An Investigation on Multi-step Bitcoin Prediction based on LSTM and GRU

Abstract

This work aims to conduct an investigation on 2-, 5- and 10 –output-step with 5 fixed input-step Bitcoin price prediction, using gated recurrent unit (GRU) and long short-term memory (LSTM). The effects of using 2 and 6 layers with regard to LSTM and GRU (2L- and 6L- LSTM and GRU) on the prediction performance are also examined. Two datasets with statistically distinct features, e.g., rather monotonic and non-monotonic, consecutively referred to Binance and Poloniex, the world's leading crypto trading and cryptocurrency exchange platforms are experimented for intensifying the investigation. Prediction performance evaluations include root mean square error (RMSE) and mean absolute error (MAE) along with Pearson correlation coefficient (Corr) are employed here. The best averaged results of all the measures are generated by 2L-GRU. 0.9873, 0.9777 and 0.9593 Corr means are generated by 2-, 5- and 10- output-step; and 0.9758, 0.9575 and 0.9259 Corr means are resulted by the same numbers of steps, respectively for Binance and Poloniex. Overall prediction performance based on more-simpler, monotonic Binance data is rather better than more – complicate, non-monotonic Poloniex data.