This study considers the ability of the Component-GARCH model to capture the stylized features of volatility in 14 stocks traded on the Stock Exchange of hong Kong. The relative merits of several GARCH models nested in the Component-GARCH are investigated using the standard likelihood ratio test. The results suggest that the decomposition of the conditional variance into a permanent or long-run and a transitory or short-run component significantly improves the goodness-of-fit. A roll-over estimation method is then used to present out-of-sample tests of the forecasting ability of both the GARCH and Component-GARCH models. Although the traditional ARCH model slightly outperforms the Component-GARCH when forecasting short-term volatility, it is shown that only the latter model provides accurate volatility forecasts at longer time horizons. Similar findings were obtained using weekly retuns, confirming the robustness of the results.