Abstract We use generative adversarial TimeGAN networks to create synthetic time series datasets that mirror statistical and temporal patterns of several anomalous diffusion models, with particular interest in Fractional Brownian Motion, which is prevalent in financial markets. We assess the quality of these new datasets in terms of classification accuracy of the diffusion type and regression of the anomalous diffusion exponent, which is linked to market volatility, by random forest models. We also use Standard & Poor 500 stock price data from 2012 to 2024, aiming to enhance predictive tools for financial strategies. Random forests achieve a 90% accuracy in the classification of the anomalous diffusion model and an MAE of 0.12 in the anomalous diffusion exponent regression when capturing the most significant features for each model. Random forests also show classifications and predictions consistent with existing literature when evaluating S&P 500 time series. When evaluating these models on TimeGAN-generated data, we appreciate reduced variability and inability to replicate complex dynamics like jumps, yielding a lower classification accuracy of 32%. These findings highlight the potential of synthetic data for training models in markets with limited historical records, offering a scalable approach to financial simulations. However, limitations in replicating volatility underscore the need for refined generative techniques to fully capture complex market dynamics, paving the way for future improvements in computational resources and model architectures.
Arévalo et al. (Fri,) studied this question.