Realized Volatility Forecasting with Heterogeneous Autoregressive Models

We extend the Heterogeneous Autoregressive model for Realized Variance (HAR-RV) to incorporate corrections for microstructure noise present in high-frequency intraday data. The study develops a noise-robust realized kernel estimator and evaluates forecast accuracy using a battery of out-of-sample loss functions under realistic transaction-cost assumptions. The methodology is applied to a diverse set of equity and index instruments, with an emphasis on model stability across varying liquidity regimes.