Convolutional Neural Network Portfolio Management System with Heterogeneous Input

We implement a cryptocurrency portfolio management system based on a convolutional neural network architecture. We train and test several models, each augmented with data from various sources - past market information (price, volume, market capitalization), sentiment information (positive, neutral, negative sentiment scores extracted from online forums), and blockchain technical data (number of blocks and transactions per trading unit, amount paid in fees, block difficulty etc.). We show that augmenting the model with transaction volume history can lead to larger profits and higher Sharpe ratio, and augmenting the model with sentiment information can lead to better risk management.