On Technical Trading and Social Media Indicators in Cryptocurrencies'\n Price Classification Through Deep Learning

Abstract

This work aims to analyse the predictability of price movements of\ncryptocurrencies on both hourly and daily data observed from January 2017 to\nJanuary 2021, using deep learning algorithms. For our experiments, we used\nthree sets of features: technical, trading and social media indicators,\nconsidering a restricted model of only technical indicators and an unrestricted\nmodel with technical, trading and social media indicators. We verified whether\nthe consideration of trading and social media indicators, along with the\nclassic technical variables (such as price's returns), leads to a significative\nimprovement in the prediction of cryptocurrencies price's changes. We conducted\nthe study on the two highest cryptocurrencies in volume and value (at the time\nof the study): Bitcoin and Ethereum. We implemented four different machine\nlearning algorithms typically used in time-series classification problems:\nMulti Layers Perceptron (MLP), Convolutional Neural Network (CNN), Long Short\nTerm Memory (LSTM) neural network and Attention Long Short Term Memory (ALSTM).\nWe devised the experiments using the advanced bootstrap technique to consider\nthe variance problem on test samples, which allowed us to evaluate a more\nreliable estimate of the model's performance. Furthermore, the Grid Search\ntechnique was used to find the best hyperparameters values for each implemented\nalgorithm. The study shows that, based on the hourly frequency results, the\nunrestricted model outperforms the restricted one. The addition of the trading\nindicators to the classic technical indicators improves the accuracy of Bitcoin\nand Ethereum price's changes prediction, with an increase of accuracy from a\nrange of 51-55% for the restricted model, to 67-84% for the unrestricted model.\n