| 摘要: | 本文探究投資人情緒是否會影響比特幣交易行為、以及比特幣日報酬及日報酬波動(7日、30日)會受到哪些因素影響。因此，本文延用了Antweiler and Frank (2004)、Cookson and Niessner (2020)所使用的投資人看漲訊號指標(Bullishness Signal, BS)與投資人看法分歧度指標(Agreement Index, AI)，透過2013年10月至2020年12月共53,415筆Bitcointalk論壇文章，以Python的VADER套件為工具做文字分析，並運用機器學習LSTM模型建構了共四組情緒指標，根據文獻納入區塊鏈資訊、總體經濟、全球貨幣匯率、情緒指標共28個變數建構模型，探究變數對比特幣日報酬及日報酬波動(7日、30日)之間的關聯性。
實證結果發現，區塊鏈資訊、總體經濟、全球貨幣匯率加上四組情緒指標共28個變數中，共計5個變數為比特幣日報酬預測之重要變數，分別為比特幣算力(BitcoinHashRate)、黃金日報酬(Gold)、恐慌指數(VIX)、英鎊匯率變化率(GBP)、人民幣匯率變化率(CNY)。共計9個變數為比特幣日報酬波動(7日、30日) 預測之重要變數，分別為上海證交所綜合股價指數變化率(SSE)與四組情緒指標(投資人看漲訊號、投資人看法分歧度指標為一組，共計四種建立方式)。
在建構情緒指標方面，本文運用機器學習LSTM模型，訓練模型的準確性達到95.19%，驗證模型的準確性達到91.13%，並將模型建立的四組情緒指標作為解釋變數探究對比特幣日報酬、日報酬波動預測性。由實證結果發現，本文透過投資人論壇文章建構的情緒指標能夠有效地預測比特幣日報酬波動。
;This study is intended to explores whether investor sentiment would affect bitcoin trading behavior, and what factors would affect bitcoin daily return and daily return volatility. Therefore, this study cites two sentiment factors used by Antweiler and Frank (2004) and Cookson and Niessner (2020): Bullishness Signal (BS) and Agreement Index (AI). The factors construction process uses the VADER package in Python as a text analysis tool, and uses the Long Short-Term Memory(LSTM) model to construct a total of four sets of sentiment factors. The sample data is a total of 53,415 Bitcointalk forum articles from October 2013 to December 2020. According to the literature, we included 28 variables of blockchain information, macro economic development, global currency exchange rates, and sentiment factors to construct a regression model.
The empirical results found that out of 28 variables including blockchain information, macro economic development, global currency exchange rates, and sentiment factors, a total of 5 variables are important variables for Bitcoin daily return prediction, including BitcoinHashRate. , Golden Daily Return (Gold), VIX, British Pound Exchange Rate (GBP), RMB Exchange Rate (CNY). A total of 9 variables are important variables for Bitcoin daily return volatility prediction, including the Shanghai Stock Exchange Composite Index return (SSE) and four sets of sentiment factors.
In terms of constructing sentiment factors, this study uses machine learning LSTM model. The accuracy of the training model reached 95.19%. The empirical results found that the sentiment fators constructed in this study through Bitcointalk forum articles can effectively predict the daily return volatility of Bitcoin. |
