Financial time series are often clustered based on conditional volatility, estimated from GARCH models. However, realized measures based on high-frequency data provide a more accurate estimation of the latent volatility process. In this paper, we assess the similarity of realized volatility dynamics using an autoregressive metric and the decomposition of volatility into good and bad components. In particular, we introduce a novel weighted algorithm for improving the hierarchical clustering ap- proach and apply it to the U.S. stocks traded in the Dow Jones Industrial Average (DJIA) index.