Generation of synthetic data to improve financial prediction models

Abstract

The main issue at the heart of this dissertation is the improvement of ML financial prediction systems through the use of augmented training data. An overview and assessment of various financial data synthesis methods has been carried out. To be able to accomplish this assessment, a review of evaluation methods had to be undertaken keeping in mind that asset price time sequences have characteristics that will be missed if only the typical statistical measures are used. This is because asset time sequences have particular patterns that play out over the time dimension in addition to the magnitude dimension. Augmented asset price time sequences may be useful in a number of scenarios. High fidelity synthetic data may be needed to substitute real data to preserve confidentiality. The temporal portions of the data that are of interest may be short as is the case for financial bubbles or short lived changes in the behavior of market participants which require the training data to exhibit these particular characteristics repeatedly. In addition to this, ML systems improve in their ability to deal with unseen data if they are trained on larger datasets. The objective of this work is to choose the best performing (in terms of fidelity) financial data synthesis method and to verify that the data thus generated actually improves the performance of an asset price prediction system. We undertook to use a number of qualitative visual evaluation metrics that helped confirm the quantitative assessments carried out. This resulted in SigCWGAN being selected. The asset time series generated using this GAN were then used to train an ARIMA/RNN asset price prediction system. The data obtained using SigCWGAN did, in fact result in an improvement in the performance of the predictive system we used in terms of MAE, MSE and MAPE metrics.