The Effect of the Different Data Aggregation Methods and their Detail Levels to the Prediction of Bitcoin's Exchange Rate

Abstract

Recently a growing interest can be observed in the field of financial forecasting and especially in the field of cryptocurrency market forecasting. This proved to be an outstandingly complex problem because of the many special characteristics of these markets. Making accurate predictions requires the proper usage and fine tuning of the most modern algorithms. The goal of our research was to find the optimal data division method for the LSTM neural network-based prediction of the exchange rate of Bitcoin and fine-tune the model to achieve the lowest mean percentage error possible. To fulfill this goal, two binning methods, namely transaction time-based, and transaction quantity-based binning methods were evaluated from the viewpoint of the Bitcoin exchange rate prediction. We came to the conclusion that time-based binning method outperforms the other tested method and the granularity of the optimal time division was also established. Experimental results show, that the 20-minute and 30-minute prediction interval are the most suitable choices in case of a limited amount of training data and for making more trading decisions. In case of markets with a higher commission, or when more training data are available the 2-hour prediction is recommended. Our results show that on the proper time division-based LSTM prediction method is suitable for developing successful short term trading strategies for Bitcoin markets.