The class of mixed normal conditional heteroskedastic (MixN-GARCH) models, which couples a mixed normal distributional structure with GARCH-type dynamics, has been shown to offer a plausible decomposition of the contributions to volatility, as well as excellent out-of-sample forecasting performance, for financial asset returns. In this paper, we generalize the MixN-GARCH model by relaxing the assumption of constant mixing weights. Two different specifications with time--varying mixing weights are considered. In particular, by relating current weights to past returns and realized (component-wise) likelihood values, an empirically reasonable representation of Engle and Ng's (1993) news impact curve with an asymmetric impact of unexpected return shocks on future volatility is obtained. An empirical out-of-sample study confirms the usefulness of the new approach and gives evidence that the leverage effect in financial returns data is closely connected, in a non-linear fashion, to the time--varying interplay of mixture components representing, for example, various groups of market participants.