Introduction: Demand prediction is essential part of business planning. An accurate and timely sales prediction is essential for stock management. That is a crucial part for food wholesales and retail profitability. The stock includes large assortment of goods, some of them require special storage conditions, some are quickly perishable. There are general and product specific causes of the demand fluctuation. The variations in consumer demand may be influenced by price (change), promotions, changing (rapid or gradual, global or local) consumer preferences or weather changes. Furthermore, a large share of the products is sensitive to some form of a seasonal change. Seasonal changes occur due to different cultural habits, national, school or religious holidays, fasting. All these factors imply that some types of products have high sales during a limited period of time. Although seasonal patterns are expected, the predictive features that define these seasons are not always directly observed. Therefore, fluctuations in sales which are accommodated by the changing seasons are often difficult to predict. Besides, the historical data is often highly imbalanced, i.e. occasion specific products would have only a few weeks of the sales peaks per year. The main idea of the context aware approach that we propose in this work is to select the predictor based on the structural properties of time series. Different products have different sales behavior and different dependence on calendar events (seasonality). If we can identify and extract distinct categories of products, specific input data construction procedures and specific predictors could be employed for each category. One could argue, that an ensemble approach with a rigorous feature selection and predictor selection  does that automatically. However, this approach has limitations with respect to a given food sales prediction problem. First of all, the data is noisy and relatively short. For example, a particular food wholesaler company keeps only two years record in their transactional database. If a product is seasonal and peaks only once per year for a particular event, we would have only one or two positive examples in the historical data. By defining the context, we filter out a part of noise. Secondly, some series share common patterns. For example, New Year peaks are common for a large subset of products. By categorizing the time series based on their structural properties, we narrow down the job for the particular predictor, allowing to focus on the peculiarities of a particular series.
|Title of host publication||Proceedings 21st Benelux Conference on Artificial Intelligence (BNAIC'09, Eindhoven, The Netherlands, October 29-30, 2009)|
|Editors||T. Calders, K. Tuyls, M. Pechenizkiy|
|Publication status||Published - 2009|