Rank-Invariance Conditions for the Comparison of Volatility Forecasts

The paper derives four conditions which guarantee rank–invariance, i.e. that the empirical rankings (based on measurement error–affected variance proxies) of competing volatility forecasts be consistent with the true rankings (based on the unobservable conditional variance). The first three establish bounds beyond which the separation between the forecasts is enough for their rankings not to be affected by the measurement error. The conditions’ ability to establish rank-invariance with respect to forecast characteristics, such as bias, variance and correlation, is studied via Monte Carlo simulations. An additional moment condition identifies the functional forms of the triplet {model, estimation criterion, loss} for which the effects of measurement errors on the rankings cancel altogether. Both theoretical and empirical results show the extension of admissible loss functions achieving ranking consistency in forecast evaluations.