隨機森林演算法於GARCH模型波動性預測之改進及關鍵指標分析 - 以比特幣為例

Abstract

本研究以機器學習方法對比特幣報酬的波動率進行相關研究，並比 較一般化自回歸異質變異數模型(GARCH model)與機器學習模型對 比特幣報酬波動性的預測結果和所得出的重要影響指標探討。首先根 據過去文獻整理影響比特幣報酬波動性的外生指標並依據三個步驟進 行模型的建構和外生指標的分析。第一，運用日內資料進行實際波動 率的計算，並以隨機森林重要性排序(Random forest importance)的 方式對此實際波動率進行外生指標的挑選，依據此挑選結果進行模型 的建構和指標的分析;第二，使用 GARCH(1,1) 模型捕捉比特幣報酬 全樣本的波動性，並分別以 GARCH(1,1) 模型和機器學習模型對此波 動性進行樣本外的預測，並比較模型之間的預測結果，找出能夠最準 確對比特幣報酬波動性進行預測的模型;第三，依據具有最優預測結 果模型中的外生指標進行分析，了解影響比特幣報酬波動性預測之外 生指標及其原因。本研究實證結果顯發現，機器學習模型對預測結果 的改進可以達到預測誤差最小的效果，此外，在選擇預測比特幣報酬 波動性所使用的外生指標時，引入機器學習的相關方法可以找出具有 關鍵影響力的外生指標。

This study uses machine learning methods to study the volatility of bitcoin re- turns,compares the prediction results of the Generalized Autoregressive Heteroge- neous Variance model (GARCH model) and the machine learning model.The im- portant indicator will also be discussed.According to the past literature, the exoge- nous indicators that affect the volatility of Bitcoin’s return are sorted out. First, the realized volatility is calculated by the intraday data and sort the exogenous indica- tors of this actual volatility by Random forest importance selection; Second, use the GARCH(1,1) model and machine learning model to predict the volatility out of sample, and compare the prediction results between these models to find the model have the best prediction; Third, analyzing the exogenous indicators in models with optimal predictive outcomes to understand the affection of exogenous indicators . The empirical results shows that the improvement by machine learning method can obtain the minimize prediction error. In addition, when selecting the exogenous indicators used to predict the volatility of Bitcoin’s return, the related methods of machine learning can find the exogenous indicators with key influence.