Introduction
Many previous studies have shown the importance of basic features such as transaction amount and the previous price in predicting the bitcoin price. Moreover, as already observed in previous studies that reveal the relationship between the transaction amount, and market price with the underlying graph structure, the topological data analysis (TDA) can capture these higher-order interactions\cite{abay2019chainnet}. However, in 2020, due to the rapid development of the cryptocurrency market, the market price of Bitcoin transactions has increased over sixfold, and massive large-scale international capital has poured into this market. In addition, blockchain technology based on Bitcoin has been gradually optimized, and a variety of new cryptocurrencies with different characteristics are constantly being introduced to the market, such as Ethereum, Litecoin, and Dogecoin. Therefore, the general strength of TDA in the analysis of various cryptocurrency transactions urgently needs to be further verified.
In contrast, different research has shown the high performance of vanilla Machine Learning algorithms such as Random forest, Gaussian Process, and ElasticNet in the long-term bitcoin price prediction. However, the latest efficient deep learning model in time series data prediction like RNN, LSTM, and other neural networks did not significantly improve\cite{abay2019chainnet}. Since 2017 with the introduction of the attention mechanism in neural networks, the vanilla transformer’s performance has achieved satisfactory results in many research fields \cite{vaswani2017attention}. Then in 2020, based on the vanilla Transformer, Haoyi Zhou proposed a more efficient model, Informer, which successfully enhanced the prediction capacity in the LSTM problem and used the ProbSparse Self-attention mechanism to replace the canonical self-attention for the reduction of the time complexity and memory usage\cite{zhou2020informer}.
In this research, we focus on topological properties on the dynamics graph of cryptocurrencies of Bitcoin and Ethereum and use the latest attention neural network model Informer to improve the price prediction of cryptocurrencies. At last, we will also explore the different influences of the topological properties on the various cryptocurrency datasets based on their characteristics.