Print Email Facebook Twitter Impact of graph-based features on Bitcoin prices Title Impact of graph-based features on Bitcoin prices Author van Schetsen, Anouk (TU Delft Electrical Engineering, Mathematics and Computer Science) Contributor Oosterlee, Kees (mentor) Degree granting institution Delft University of Technology Programme Applied Mathematics | Financial Engineering Date 2019-06-28 Abstract Predicting the trends in Bitcoin market prices is a very challenging task due to the many uncertainties and variables influencing the market value. The market is susceptible to quick changes, causing seemingly random fluctuations in the Bitcoin price. Due to the chaotic and highly volatile nature of Bitcoin behavior, investments come with high risk. To minimize the risk involved, knowledge of the Bitcoin price movement in the future is desirable. Different studies have shown that Machine Learning algorithms can predict, to varying degrees, the price fluctuations of Bitcoin. However, most researches do not explore the relationship between the price and other features outside the transaction network, such as market capitalization, Bitcoin mining speed, or entity behavior. Also, most of the features are extracted from the network level, which means obtaining the number of transactions, users, Bitcoins mined, etc. In this research, we focus on additional features, such as features outside the transaction network and node-based features inside the transaction network, which could improve the price prediction of Bitcoin. The investigated features are the “fairness and goodness” measure and the “1-ARW-betweenness cen- trality” measure. Fairness and goodness are entity behavior measures. The goodness of a Bitcoin address captures how much this address is liked/trusted by other addresses, while the fairness of a Bitcoin address captures how fair the address is in rating other addresses’ likeability or trust level. The 1-ARW-betweenness centrality is a feature based on absorbing random walks. The feature captures the extent to which a Bitcoin address has control over the money flow between different addresses. A benchmark, based on the machine learning algorithm Random Forest with commonly used features, is used to test the impact of the additional features. The Random Forest tries to predict the sign (up-down movement) of the price per day, using data from the two previous days. Comparing this benchmark with a similar model, but then including the additional features, will gain more information about how these additional features influence the Bitcoin price. Subject Machine LearningRandom ForestBitcoinrandom walkgraph analysis To reference this document use: http://resolver.tudelft.nl/uuid:363d443c-64f6-4c35-9671-4092aa334923 Part of collection Student theses Document type master thesis Rights © 2019 Anouk van Schetsen Files PDF Thesis_AnoukvanSchetsen.pdf 5.79 MB Close viewer /islandora/object/uuid:363d443c-64f6-4c35-9671-4092aa334923/datastream/OBJ/view