This paper discusses how lognormal, Gumbel and log-Pearson type 3 distributions were eliminated as candidates for modelling at-site annual maximum discharge series in Bangladesh. Among candidate distributions, the generalized extreme value distribution best represented the statistical characteristics of observed data. Comparisons of distributions were based upon the root mean square deviation in fit, the probability plot correlation coefficient andL-moment ratio diagrams. The last method assesses whether a candidate distribution is capable of simulating the site-to-site variation in statistical characteristics that is present in observed flood samples.