Outlier detection as the primary step for promotion planning in retail

Forecasting the increase of customer demand during discount promotions is a fundamental business assignment in retail, which can nowadays be performed by sophisticated data mining algorithms. The calculations are based on data gained during previous promotions. Apart from choosing the right mining algorithm, the quality of prediction models strongly depends on the quality of the training data. Outliers are points of data that do not conform to a defined notion of normal behavior and are therefore either excluded from the training set, or their impact on the model is weighted in a manner different from other data. In this paper we propose a new approach to outlier analysis, with the aim of distinguishing between the outliers associated with an outlier-generating store or product and the outliers that can be classified as noisy data. Outlier analysis is performed with a multidimensional perception of the dataset typical for data warehousing and OLAP. We also introduce measures that estimate the probability of store or product being an outlier generator and conduct experiments to determine their critical threshold values.