Predicting the Bitcoin trend using technical indicators for deep learning algorithmic features

Abstract

This research focused on utilization of technical indicators for predicting the trend of Bitcoin/USD price with different deep learning algorithms. The goal was to achieve better trend prediction accuracy results using technical indicators compared to using only close, open, high, low and volume (OHLCV) data for Bitcoin/USD parity. Through achieving this goal, three different deep learning algorithms, Deep Neural Networks (DNN), Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) were used because of the performance they exhibit in literature for financial stock prediction domain and their theoretical convenience. 156 technical indicators, mathematical transformations and financial patterns were used in feature set to test against OHLCV data of Bitcoin/USD. Experiments in this research showed that, utilization of technical indicators produced better accuracy results compared to OHLCV data for all three prediction models. For the imbalanced dataset distribution produced by a one-way transaction cost to decide buy, hold or sell operations, LSTM performed best among the models used in this research with achieving 56.33% accuracy score with reasonable individual class prediction rates whereas raw data could achieve 53.26%. In the scenario for which the one-way transaction cost is tuned to have a uniformly distributed dataset, GRU performed best with achieving 52.19% accuracy score whereas raw data could achieve 39.85%.