Forecasting Time Series from Clusters

Forecasting large numbers of time series is a costly and time-consuming exercise. Before forecasting a large number of series that are logically connected in some way, we can first cluster them into groups of similar series. In this paper we investigate forecasting the series in each cluster. Similar series are first grouped together using a clustering procedure that is based on a test of hypothesis. The series in each cluster are then pooled together and forecasts are obtained. Simulated results show that this procedure for forecasting similar series performs reasonably well.