統計與機器學習方法的資產價格預測—以比特幣為例

Abstract

資產價格的預測性在市場效率假說提出後便受到眾多討論，至今亦有許多研究係利用此概念在不同的金融資產上進行實證分析。隨著近年加密貨幣的快速發展，其價格劇烈的增長吸引許多投機客進入該市場，而使得發展及波動更加劇烈，在市值大幅成長的同時也引來學者的警告，諾貝爾經濟學獎得主Joseph Stiglitz 與Robert Shiller便稱其為「危險的投機泡沫」。本研究同時利用傳統時間序列模型與機器學習模型，對比特幣報酬與價格進行預測並比較預測表現。我們發現機器學習模型對報酬進行預測，有較小的預測誤差，表現優於傳統時間序列模型，然而透過繪圖則顯現出其較佳的預測效果係來自較低幅度的波動，實際上預測表現並不理想。而後利用組合預測法改善，根據Diebold-Mariano 檢定的結果，雖然比起單一傳統時間序列模型來得要好，卻也是源自低幅度的波動所致。基於上述結果，我們認為比特幣的價格或報酬的可預測性較低。The predictability of asset prices has been widely discussed since the introduction of the market efficiency hypothesis and has been applied to various financial assets. With the rapid development of cryptocurrencies in recent years, their volatile price growth has attracted many speculators to enter the market, leading to even more intense development and fluctuations. As a result, scholars have issued warnings, with Nobel laureates in economics Joseph Stiglitz and Robert Shiller referring to it as a"dangerous speculative bubble". This study simultaneously employs both traditional time series models and machine learning models to predict the returns and prices of Bitcoin, comparing their predictive performances. We found that machine learning models have smaller prediction errors than traditional time series models in predicting returns. However, visualizations show that superior predictive performance of machine learning models is due to lower magnitude volatility. Based on the results, weconclude that predicting prices or returns of Bitcoin is much infeasible.