TECHNICAL REPORT: On Recognition of Seasonal Predictability in SLIGRO product sales

A summary of new results on SLIGRO sales prediction. 1 The task We aim to predict one week ahead the quantity of weekly product sales, aggregated over all SLIGRO locations. 2 What’s new since ICDM paper The modifications since ICDM paper are the following: 1. Product categorization into 2 classes: ‘predictable’ and ‘unpredictable’ (previously we had 4). 2. Product categorization using the Linear Discriminant Classifier. Reasons: interpretable weight vector (importance of features), unique model for a given dataset, easy to manipulate prior probabilities and costs of missclassification. We train the model assuming equal prior probabilities, but in fact we have about 4 times more of ‘unpredictable’ products. Then we bias the discrimination threshold towards ‘unpredictable’ class, using a threshold −0.002 which was learned experimenting with the training data to match the prior probabilities of the classes. 3. We changed normalization and categorization of the time series to more domain driven approach. Now normalization works as follows: x∗ t = xt mean(x1, . . . , xT ) , (1) here xt is the original signal at time t and xT is the last instance of the training data.