Paper: arXiv 2410.09850

Authors: Sohyeon Kwon, Yongjae Lee

Abstract

In the financial sector, a sophisticated financial time series simulator is essential for evaluating financial products and investment strategies. Traditional back-testing methods have mainly relied on historical data-driven approaches or mathematical model-driven approaches, such as various stochastic processes. However, in the current era of AI, data-driven approaches, where models learn the intrinsic characteristics of data directly, have emerged as promising techniques. Generative Adversarial Networks (GANs) have surfaced as promising generative models, capturing data distributions through adversarial learning. Financial time series, characterized ‘stylized facts’ such as random walks, mean-reverting patterns, unexpected jumps, and time-varying volatility, present significant challenges for deep neural networks to learn their intrinsic characteristics. This study examines the ability of GANs to learn diverse and complex temporal patterns (i.e., stylized facts) of both univariate and multivariate financial time series. Our extensive experiments revealed that GANs can capture various stylized facts of financial time series, but their performance varies significantly depending on the choice of generator architecture. This suggests that naively applying GANs might not effectively capture the intricate characteristics inherent in financial time series, highlighting the importance of carefully considering and validating the modeling choices.

Complexity vs Empirical Score

  • Math Complexity: 5.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced mathematical concepts including stochastic processes (Ornstein-Uhlenbeck, Heston, Jump Diffusion) and KL/Jensen-Shannon divergence, but it is also heavily empirical with extensive experiments on multiple GAN architectures, hyperparameter optimization, and specific metrics for financial stylized facts.

Research Flowchart

  flowchart TD
  A["Research Goal: Assess GANs' ability to learn stylized facts of financial time series"] --> B["Data Preparation"]
  B --> B1["Univariate & Multivariate Financial Time Series"]
  B --> B2["Identify Stylized Facts<br>Random Walk, Mean Reversion, Jumps, Volatility"]
  B --> C["Methodology: GAN Architecture Comparison"]
  C --> D["Generative Process<br>Generator vs Discriminator<br>Adversarial Training"]
  D --> E{"Evaluation of Generated Series"}
  E -->|"Capture Stylized Facts?"| F["Key Findings"]
  E -->|"Inadequate"| G["Refine Model / Select Architecture"]
  G --> D
  F --> F1["GANs capture stylized facts but performance varies"]
  F --> F2["Generator architecture is critical"]
  F --> F3["Naive application is insufficient for complex time series"]

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 investigates whether GANs can learn the stylized facts of financial time series, specifically random walks, mean reversion, jumps, and time‑varying volatility, for both univariate and multivariate data.

Source sentence
This study examines the ability of GANs to learn diverse and complex temporal patterns (i.e., stylized facts) of both univariate and multivariate financial time series. In this paper, we aim to verify the ability of GANs to learn various stylized facts inherent in financial time series: random walks, mean reversions, jumps, and time-varying volatilities.

Assumes 3

  • The five stochastic processes (BM, GBM, OU, Jump diffusion, Heston) adequately represent the stylized facts of real financial time series. model
    Source sentence
    EXPERIMENT 4.1 Data description For modeling temporal structures of financial time series, this study employs five stochastic processes as introduced in Section 2.1: Brow-nian motion (BM), Geometric Brownian motion (GBM), Ornstein-Uhlenbeck process (OU), Jump diffusion model (JD), and Heston model (HT).
  • Synthetic data generated from these stochastic processes can be used to evaluate GAN learning ability. data
    Source sentence
    However, with the availability of parame-ters, stochastic processes can be simulated consistently, providing a valuable tool for generating synthetic financial time series data that can capture the characteristics of the real-world data.
  • Short training sequences of the Wiener process are sufficient for GAN training despite parameter estimation challenges. statistical
    Source sentence
    As the length of the Wiener process used as training data in this experiment is comparatively short, directly estimating the values of its parameters from the generated data is challenging.

Data

  • Synthetic Brownian motion, Geometric Brownian motion, Ornstein‑Uhlenbeck, Jump diffusion, synthetic unknown
    Source sentence
    EXPERIMENT 4.1 Data description For modeling temporal structures of financial time series, this study employs five stochastic processes as introduced in Section 2.1: Brow-nian motion (BM), Geometric Brownian motion (GBM), Ornstein-Uhlenbeck process (OU), Jump diffusion model (JD), and Heston model (HT).
  • Synthetic bivariate Geometric Brownian motion synthetic unknown
    Source sentence
    For the multivariate case, a bivariate Geometric Brownian motion is generated with two different correlation settings, high and low.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • GANs effectively learn the distributions and features of Brownian Motion and Geometric Brownian Motion. performance
    Source sentence
    The results suggest that the GAN framework effectively learns both the distributions and features of Brownian Motion and Geo-metric Brownian Motion, which represent the randomness of stock behavior.
  • The TCN generator struggles to learn the reverting speed of the Ornstein‑Uhlenbeck process, performing worse than MLP and LSTM. performance
    Source sentence
    Among the various generator architectures, MLP and LSTM are the most effective at reproducing a reverting speed close to that of the original data, while TCN shows limited perfor-mance in learning this feature.
  • Generated Jump Diffusion data shows the largest distribution distance to training data (Jensen‑Shannon divergence ≈0.52). statistical
    Source sentence
    As clearly shown in Figure 2, the log return distribution of generated data in the case of the Jump Diffusion model shows the most prominent distribution distance to that of the training data, with a value of 0.52243, regardless of the time series length.
  • In the multivariate setting, the GAN captures some correlation structure but fails to accurately reproduce individual distributions and dependencies. performance
    Source sentence
    These results indicate that while the GAN can capture some aspects of the correlations between variables, it struggles to accurately reproduce the individual distributions and dependencies in the multivariate setting.
  • Overall, GAN models capture major temporal structures in univariate data but struggle with detailed features and multivariate replication. other
    Source sentence
    In univariate data generation, the GAN models generally capture the different tempo-ral structures but struggle with detailed features. In summary, GAN models have shown potential as financial simulators without mathematical assumptions, but the detailed replication of temporal structures or multivariate time series still requires additional attention.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 4 papers in the archive (4 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 (21)

  1. Generating synthetic data in finance: opportunities, challenges and pitfalls (2020)
  2. Generative adversarial networks (2020)
  3. A style-based generator architecture for generative adversarial networks (2019)
  4. Zero-shot text-to-image generation (2021)
  5. Seqgan: Sequence generative adversarial nets with policy gradient (2017)
  6. Time-series generative adversarial networks (2019)
  7. Modeling financial time-series with generative adversarial networks (2019)
  8. Quant GANs: deep generation of financial time series (2020)
and 13 more
  1. Stock price distributions with stochastic volatility: an analytic approach (1991)
  2. Option pricing when underlying stock returns are discontinuous (1976)
  3. Jump-diffusion models for asset pricing in financial engineering (2007)
  4. A closed-form solution for options with stochastic volatility with applications to bond and currency options (1993)
  5. Generative adversarial networks for financial trading strategies fine-tuning and combination (2021)
  6. Mitigating overfitting on financial datasets with generative adversarial networks (2019)
  7. Fin-gan: Forecasting and classifying financial time series via generative adversarial networks (2024)
  8. Tail-gan: Nonparametric scenario generation for tail risk estimation (2022) — link
  9. Real-valued (medical) time series generation with recurrent conditional gans (2017) — link
  10. Corrgan: Sampling realistic financial correlation matrices using generative adversarial networks (2020)
  11. A GANs-Based Approach for Stock Price Anomaly Detection and Investment Risk Management (2023)
  12. Enhancing mean–variance portfolio optimization through GANs-based anomaly detection (2024)
  13. An Overview of Machine Learning for Asset Management. (2023)

0 of 21 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.