基於深度學習之比特幣交易費預測

Abstract

隨著區塊鏈(blockchain)發展越來越成熟，區塊鏈技術在資訊、金融等各個產業帶來許多創新，特別是金融產業由於區塊鏈技術的發展，讓全世界的使用者能不必再受限於傳統金融體系或是政府，自由的進行去中心化(decentralization)的交易。但是區塊鏈技術的交易費是能夠調整的，交易費的多寡影響出塊的優先度，能優先出塊的交易費預測成為一項令人頭痛的問題。以區塊鏈技術最大的應用比特幣(Bitcoin)為例，若能夠精準地預測比特幣交易費，不但用戶本身能夠節省大筆的交易費，虛擬貨幣錢包與交易所也能因精準預測並或提供用戶建議而達到獲利。 本研究提出(1)利用重建内存池(mempool)狀態來估算交易何時被確認(confirmed)，(2)運用深度學習(deep learning)技術來建立預測模型，結合以上來進行交易費預測。首先運行Bitcoin Core 來紀錄每筆交易在進入内存池時的狀態，並模擬重建内存池來按照費率大小排序出當交易進入内存池後會在第幾塊後被確認。最後根據收集到的資料，運用深度學習中的長短期記憶模型(Long Short-Term Memory, LSTM) 進行模型訓練，並利用此模型進行比特幣交易費預測。

With the development of blockchain technology becoming increasingly mature, it has brought many innovations in various industries, such as information and finance. Especially in the financial industry, users worldwide can freely make decentralized transactions without being controlled by the traditional financial system or the government. Bitcoin is one of the largest and most popular applications of blockchain technology, but the prediction of bitcoin transaction fees is a problem. Being able to accurately predict bitcoin transaction fees can save a lot of costs for bitcoin users, crypto wallets, and exchanges. Besides, crypto exchanges can profit by providing users with accurate bitcoin transaction fee predictions. This research(1) will use the memory pool (mempool) state to estimate when the transaction is confirmed, and then(2) use deep learning technology to build a prediction model. It combines these two methods to predict bitcoin transaction fee. First, running Bitcoin Core to record the state of each transaction when it enters the mempool, and then simulating the rebuilding of the mempool to sort the transactions by the fee rate. We can get which block each transaction will be confirmed after the transaction is sorted according to the fee rate. Finally, according to the data, we use the long short-term memory (LSTM) model in deep learning to predict the transaction fee.