARIMA-GARCH: Captured linear trends, missed nonlinear dynamics<br>VAE: Smooth trajectories, underestimated extreme events<br>TimeGAN: Best trade-off (Lowest MMD: 1.84e-3), high temporal coherence"]
  
  Start --> Data
  Data --> Models
  Models --> Process
  Process --> Outcomes

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 a unified multi-criteria evaluation framework for synthetic financial data covering distributional similarity, temporal coherence, volatility patterns, downstream task performance, and visual structure.

Source sentence
A multi-criteria evaluation framework assessing distributional similarity, temporal coherence, volatility patterns, downstream task performance, and visual structure via PCA.

Assumes 4

  • Using a single asset provides a controlled setting for comparison. data
    Source sentence
    Using a single asset provides a controlled setting for comparison.
  • Log‑returns transformation yields a stationary series suitable for models. model
    Source sentence
    The raw price series was converted into a stationary and numerically stable representation suitable for both econometric and deep learning models.
  • Temporal train/validation/test split avoids forward‑looking leakage. execution
    Source sentence
    We adopt a temporal split to avoid forward-looking leakage.
  • Synthetic datasets are generated with the same size as the real dataset. data
    Source sentence
    Each model generates synthetic datasets matched in size to the real dataset.

Data

  • S&P 500 daily closing prices · S&P 500 index market-prices public
    January 2000 to March 2024 · S&P 500 index
    Source sentence
    We use daily closing prices of the S&P 500 index from January 2000 to March 2024.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • TimeGAN achieves the lowest Maximum Mean Discrepancy (MMD) of 1.84e-3, indicating the best overall realism among the models. statistical
    Source sentence
    TimeGAN attained the lowest MMD: 1.84e-3, average over 5 seeds
  • In portfolio optimization, TimeGAN’s synthetic data yields a Sharpe ratio difference of -0.03 and a weight cosine similarity of 0.91, closest to real‑data performance. performance
    Source sentence
    TimeGAN achieves the closest replication of the real-data portfolio, with a cosine similarity of 0.91 and a Sharpe difference of only -0.03.
  • For volatility forecasting, TimeGAN reduces RMSE to 0.103 and improves directional accuracy to 63%, outperforming ARIMA‑GARCH and VAE. predictive-accuracy
    Source sentence
    TimeGAN again achieves the best alignment with real-data volatility, reducing RMSE by 17% relative to ARIMA–GARCH and improving directional accuracy to 63%.
  • ARIMA‑GARCH captures linear trends and conditional volatility but fails to reproduce nonlinear dynamics. negative
    Source sentence
    ARIMA‑GARCH captures linear trends and conditional volatility but fails to reproduce nonlinear dynamics
  • VAEs produce smooth trajectories that underestimate extreme events. negative
    Source sentence
    VAEs produce smooth trajectories that underestimate extreme events

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 55 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 (33)

  1. The synthetic data vault (2016)
  2. Synthetic data: closing the gap between development and real-world deployment (2022)
  3. Modeling tabular data using conditional GAN (2019)
  4. Synthetic data—a privacy-friendly way to share and analyze data (2022)
  5. Privacy-preserving synthetic data generation: A survey (2022)
  6. Time Series Analysis: Forecasting and Control (2015)
  7. Generalized autoregressive conditional heteroskedasticity (1986)
  8. Auto-encoding variational Bayes (2014) — link
and 25 more
  1. Time-series generative adversarial networks (2019)
  2. Empirical properties of asset returns: stylized facts and statistical issues (2001)
  3. Analysis of Financial Time Series (2010)
  4. Synthetic financial time series generation for risk and portfolio management (2022)
  5. Evaluation of generative models for synthetic time-series data: a systematic approach (2021)
  6. Regulation (EU) 2016/679 of the European Parliament and of the Council — General Data Protection Regulation (GDPR) (2016) — link
  7. Application of Financial Data in Risk Management and Portfolio Optimization (2012)
  8. The Synthetic Data Vault (2016) — link
  9. Robust De-anonymization of Large Sparse Datasets (2020) — link
  10. Portfolio Selection (1952) — link
  11. Risk Management (2000)
  12. Statistical Fraud Detection: A Review (2002) — link
  13. Algorithmic Trading: Winning Strategies and Their Rationale (2013)
  14. Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs (2017) — link
  15. QuantGANs: Deep Generation of Financial Time Series (2020) — link
  16. DPGAN: Differentially Private Generative Adversarial Network (2019) — link
  17. Generation and Evaluation of Synthetic Data: A Survey (2020) — link
  18. Deep Generative Models for Synthetic Data Generation: A Comparative Study (2022)
  19. Time-series Generative Adversarial Networks (2019)
  20. Generative Adversarial Nets (2014)
  21. Auto-Encoding Variational Bayes (2013) — link
  22. TabDDPM: Modelling tabular data with diffusion models (2023)
  23. Diffusion models for time-series and tabular data: A survey (2022) — link
  24. Synthcity: A benchmark framework for diverse algorithms and data modalities in synthetic data generation (2023)
  25. Evaluation of deep generative models for financial time series (2023)

0 of 33 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.