Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthetic Financial Data Generation for Enhanced Financial Modelling

Christophe D. Hounwanou, Yae Ulrich Gaba, Pierre Ntakirutimana

TL;DR

The paper tackles data scarcity and confidentiality in finance by proposing a unified, multi-criteria framework to evaluate synthetic financial time series. It systematically compares three representative paradigms—ARIMA-GARCH, Variational Autoencoders, and TimeGAN—on S&P 500 daily returns across distributional fidelity, temporal coherence, downstream task performance, and privacy leakage. TimeGAN consistently provides the best overall realism and downstream utility, while ARIMA-GARCH offers interpretability and efficiency, and VAEs yield stable but over-smoothed sequences; these trade-offs support a practical trilemma in model selection. The work delivers actionable guidelines and a reproducible benchmarking pipeline to standardize synthetic data research in finance, enabling safer benchmarking, stress testing, and confidential experimentation.

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.

Synthetic Financial Data Generation for Enhanced Financial Modelling

TL;DR

The paper tackles data scarcity and confidentiality in finance by proposing a unified, multi-criteria framework to evaluate synthetic financial time series. It systematically compares three representative paradigms—ARIMA-GARCH, Variational Autoencoders, and TimeGAN—on S&P 500 daily returns across distributional fidelity, temporal coherence, downstream task performance, and privacy leakage. TimeGAN consistently provides the best overall realism and downstream utility, while ARIMA-GARCH offers interpretability and efficiency, and VAEs yield stable but over-smoothed sequences; these trade-offs support a practical trilemma in model selection. The work delivers actionable guidelines and a reproducible benchmarking pipeline to standardize synthetic data research in finance, enabling safer benchmarking, stress testing, and confidential experimentation.

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.
Paper Structure (72 sections, 14 equations, 7 figures, 11 tables)

This paper contains 72 sections, 14 equations, 7 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. The generative goal: a unified formal perspective
  4. The statistical foundation: Interpretability at a cost
  5. Bridging the gap: temporal and hybrid models
  6. The unresolved challenge: A fragmented landscape
  7. Summary of synthetic data generation methods
  8. Methodology
  9. Dataset and preprocessing
  10. Data source
  11. Preprocessing steps
  12. Theoretical foundations of generative methods
  13. Experimental setup and implementation details
  14. Generative models
  15. ARIMA - GARCH (statistical baseline)
  16. ...and 57 more sections

Figures (7)

  • Figure 1: Full methodology pipeline, including preprocessing, model training, synthetic data generation, and evaluation.
  • Figure 2: Short sequences of real and synthetic S&P 500 log-returns.
  • Figure 3: PCA projection of real and synthetic datasets into the 2D subspace defined by the first two principal components of the real data.
  • Figure 4: QQ-plots of normalized log-returns. Left: Real data; Middle: TimeGAN; Right: VAE. Deviations from the diagonal line indicate differences in distribution tails.
  • Figure 5: Autocorrelation function (ACF) for real returns and squared returns, showing linear and volatility persistence structures. Similarity in ACF patterns indicates temporal fidelity of synthetic sequences.
  • ...and 2 more figures