Interpreting conflicting rankings of volatility forecasts out of sample

Ascertain robust and interpretable evaluation frameworks for out-of-sample realized variance forecasting that reconcile conflicting rankings across different loss functions (RMSE, MAE, QLIKE, MZ-regression R^2) and aggregation schemes (average firm, average cross-section, pooled panel).

Background

Using multiple loss functions and aggregation approaches, the authors find that different volatility forecasting models are ranked inconsistently across panels and metrics, reflecting known ambiguities in volatility forecast evaluation. This inconsistency hampers clear model selection and motivates improved interpretive methods.