The predictive power of the Blockhain transaction networks: Towards a new generation of network science market indicators

TL;DR

Cryptocurrency markets on public blockchains offer open data that enable network-based market indicators. The authors compare a Base Model using technical analysis and social signals with a Full Model that adds Ethereum transaction-network properties, showing that network information improves price-direction predictions. Using XGBoost on categorized daily returns, the Full Model achieves higher accuracy and favorable class-specific metrics, with network features often among the most important predictors. The work argues for a new generation of network-based financial econometrics and points to multiplex network representations as a future direction to capture multi-token, multi-application interactions on Ethereum.

Abstract

Currently cryptocurrencies and Decentralized Finance (DeFi), which enable financial services on public blockchains, represents a new growing trend in finance. In contrast to financial markets, ruled by traditional corporations, DeFi is completely transparent as it keeps records of all transactions that occur in the network and makes them publicly available. The availability of the data represents an opportunity to analyze and understand the market from the complexity that emerges from the interactions of the actors (users, bots and companies) operating in the embedded market. In this paper we focus on the Ethereum network and our main goal is to show that the properties of the underlying transaction network provide further and useful information to forecast the evolution of the market. We aim to separate the non redundant effects of the blockchain transaction network properties from classic technical indicators and social media trends in the future price of Ethereum. To this end, we build two machine learning models to predict the future trend of the market. The first one serves as a base model and considers a set of the most relevant features according to the current scientific literature including technical indicators and social media trends. The second model considers the features of the base model, together with the network properties computed from the transaction networks. We found that the full model outperforms the base model and can anticipate 46 more rises in the price than the base model and 19 more falls.

Figures (9)

• Figure 1: This figure summarizes the relations between the variables of the three family of indicators: network properties, technical analysis and social media. Panel A show the network resulting from the correlation between all the variables considered in the study. Panel B shows a filtered version of the network, only including the variables selected by the final model. Note that only significant correlation are represented as links ($\text{p-value}<0.05$)
• Figure 2: This figure compares the performance of the BM and FM models. Panel A shows a comparison between the confusion matrix of both models, reflecting that the FM model achieves a higher accuracy. Panel B compares the precision, recall and F1-score of both models for all classed, uptrends and downtrends. As it can be seen the FM model outperforms in all cases.
• Figure 3: Network representation of two transactions in the Ethereum system. Nodes represent addresses and the links ---denoted with arrows--- represent transactions between two addresses. The property of the edges is the timestamp in which transactions were added to the Blockchain.
• Figure 4: Panel a) shows the time intervals (vertical lines) and similarity score (green lines) computed with Dynamic time-slicing method. The blue line (y-axis) represents the average number of weighted events by the number of transactions. Panel b) shows the evolution of the similarity score (green line) with the close price of Ethereum (blue line).
• Figure 5: Feature importances measure by the average gain (improvement in the score) of splits which use the feature.