Generating Synthetic Sequential Data for Enhanced Model Training: A Generative Adversarial Net Framework

Financial data is highly noisy with often drastic and rapid distribution changes. This causes financial models with high complexity overfit the data easily and render the model unproductive when deployed in the real-life applications. To solve this problem, we present a framework, attention generative adversarial net (A-Gan), a novel architecture to better learn the distribution of the sequence data and generate realistic sequence data. A-Gan utilizes various attention-based and residual net-based as well as combined architecture as the generator and discriminator. When used for data augmentation, we find that it increases the generalizing ability and model robustness. For the framework, we introduce a number of different architectures, including transformer-based GAN, ResNet GAN, Transformer-FCN GAN as well as Transformer-ResNet GAN. We find that, out of the 50 UCR sequence datasets, 35 show performance improvements, some quite significant. The results indicate that the model was able to learn from the original data and generate new data of similar distributions which can be used to strengthen the model. We also introduce a new metric to examine the generated data and gauge the synthetic data quality, which can further enhance.