本文採用異質性自我迴歸(HAR-RV)模型,以精簡原則來配適具有緩長記憶性的實現波動率序列,來預測波動率並且能提供統計上的結果去解釋台股指數波動率的特性。我們希冀能配適一個能夠充分預測價格變動的波動模型,進而幫助投資人在風險管理或交易策略上提供有效的決策依據。 實證結果發現,加入了槓桿效果與交易量作為解釋變數的LHAR-RV-cum-Vol模型提供了最佳的預測模型。其中,槓桿效果顯示了具有異質性結構,而且槓桿效果在短期是由日跳躍所引起,但長期下卻不由日跳躍所引起。此外,研究結果顯示只有日交易量對於未來波動率具有顯著的解釋能力,特別是以成交筆數當做訊息流動之替代變數時,才能提供最好的預測能力。 再者,我們發現使用Corsi, Pirino and Reno(2009)的CTBPV方法來分離連續與跳躍並加入HAR-RV模型其所提供的預測能力優於Barndorff-Nielsen and Shephard(2004)的BPV方法,但兩者的差異不具有統計顯著性。最後,本研究將市場區分成多頭市場與空頭市場時,研究結果發現門檻連續與跳躍(TCJ)做為解釋變數不論在多頭或是空頭市場皆提供最好的預測能力,但是已實現冪次變異(RPV)做為解釋變數時只有在空頭市場提供最佳的預測能力。當市場處於空頭市場時,其所隱含的交易資訊會上升,進而提升了實現波度率的預測績效。
This paper employs a ’Heterogeneous Autoregressive’ (HAR) model which is suitable to parsimoniously model long memory in realized volatility time series. The purpose is to use this model to predict the future volatility and provide some statistical results to explain volatility behavior in Taiwan stock index market. We hope to provide an accurate predictive model on the volatility and then help investors with regards to risk management or trading strategies. The empirical results verify that the “best” model for volatility prediction is the LHAR-RV-cum-Vol model which includes the leverage effect and trading volume as regressors. Particularly, the leverage effect unveils a heterogeneous structure and this effect is induced by jumps for short-run prediction horizons but not for long-run prediction horizon. Besides, results reveal only daily trading volume has significant effect on future volatility, especially the number of transactions as a proxy for information flows provides the best predictive ability on the volatility. The empirical results also reveal that the HAR model adds continuous components (C) and jump components (J) extracted by Corrected Threshold Bi-power Variation (Corsi et al. 2009) to predict volatility better than Bi-power Variation (Barndorff-Nielsen and Shephard, 2004). However, we do not get significant gain derived by dividing the continuous and jump components. Lastly, this study separates the market into up-market days and down-market days. We find that the threshold continuous and jump (TCJ) as a regressor is the top forecaster in both markets, while realized power variation (RPV) only performs best on down-market days. When the market is down the amount of market information increases, the predictive ability of future volatility also increases.