Synthetic Stocks and Market Movement Estimation: An Analysis with Gaussian Process, Drawdown, and Data Science

In this study, Gaussian Process regression (GPR) is implemented as a method to denoise observed financial data and to make predictions at densely interpolated time points. The methodology is applied to various empirical financial datasets. A Gaussian kernel with data-dependent initialization is implemented to derive predictive means and confidence intervals. With this procedure, synthetic data points are created from the denoised dataset for better prediction accuracy. The drawdown of the denoised dataset is calculated using the predicted mean derived by GPR, and the model for predictions is trained with those values. Finally, various machine/deep learning-based approaches are implemented to show that the data densification with GPR improves the detection of upcoming significant fluctuation of a given financial dataset. It is shown that with shorter time intervals of the data, the improvements are more significant.