How accuracy of time-series prediction for cryptocurrency pricing is affected by the sampling period

Abstract: Cryptocurrencies and their pertaining market currently succeeds $300billion and has had an all time high surpassing $850 billion. Being able to predict market movements and future valuations for a cryptocurrency would be an invaluable and very profitable tool for designing successful investment strategies. This thesis compares how time series predictions on the cryptocurrency Ether using a long short-term memory (LSTM) neural network is affected by altering the sampling period. Specifically we look at how the sampling periods of 30 minutes, 2 hours and 4 hours affect a prediction horizon of 4 hours. The results are also verified across a varying number of neurons (10, 20 and 40) for each of the two LSTM layers of the model. The results indicate that the accuracy of predictions can be improved by decreasing the sampling period of data. However there does not seem to be any clear trend how changing the number of neurons per LSTM layer affect prediction accuracy.