Forecasting Sales Through Time Series Clustering

Abstract

We study the effect of decomposing time series into multiple components like trend, seasonal and irreg ular and performing the clustering on those components and generating the forecast values of each component separately. In this project we are working on sales data. Multiple forecast experts are used to foreca st each component series. Statistical method ARIMA, Holt winter and exponential smoothing are used to forecast these components. We performed clustering for forec asting and discovered a set of best, good and bad forecasters. Selection of best, good and bad forecasters is performed on the basis of count and rank o f expert id’s generated. Since we have thousands of e xperts, we experiment with combining method to get better forecast. Finally absolute percentage error (APE) is used for comparing forecast.