Generative Networks for Financial Time Series Modelling and Forecasting in FinTech
Since their introduction in 2014, Generative Adversarial Networks (GANs) have attracted growing interest across a wide range of computer vision tasks, from image generation and restoration to text-to-image applications. However, comparatively little research has been conducted on the generation of time series using GANs, particularly in the field of financial data.
This thesis proposes a GAN-based model designed to generate realistic financial time series. The specific objective is to generate stock price trends in order to simulate a large number of economic scenarios under both real-world and controlled conditions.
The novelty of the proposed model, compared with the original GAN architecture, lies in its design. The generator is implemented as a sequence-to-sequence network trained to generate the final portion of a financial time series conditioned on the initial segment provided as input. The discriminator, on the other hand, is a convolutional neural network trained to determine whether the complete univariate time series presented as input is real or generated by the generator.