This article presents the first results of a study which deals with the detection and the correction of abnormal values in data series intended to forecast demand. This work fits in the broader context of performance management for proximity retailers. Indeed, when this kind of point of sales (POS) is studied, sales volumes are often too small to be effectively exploited by statistical processing methods. It is therefore useful to consolidate the information with expertise and additional knowledge resulting from similar POS. It is also relevant to take into account the inherent uncertainty of such information. The proposal of this paper is a methodological contribution which uses consolidated knowledge to detect and correct abnormal values and to improve the quality of data used to implement forecast methods.