Beating the baseline prediction in food sales: How intelligent an intelligent predictor is?

Sales prediction is an essential part of stock planning for the wholesales and retail business. It is a complex task because of the large number of factors affecting the demand. Designing an intelligent predictor that would beat a simple moving average baseline across a number of products appears to be a non-trivial task. We present an intelligent two level sales prediction approach that switches the predictors depending on the properties of the historical sales. First, we learn how to categorize the sales time series into ‘predictable’ and ‘random’ based on structural, shape and relational features related to the products and the environment using meta learning approach. We introduce a set of novel meta features to capture behavior, shape and relational properties of the sales time series. Next, for the products identified as ‘predictable’ we apply an intelligent base predictor, while for ‘random’ we use a moving average. Using the real data from a food wholesales company we show how the prediction accuracy can be improved using this strategy, as compared to the baseline predictor as well as an ensemble of predictors. In our study we also show that by applying an intelligent predictor for the most ‘predictable’ products we can control the risk of performing worse than the baseline.

► We aim to predict the quantity of food product sales on a weekly basis. ► We present a two level sales prediction approach with meta learning elements. ► The first level assigns a given product to a behavioral category. ► The second level applies the predictor that is specific for this category.