This dissertation consists of two chapters that examine the construction of financial market volatility indexes and their forecasting efficiency across predictive regression models. Each of the chapter is devoted to diferent volatility measures which are related and evaluated in thframework of forecasting regressions. The first chapter studies the sampling and liquidity issues in constructing volatility indexes, VIX and VXO, in emerging options market like Taiwan. VXO and VIX have been widely used to measure the 22-day forward volatility of the market. However, for an emerging market, VXO and VIX are difficult to measure with accuracy when tradings of the second and next to second nearby options are illiquid. The chapter proposes four methods to sample the option prices across liquidity proxies ¡V five different days of rollover rules ¡V for option trades to construct volatility index series. The paper finds that, based on the sampling method of the average of all midpoints of bid and ask quote option prices, the volatility indexes constructed by minute-tick data have less missing data and more efficient in volatility forecast than the method suggested by CBOE. Additionally, illiquidity in emerging options market does not, based on different rollover rules, lead to substantial biases in the forecasting effectiveness of the volatility indexes.Finally, the forecasting ability of VIX, in terms of naive forecasts and forecasting regressions, is superior to VXO in Taiwan. The second chapter uses high-frequency intraday volatility as a benchmark to measure the efficacy of model-free and model-based econometric models. The realized volatility computed from intraday data has been widely regarded as a more accurate proxy for market volatility than squared daily returns. The chapter adopts several time series models to assess the fore-casting efficiency of future realized volatility in Taiwan stock market. The paper finds that, for 1-day directional accuracy forecast performance, semiparametric fractional autoregressive model (SEMIFAR, Beran and Ocker, 2001) ranks highest with 78.52% hit accuracy, followed by multiplicative error model (MEM, Engle, 2002), and augmented GJR-GARCH model. For 1-day forecasting errors evaluated by root mean squared errors (RMSE), GJR-GARCH model augmented with high-low range volatility ranks highest, followed by SEMIFAR and MEM model, both of which, however, outperform augmented GJR-GARCH by the measure of mean absolute value (MAE) and p-statistics (Blair et al., 2001).