The prediction of volatility is of primary importance for business applications in risk management, asset allocation and pricing of derivative instruments. This paper proposes a novel measurement model which takes into consideration the possibly time-varying interaction of realized volatility and asset returns, according to a bivariate model aiming at capturing the main stylized facts: (i) the long memory of the volatility process, (ii) the heavy-tailedness of the returns distribution, and (iii) the negative dependence of volatility and daily market returns. We assess the relevance of \volatility in volatility" and time-varying \leverage" effects in the out-of-sample forecasting performance of the model, and evaluate the density forecasts of the future level of market volatility. The empirical results illustrate that our specification can outperform the benchmark
HAR-GARCH model, both in terms of point and density forecasts.