Abstract Bitcoin has became one of the most popular investment asset recent years. The volatility of bitcoin price in financial market attracting both investors and researchers to study the price changing manners of bitcoin. Existing works try to understand the bitcoin price change by manually discovering features or factors that are assumed to be reasons of price change. However, the trivial feature engineering consumes human resources without the guarantee that the assumptions or intuitions are correct. In this paper, we propose to reveal the bitcoin price change through understanding the patterns of bitcoin blockchain transactions without feature engineering. We first propose k-order transaction subgraphs to capture the patterns. Then with the help of machine learning models, Multi-Window Prediction Framework is proposed to learn the relation between the patterns and the bitcoin prices. Extensive experimental results verify the effectiveness of transaction patterns to understand the bitcoin price change and the superiority of Multi-Window Prediction Framework to integrate multiple submodels trained separately on multiple history periods.
展开▼
