ABSTRACT The main task of flood frequency analysis is to obtain design flood magnitudes from a streamflow record. The gauged record is rarely long enough to yield an estimate of an extreme flood which is sufficiently accurate to be applied with confidence in hydraulic engineering. Therefore extending a data record back in time using historical or palaeoflood data has the potential to provide a considerable amount of additional information on very large floods. Parametric estimation methods are readily applicable to flood frequency analysis when pre-gauging data is available. However, all parametric approaches need an assumption about the underlying parent distribution which is never known in hydrologic processes. A new nonparametric kernel estimation model is proposed and developed. With limited real data and simulation experiments, results show that quantiles estimated by nonparametric methods are better than those obtained by some selected parametric estimators both in terms of the descriptive ability and predictive ability. The choice of the optimum kernel function, the uncertainty of the threshold of perception value and the difference between fixed kernel and variable kernel estimators are also discussed. It is expected that the nonparametric approach will be widely used in practice as it is free of serious limitations of classical parametric models.