Financial Time Series Prediction Using Recurrent Neural Network

OData support
Petróczi Attila István
Department of Telecommunications and Media Informatics

In economy the choice between the available investment possibilities is closely related to the expected value of their future development. If we can estimate the changes of a financial instrument’s future value, we can reduce the risks and increase the profit of our investments. Due to the complexity of economic processes it is difficult to make simple heuristic considerations about investment decisions. To achieve true financial success, an intelligent approach is needed.

Nowadays, the application of decision support systems is a widespread approach for tradesman to achieve better results. Also the employment of fully autonomous robot dealers isn’t considered to be a novel technique and the solutions based on artificial neural networks are very common among these. The ability to learn so to independently pick up profitable trade tricks is an attractive feature of artificial neural networks. The profit that can be made using automated systems constructed to use predictions based on neural nets can easily surpass the result of usual investment opportunities.

In my thesis, I am pursuing the data mining approach to examine the possibility of how neural nets based artificial intelligence can be implemented to completely eliminate the human involvement required for investments. I am designing a system that can trade stock shares profitably on its own. In my work, I pay special attention to the real life applicability of my solution. I demonstrate how costs of trading and the individual behaviour of exchange rates can be taken into consideration when creating a forecasting model. I use evolutionary algorithms to find the best modelling parameters. I will endeavour to extend the conventional neural models in a way that we can include our explicit knowledge about economical processes in their operation. I focus my attention on the phenomenon of repetitive price change behaviours ignored by conventional neural solutions in the hope that this can improve my trading results.

I implement the designed system and measure its performance using exchange rate data derived from real stock markets. In this form, I can finally describe the investment success of the model I have created in terms of earned money.


Please sign in to download the files of this thesis.