Extreme value analysis is one of the cornerstones of hazard quantication and risk assessment. Its basic objective is to estimate the distribution of some environmental variable X, e.g. annual maximum of the areal rainfall over some catchment, annual maximum ood, etc. This distribution can be used to estimate the exceedance probability of a given value of X (often expressed in terms of return period), or alternatively, to estimate the p-quantile of X. The estimation of quantiles is of primary importance since they are used to design civil engineering structures (e.g. dams, reservoirs, bridges) or to map hazard-prone areas where restrictions may be enforced (e.g. building restrictions in ood zones). Extreme value analysis has been the subject of extensive research, yielding an abundance of approaches. In Hydrology, several families of methods exist, including (but not limited to): Standard application of extreme value theory (EVT), i.e. estimation of an extreme value distribution based on a sample of block maxima or peaks over a high threshold Climate/Weather-informed application of EVT. This family of methods uses additional meteorological (e.g., weather type, [1]) or climatic (e.g. Interdecadal Pacic Oscillation IPO, [2]) information. Regional approaches, conjointly using data from several sites to perform the inference, which may improve the precision of estimates. Model-based approaches, using a simulation model reproducing the main characteristics of the environmental variable (e.g. [1]). In practice, users and practitioners of extreme value analyses may feel lost facing such an abundance of methods. Consequently, it is necessary to provide them with practical guidelines to choose and implement adequate methods, depending on the conditions of application (e.g. availability of long series, geographical area, type of hydrological regime, etc.). This presentation describes a methodological framework to perform a data-based comparison of competing approaches for predicting extremes. This framework is based on the following principles: The objective is to assess the predictive performance of competing methods (as opposed to standard goodness-of- evaluations). This requires decomposing the available dataset into estimation / validation sub-samples. Reliably quantifying uncertainties is recognized as a primary objective, and the issue of scrutinizing uncertainty estimates is discussed. To this aim, we make use of predictive distributions for extremes, obtained by integrating out parameter uncertainty. Such predictive distributions are standard in a Bayesian context [4] but can also be derived in a frequentist context [5]. Reliability indices are derived in order to compare the performances of competing methods on an objective basis. In a second step, this framework is used to perform a thorough comparison between approaches currently used in France for extreme prediction. The comparison is based on an extensive dataset of long series of rainfall and runo (about 40-50 years of daily data), available for hundreds of sites over France. Results demonstrate the ability of the comparison framework to distinguish between "good" and "bad" approaches, and yield valuable insights into the optimal ambit of each approach.