Predicting Emerging and Frontier Stock Markets Using Deep Neural Networks

Abstract

Investors, researchers and finance practitioners are continuously looking for the best technique that can assist them in accurately predicting the stock markets. The ability to predict stock prices contradicts the efficient market hypothesis (EMH) and can yield substantial monetary rewards for investors. Various stock price prediction techniques are used to predict the stock market and they range from statistical to machine learning methods. Statistical models fall short in handling nonlinear data which characterizes most stock markets. Artificial Neural Networks (ANNs), one of the widely used techniques are able to handle nonlinear data but have low prediction accuracy due to their inability to handle long term dependencies and memory capacity handling. Prediction models that have an ability to learn long-term dependency information are ideal for stock market prediction. The current study uses deep learning techniques, namely, Long Short Term Memory (LSTM), Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs), Bidirectional LSTM (BLSTM), Bidirectional RNN (BRNN), Bidirectional GRU (BGRU) to predict stock markets in ten sub-Saharan African countries. The prediction techniques were run on a python 3.5 environment using Theano and Keras libraries. Limited computing capacity was of great concern. However, for the purpose of this study, access to high performance computing facilities was granted in order to run the experiments. Experimental results show that both unidirectional and bidirectional architectures greatly improved prediction accuracy in this research. However, both architectures were found not to be significantly different in predicting the stock markets of the ten African countries. In general, LSTMs followed by BGRUs proved to be the best models in predicting the African stock markets.