Application of Neural Networks for Financial Time Series Forecasting

The thesis aims to design, implement and evaluate a solution for financial time series prediction using neural networks on real data. Specifically, it focuses on the design and implementation of neural network models for predicting price movements of selected assets and creating a trading strategy based on these predictions. The objective of the work is to determine whether the proposed strategy has the potential to achieve long-term profitability. The theoretical part of the thesis introduces the basic concepts of the financial market, working with financial time series and the basics of neural networks. The practical part includes the acquisition and preprocessing of intraday minute data of 15 US stocks for the period from April 16, 2021 to September 1, 2023. Furthermore, the design and implementation of TFT and N-HiTS neural network models for predicting asset price movements is carried out, with a total of 30 N-HiTS and 30 TFT models. Two TFT and N-HiTS models were created for each of the 15 stocks, one working with resampled data at a frequency of 5 minutes and the other with data at a frequency of 15 minutes. The last part of the practical section describes the implementation of a trading strategy driven by the predictions of the TFT and N-HiTS models. The performance of this strategy is measured by backtesting against a proposed passive investment strategy that serves as a benchmark. The results of the study indicate the short-term profitability of the proposed solution on a 9-week backtesting horizon, which performed better on average than the passive investment strategy. The paper concludes with a discussion of the next steps for a comprehensive evaluation of the long-term profitability of the current solution and further directions for its development.