This paper compares model-based and reduced-form forecasts of financial volatility when high-frequency return data are available. We derived exact formulas for the forecast errors and analyzed the contribution of the “wrong” data modeling and errors in forecast inputs. The comparison is made for “feasible” forecasts, i.e., we assumed that the true data generating process, latent states and parameters are unknown. As an illustration, the same comparison is carried out empirically for spot 5-minute returns of DM/USD exchange rates. It is shown that the comparison between feasible reduced-form and model-based forecasts is not always in favor of the latter in contrast to their infeasible versions. The reduced-form approach is generally better for long-horizon forecasting and for short-horizon forecasting in the presence of microstructure noise.