本論文包含兩個財務實證研究,第一篇將減幅固定彈性波動度 (damped constant elasticity variance; DCEV)、對數波動度 (log variance; LV) 與其他文獻上提出具優勢的隨機波動度模型做比較性分析,第二篇應用 Elton, Gruber, 與 Michaely (1990) 提出的方法找出 Heterogeneous Autoregressive (HAR) 模型的關鍵已實現波動度 (realized variance; RV)。 本文第一個議題將比較文獻上廣泛使用之隨機波動度模型的實證表現,包含LV, CEV, 非線性飄移項模型與 Hung et al. (2021) 提出的 DCEV 模型,實證結果發現DCEV 模型具有比其他模型更佳的配適度,其改善效果在市場處於劇烈波動的情況下更為明顯。 Corsi (2009)提出HAR模型,該模型架構為一簡單的迴歸模型,僅利用過去1天、1周 (5天)、1個月 (22天) 三種時間間隔的RV作為自變數,文獻多已證明其波動度預測表現顯著優於複雜的隨機波動度模型。本文第二個議題即以HAR模型為基礎,應用Elton, Gruber, 與Michaely (1990)方法,透過系統性地萃取過去35個交易日內RV 結構的資訊,嘗試找出能取代1天、1周、1個月RV 的組合。實證結果顯示當樣本期間夠長,HAR模型透過選取關鍵已實現波動度作為自變數,可有效提升預測能力。
This thesis includes two financial empirical studies. The first is a comparative analysis of the damped constant elasticity variance (DCEV) model, the log variance (LV) model, and other state-of-the-art stochastic volatility models. The second applies the model of Elton et al. (1990) to identify the key realized variance (RV) in the heterogeneous autoregressive (HAR) model. In the first study, we comprehensively compare the empirical performance of the DCEV model proposed by Hung et al. (2021) with that of the LV, CEV, and nonlinear drift (NLD) models used widely in the literature. The empirical results indicate that the DCEV model still exhibits fitting performance superior to that of competing models, especially for crisis periods in the market. Corsi (2009) proposes the HAR model, a simple regression based on three explanatory variables, including one-day, one-week (5-day), and one-month (22-day) RVs. Many studies verify that the forecasting ability of the HAR model with an indexed RV lag of (1, 5, 22) is superior to that of complicated stochastic volatility models. In the second study, we propose applying the model of Elton et al. (1990) to identify the key RVs by systematically extracting information from the past 35-day RV structure instead of using lag (1, 5, 22). The empirical results reveal that given a sufficient sample period, the HAR model with key RVs as the explanatory variables efficiently improves on the forecasting ability of the original HAR model.