Paper: arXiv 2410.18897

Authors: Tomonori Takahashi, Takayuki Mizuno

Abstract

Despite its practical significance, generating realistic synthetic financial time series is challenging due to statistical properties known as stylized facts, such as fat tails, volatility clustering, and seasonality patterns. Various generative models, including generative adversarial networks (GANs) and variational autoencoders (VAEs), have been employed to address this challenge, although no model yet satisfies all the stylized facts. We alternatively propose utilizing diffusion models, specifically denoising diffusion probabilistic models (DDPMs), to generate synthetic financial time series. This approach employs wavelet transformation to convert multiple time series (into images), such as stock prices, trading volumes, and spreads. Given these converted images, the model gains the ability to generate images that can be transformed back into realistic time series by inverse wavelet transformation. We demonstrate that our proposed approach satisfies stylized facts.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 5.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematical concepts like diffusion models and wavelet transformations, indicating high complexity, while demonstrating the approach with concrete experimental setups and stylized fact reproduction, showing significant empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: Generate realistic synthetic<br>financial time series"] --> B["Input: Real Financial Data<br>(Prices, Volumes, Spreads)"]
  B --> C["Wavelet Transform: Convert time series to images"]
  C --> D["Diffusion Model DDPM Training<br>on converted images"]
  D --> E["Model Generation: Create synthetic images"]
  E --> F["Inverse Wavelet Transform: Convert back to time series"]
  F --> G["Outcome: Validated synthetic data<br>Satisfies stylized facts"]

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 denoising diffusion probabilistic models (DDPMs) combined with wavelet transformation to generate synthetic financial time series.

Source sentence
We alternatively propose utilizing diffusion models, specifically denoising diffusion probabilistic models (DDPMs), to generate synthetic financial time series. This approach employs wavelet transformation to convert multiple time series (into images), such as stock prices, trading volumes, and spreads.

Assumes 6

  • The methodology assumes simultaneous generation of three synchronized observed financial time series: prices, bid/ask spreads, and trading volumes. data
    Source sentence
    In the subsequent description, we assume the simultaneous generation of three synchronized observed financial time series: prices, bid/ask spreads, and trading volumes.
  • Time series are expanded by mirror expansions at both ends to achieve a length of 2n before wavelet transformation. model
    Source sentence
    Initially, because sequence data with size 2n simplify the subsequent discrete wavelet transformation, we expand the time series by mirror expansions at both ends to align the length of the time series to 2n.
  • The Haar wavelet is used as the mother wavelet for the discrete wavelet transformation. model
    Source sentence
    For the transformation, we employed the Haar wavelet as the mother wavelet.
  • Power conversion parameters are set to p = 1.5 for price log returns and p = 1.0 for spreads and trading volumes. model
    Source sentence
    Parameters: In preprocessing, we adopt power conversion parameter p = 1.5 for price log returns and p = 1.0 for the spreads and trading volumes.
Show 2 more
  • Winsorization is applied with a level of 10.0, replacing outliers beyond 10σ with 10σ. model
    Source sentence
    The win-sorization level is set at 10.0, meaning that outliers beyond 10σ are replaced by 10σ.
  • The three time series are treated as three channels of an RGB image for generation. model
    Source sentence
    To leverage its advantages (excellent image generation), we convert the time series data into images to train the model. After training, the model generates images, which we then convert back into the original time series data.

Data

  • AAPL minute-based stock prices, spreads, and trading volumes · Refinitiv (LSEG) market-prices commercial
    January 2005 to December 2014 · AAPL.O on NASDAQ
    Source sentence
    Data: In our experiment, we selected minute-based stock prices, spreads, and trading volumes of AAPL.O traded on NASDAQ from January 2005 to December 2014. We purchased the data from Refinitiv (LSEG), a financial data provider.

Recognised sources: Refinitiv / LSEG (Datastream, Eikon, Tick History)

Costs

Not applicable The paper reports no trading or execution result.

Finds 1

  • The DDPM with wavelet imaging more effectively replicates financial stylized facts—including fat tails, slow decay of autocorrelations (volatility clustering), intraday seasonality, and cross‑correlation structures—than TimeGAN, QuantGAN, and a simple DDPM wit statistical
    Source sentence
    The DDPM and wavelet imaging approaches more effectively replicated characteristics commonly observed in financial time series, such as the fat tails, the slow decay of autocorrelations including volatility clustering, the intraday seasonal-ity patterns, and the cross correlation coefficients among time series, compared to TimeGAN and QuantGAN approaches and the simple application of DDPM without wavelet imaging.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 5 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 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.