Statistical flood studies allow the determination of structural or land use actions against floods. The aim of these hydrological studies is to estimate flood quantiles, relative to fixed probabilities (10-2 to 10-4). Because of the limited amount of information in a single annual maximum flood series, the extrapolation of the distribution towards large return periods gives very inaccurate estimates. So, a possible improvement concerns the use of the historical information, prior to the development of the conventional gauging network (in France, mainly after the second World War). Many authors presented theoretical analyses which concluded that even if the historical flood errors are greater than the gauging flood errors, the final uncertainty remains smaller when using additional historical information. The aim of this paper is to present a study, made in 1996 and 1997 on the Guiers river in France (near Grenoble). A collaboration between hydrologists and historians allowed to collect historical flood information since 1700, to identify the ten most important floods for three centuries, to propose an inventory-guide which gives a classification of the archives (by documents or key-words), and to make a first statistical analysis of the historical information. A review of statistical methods incorporating historical information into frequency analysis allows the treatment of this kind of historical data.