The creation and effective practical application of a brain-like artificial intelligence is one of today’s fastest growing and developing research area of artificial neural networks. During the last few years, a number of models were born that efficiently managed previously unsolvable tasks.
The theory of recurrent neural networks was a big step forward towards modeling nonlinear dynamic system behaviours. Increasing the networks complexity allowed the temporal internal states to follow the characteristics and predict future values of chaotic systems.
The intent of this thesis is the modeling and short-term forecasting of real-time financial market movements with recurrent neural networks. Due to dimension reduction, various denoise and decomposition methods are applied to the time-series. Apart of the results of the reduction, many technical analysis indicators are fed into network. Comparing the results of the training and testing phases, further parametric optimization is implemented in order to improve the forecasting accuracy. Finally, respectively to the predicted short-term movements, traditional trading and portfolio management strategies are applied. The success of the full system is measured by the backtest results for the past few years.