Paper: arXiv 2401.10370

Authors: Lars Ericson, Xuejun Zhu, Xusi Han, Rao Fu, Shuang Li, Steve Guo, Ping Hu

Abstract

In the financial services industry, forecasting the risk factor distribution conditional on the history and the current market environment is the key to market risk modeling in general and value at risk (VaR) model in particular. As one of the most widely adopted VaR models in commercial banks, Historical simulation (HS) uses the empirical distribution of daily returns in a historical window as the forecast distribution of risk factor returns in the next day. The objectives for financial time series generation are to generate synthetic data paths with good variety, and similar distribution and dynamics to the original historical data. In this paper, we apply multiple existing deep generative methods (e.g., CGAN, CWGAN, Diffusion, and Signature WGAN) for conditional time series generation, and propose and test two new methods for conditional multi-step time series generation, namely Encoder-Decoder CGAN and Conditional TimeVAE. Furthermore, we introduce a comprehensive framework with a set of KPIs to measure the quality of the generated time series for financial modeling. The KPIs cover distribution distance, autocorrelation and backtesting. All models (HS, parametric and neural networks) are tested on both historical USD yield curve data and additional data simulated from GARCH and CIR processes. The study shows that top performing models are HS, GARCH and CWGAN models. Future research directions in this area are also discussed.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced deep learning architectures like GANs, WGANs, and diffusion models with mathematical formulations, but also includes comprehensive backtesting frameworks, comparative model performance on real yield curve data, and specific KPIs for financial modeling readiness.

Research Flowchart

  flowchart TD
  A["Research Goal: <br>Compare Deep Generative Models<br>for Financial Time Series in VaR"] --> B
  
  subgraph B ["Data & Inputs"]
      B1["Historical USD Yield Curve Data"]
      B2["Simulated GARCH & CIR Processes"]
  end

  B --> C["Methodology: Applied & Proposed Models"]
  
  subgraph C ["Models"]
      C1["Existing Models"]
      C2["Proposed New Methods"]
  end

  C1 --> C1a["CGAN, CWGAN"]
  C1 --> C1b["Diffusion, Signature WGAN"]
  C2 --> C2a["Encoder-Decoder CGAN"]
  C2 --> C2b["Conditional TimeVAE"]

  C --> D["Comprehensive Evaluation Framework"]
  D --> E["KPIs & Backtesting"]
  
  subgraph E ["Evaluation Metrics"]
      E1["Distribution Distance"]
      E2["Autocorrelation Structure"]
      E3["VaR Backtesting"]
  end

  E --> F["Key Findings/Outcomes"]
  
  F --> F1["Top Performers: <br>HS, GARCH, CWGAN"]
  F --> F2["Proposed Methods: <br>Encoder-Decoder CGAN & TimeVAE<br>Validated for Conditional Gen"]
  F --> F3["Future Directions: <br>Further Refinement of Generative Models"]

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 two new deep learning methods for conditional multi-step time series generation: Encoder-Decoder CGAN and Conditional TimeVAE.

Source sentence
we propose and test two new deep learning methods for conditional multi-step time series generation in Section 2, which are Encoder-Decoder CGAN (Fu (2022)) and Conditional TimeVAE (VAE for short).

Assumes 3

  • Only models that generate conditional distribution are considered. model
    Source sentence
    For this study, only models that generate conditional distribution are considered.
  • Volatility clustering is an important feature that models must capture. market
    Source sentence
    Once in a high (low) volatility market episode, it is expected that the market stays in the high (low) volatility episode for certain period of time. This phenomenon is called volatility clustering.
  • Synthetic data that are too close to real data may not be useful complements to real data. data
    Source sentence
    if the synthetic data are too close to real data (e.g. having a very high correlation), then they may not be useful complements to real data.

Data

  • USD yield curve data market-prices unknown
    Source sentence
    All models (HS, parametric and neural networks) are tested on both historical USD yield curve data and additional data simulated from GARCH and CIR processes.
  • Simulated data from GARCH process synthetic unknown
    Source sentence
    All models (HS, parametric and neural networks) are tested on
  • Simulated data from CIR process synthetic unknown
    Source sentence
    All models (HS, parametric and neural networks) are tested on
  • Market risk dataset (8 key time series) market-prices unknown
    2007-2022 · SPX, VIX, 3M, 10Y, CDX IG, CDX HY, USD/EUR, Oil
    Source sentence
    we compile another dataset for market risk, which includes eight key time series spanning the period 2007-2022: SPX and VIX for equity market, 3M and 10Y forward Treasury rates for fixed income market, North American CDX IG and HY for credit market, USD/EUR exchange rate for foreign exchange market, 1M Oil futures for

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Historical simulation (HS) model performs the best across all simulated and historical datasets. performance
    Source sentence
    HS model perform the best across all simulated and historical datasets used in model testing.
  • GARCH model (with conditional t distribution) is the best parametric model, ranking only behind HS. performance
    Source sentence
    GARCH model (for returns, with conditional t distribution in particular) is consistently the best parametric model, and ranked only behind HS model.
  • CWGAN is the consistently best performing deep generative model across different datasets. performance
    Source sentence
    CWGAN is consistently the best performing deep generative model across different datasets.
  • Other deep learning models (SIG, CGAN, VAE) show good performance on some datasets but are not consistent across all. performance
    Source sentence
    Other deep learning models such as SIG, CGAN and VAE also have good performance for some datasets, however the performance is not consistent across all datasets.
  • Simulated data with very fat tails (GARCH and t distribution with three degrees of freedom) produce a different model ranking. statistical
    Source sentence
    Simulated data from GARCH and t distribution with three degrees of freedom have very different model ranking from other datasets.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 5 papers in the archive (5 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.