Following established theories from about the effects of signals on asset prices, through discerning of sentiments from social media and news, we propose and develop a causal filtering approach for deep learning-based predictions for cryptocurrency prices. Using time-series data about cryptocurrency prices and approximately 24 sentiment indices measured in time, we develop a two-stage process to predict prices. In the first stage, we apply time-series causality derived from information theory we filter out signals from noise. Next using the signals with the highest causal scores, we use the Long-Short Term Memory (LSTM) model of recurrent neural networks (RNNs). Our results depict very high predictability with extremely low loss functions for both day-to-day predictions, and for short-term interval-based predictions for Ethereum and Bitcoin. However, compared to the baseline of the 5-Day average and relative strength indicator (RSI), our methods are less performant in predicting prices.
