Forex trend forecasting based on long short term memory and its variations with hybrid activation functions

Abstract

The foreign exchange (Forex) market, as one of the most important financial markets in the globe, has attracted many investors. In order to support forex traders’ trading decisions, accurately predicting the forex prices has continued to be a popular but challenging topic. Due to the high complexity of the forex market, it is always a question of how effective the forex prediction could be. With the rapid development on machine learning in the last decades, deep learning has been applied successfully to many areas including the forex market. Consequently, numerous research papers have been published, which aim to improve the accuracy of forex prediction. The Long Short-Term Memory (LSTM) neural network, a kind of artificial neural network, has been widely used, which is specially designed to analyse time series data. Due to its strong learning capability, the LSTM neural network has now been used to predict complex forex trading based on historical data. However, there is a lack of an authoritative and commonly accepted guidance on how to conduct proper forex predictions by using LSTM. The application of deep learning to financial forecasting is still in a developing stage. This research aims to investigate the feasibility of applying deep learning, particularly the LSTM neural network to the foreign exchange market and to enhance the prediction accuracy via improved LSTM algorithms. In this thesis, all of the fundamental and technical features related to forex trading have been collected and analysed comprehensively. The influential features are then selected to be used as the inputs for forex prediction. Based on these inputs, a LSTM is specifically built to predict the trends of forex prices, which are identified as a suitable prediction target for forex traders. Notably, a new validation method is also introduced to overcome the problems in the traditional time series validation methods. Furthermore, a novel LSTM algorithm using hybrid activation functions in the same hidden layer is proposed to improve the prediction accuracy for forex trend predictions. Extensive experiments have been conducted and the experimental results have shown that the performance of the LSTM with hybrid activation functions has outperformed that of the standard LSTM. The generasalibility of the hybrid activation functions based LSTM has also been proved by its successful applications to different ANNs (e.g., RNNs) and datasets.