An Evaluation of Methods for Combining Univariate Time Series Forecasts

Abstract: This thesis presents and evaluates nineteen methods for combining up to eleven automated univariate forecasts. The evaluation is made by applying the methods on a dataset containing more than 1000 monthly time series. The accuracy of one period ahead forecasts is analyzed. Almost 3.2 million forecasts are evaluated in the study. Methods that are using past forecasts to optimally produce a combined forecast are included, along with methods that do not require this information. A pre-screening procedure to get rid of the poorest performing forecasting methods before the remaining ones are combined is evaluated. The results confirm that it is possible to achieve a superior forecast accuracy by combining forecasts. The best methods that utilize past forecasts tend to outperform the best methods that are not considering this data. Including a pre-screening procedure to remove inferior forecasts before combining forecasts from the top five ranked methods seems to increase the forecast accuracy. The pre-screening procedure consists of ranking the automated univariate forecasting methods using an independent, but relevant, dataset. The four best performing methods utilize the pre-screening procedure together with past forecasts to optimally combine forecasts. The best method computes the historical mean squared error of each individual method and weights them accordingly. Demand for automated procedures is growing as the size of datasets increases within organizations. Forecasting from a large set of time series is an activity that can take advantage of automated procedures. However, choosing which forecasting method to use is often problematic. One way of solving this is by combining multiple forecasts into a single forecast.