A clustering based forecast engine for retail sales

Efficient and accurate sales forecasting is a vital part of creating an efficient supply chain in enterprises. Times series methods are a popular choice for forecasting demand sales. A major challenge is to develop a relatively inexpensive and automated forecasting engine that guarantees a desired forecasting accuracy. Times series decomposition and Forecast combination have been two classes of methods that have attracted the interest of recent researchers. One solution has been to use decomposition followed by recombining to form a very large number of forecasting models Further many recent papers present Data Mining based methods to intelligently discover a subset of methods from a large list that can be used in combination for sales demand forecasting. These methods are computationally expensive and prohibitive if they are applied to each individual time series in a retail organization. In this paper we present a novel technique to identify similar sales series and efficiently use the best combination of methods learnt for one series to forecast for the entire set of similar series.