Predicting digital asset market based on blockchain activity data. (arXiv:1810.06696v1 [q-fin.CP])

Blockchain technology shows significant results and huge potential for serving as an interweaving fabric that goes through every industry and market, allowing decentralized and secure value exchange, thus connecting our civilization like never before. The standard approach for asset value predictions is based on market analysis with an LSTM neural network. Blockchain technologies, however, give us access to vast amounts of public data, such as the executed transactions and the account balance distribution. We explore whether analyzing this data with modern Deep Leaning techniques results in higher accuracies than the standard approach. During a series of experiments on the Ethereum blockchain, we achieved $4$ times error reduction with blockchain data than an LSTM approach with trade volume data. By utilizing blockchain account distribution histograms, spatial dataset modeling, and a Convolutional architecture, the error was reduced further by 26\%. The proposed methodologies are implemented in an open source cryptocurrency prediction framework, allowing them to be used in other analysis contexts.