This paper analyzes the problem of volatility forecast comparison using imperfect volatility proxies. It shows that standard methods for comparing conditional variance forecasts can be misleading when using conditionally unbiased but imperfect volatility proxies, such as squared returns, intra-daily range, or realised volatility. The paper derives necessary and sufficient conditions for loss functions that ensure the ranking of competing volatility forecasts is robust to noise in the volatility proxy. It also provides a class of "robust" loss functions that are not simply the quadratic loss function or minor variations thereof. The paper illustrates these results with an application to the volatility of returns on IBM over the period 1993 to 2003. The study shows that using a conditionally unbiased proxy can lead to incorrect inferences and the selection of inferior forecasts over better forecasts. The paper also discusses the implications of using different loss functions in volatility forecasting and highlights the importance of choosing appropriate loss functions to ensure accurate forecast rankings. The results show that less noisy volatility proxies, such as the intra-daily range and realised volatility, lead to less distortion in forecast rankings. The paper concludes that the use of robust loss functions is essential for accurate volatility forecast comparison.This paper analyzes the problem of volatility forecast comparison using imperfect volatility proxies. It shows that standard methods for comparing conditional variance forecasts can be misleading when using conditionally unbiased but imperfect volatility proxies, such as squared returns, intra-daily range, or realised volatility. The paper derives necessary and sufficient conditions for loss functions that ensure the ranking of competing volatility forecasts is robust to noise in the volatility proxy. It also provides a class of "robust" loss functions that are not simply the quadratic loss function or minor variations thereof. The paper illustrates these results with an application to the volatility of returns on IBM over the period 1993 to 2003. The study shows that using a conditionally unbiased proxy can lead to incorrect inferences and the selection of inferior forecasts over better forecasts. The paper also discusses the implications of using different loss functions in volatility forecasting and highlights the importance of choosing appropriate loss functions to ensure accurate forecast rankings. The results show that less noisy volatility proxies, such as the intra-daily range and realised volatility, lead to less distortion in forecast rankings. The paper concludes that the use of robust loss functions is essential for accurate volatility forecast comparison.