僅使用交易數據進行機器學習的比特幣價格趨勢預測研究

Abstract

大多數加密貨幣市場的研究論文中，通常都是使用研究人員自行選擇的特徵變數來預測價格或對趨勢進行分類。不過，如果相對於股票市場、外匯市場等其他市場而言，加密貨幣其實是不太受市場影響的。因此，我們假設可以僅藉由交易數據中的價格、數量和交易次數來獲取足夠的資訊量，以便於預測比特幣的價格。我們在這篇論文中提出了一個框架，該框架可以依序進行比特幣的價格預測、價格趨勢和交易的建議。每個過程都可以分做一個模組來解釋。在第一個模組中，我們應用LSTM機器學習方法來預測每隔30分鐘的未來24個價格。接下來，我們建構一種轉換價格到趨勢的方法，以識別價格的走向。用一個值來代表未來24個價格的上漲或下跌。最後，我們可以通過我們提出的交易策略，基於趨勢，給出交易建議。而為了檢驗訓練模型的通用性，我們選擇了2020年中，5種不同的振幅趨勢波動情況，在每個月中的每個30分鐘時，進行交易決策。結果表明，即使趨勢線下跌了4000美元，經由一系列的模組，我們交易後的利潤率達到了將近70％。

Mostly, studies on the cryptocurrency market usually use characteristic variables hand-selected by researchers to predict prices or classify trends. However, compared to other markets such as the stock market or the foreign exchange market, the cryptocurrency is actually independent without market influences. Therefore, we assume that there is enough information to be obtained by only using the price, quantity and number of transactions in the transaction data to predict the price of Bitcoin. In this paper, we propose a framework making Bitcoin price prediction, price trend, and trading suggestion. Each process is a module. In the first module, we apply the LSTM to predict the next 24 prices with 30 minutes time interval. Next, we construct a method about transforming prices to a price movement, which is a value to represent the rise or fall of the next 24 prices. Finally, through the trading strategy we proposed, we can identify a trading action based on the trend. In addition, in order to test the versatility of models, we select 5 different amplitude fluctuations in 2020, and made trading decisions every 30 minutes in a month. The results, through a series of modules, show that even if the trendline fell by $4,000, our profit rate after trading still reached nearly 70%.