Deep learning algorithms for classification of financial time series data

Abstract

Stock trading markets are infamous for being unstable and complicated, and there is much enthusiasm by many to search for a dependable, unerring model that can be used to trade the stock markets. Long short-term memory (LSTM) networks are a variant of recurrent neural networks (RNN) and are effective in modelling time series data. Specifically, LSTM is able to circumvent the issue of long-term dependency as it has a unique unit structure for storage making it a suitable choice for financial time series classification. This paper publishes the results of training a LSTM network to classify the daily movement of 10 technology sector stocks over 5 years. The results were contrasted with Gated Recurrent Unit (GRU) networks and Recurrent Neural Networks (RNN). From the results, LSTM persistently surpasses GRU and RNN and obtains higher classification accuracies.