Scholarlink

Computer-generated time series simulations have been heavily used in banks and hedge funds for risk management, pricing, volatility trading, hedging, etc. Traditionally, these simulations require binding model-based assumptions about how market prices move, which often may not be reflected in real-life, and they usually fail to capture either fat tails or non-symmetric distributions. To address this shortcoming, we propose a novel generative deep learning frameworks, built on top of generative adversarial networks (GANs) given their strong capability in simulating synthetic data, called continuous conditional transformer-based time-series GANs (CC-TTS-GAN). CC-TTS-GAN possess the ability to incorporate some market conditions (such as implied volatility or other market indicators) which are assumed to be continuous in nature, to inform their own simulations in a model-free manner. We then compare our model’s simulations to traditional benchmark models (calibrated to real data) such as geometric Brownian motion, constant elasticity of variance model and Heston stochastic volatility model, based on a wide range of evaluation metrics. The outperformance of our proposed CC-TTS-GAN over the model-based benchmarks is statistically significant. All implementation code can be found at https://github.com/Horstann/CC-TTS-GANs .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Financial Time Series Simulation with Transformer-Based Generative Models Under Continuous Conditions

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Horstann Rui Yao Ho,
  • Chi Seng Pun

摘要

Computer-generated time series simulations have been heavily used in banks and hedge funds for risk management, pricing, volatility trading, hedging, etc. Traditionally, these simulations require binding model-based assumptions about how market prices move, which often may not be reflected in real-life, and they usually fail to capture either fat tails or non-symmetric distributions. To address this shortcoming, we propose a novel generative deep learning frameworks, built on top of generative adversarial networks (GANs) given their strong capability in simulating synthetic data, called continuous conditional transformer-based time-series GANs (CC-TTS-GAN). CC-TTS-GAN possess the ability to incorporate some market conditions (such as implied volatility or other market indicators) which are assumed to be continuous in nature, to inform their own simulations in a model-free manner. We then compare our model’s simulations to traditional benchmark models (calibrated to real data) such as geometric Brownian motion, constant elasticity of variance model and Heston stochastic volatility model, based on a wide range of evaluation metrics. The outperformance of our proposed CC-TTS-GAN over the model-based benchmarks is statistically significant. All implementation code can be found at https://github.com/Horstann/CC-TTS-GANs .