Combining earnings forecasts using multiple objective linear programming

Abstract

In this study, exponential smoothing, univariate time series and (transfer function) bivariate time series models are combined to forecast annual corporate earnings for six major corporations. Consideration is given to combining forecasts generated by the same technique at different points in time as well as those generated by different techniques. Multiple objectives are incorporated into the forecasting process. Mathematical programming is utilized to generate combined forecasts that are efficient with respect to multiple objectives. Results indicate that combined forecasts outperform individual forecasts, that all three major categories of forecasting techniques are utilized in the construction of the efficient combined forecasts, that the techniques included in the combined forecasts and their relative weights can change over time and that the most recent forecasts do not always receive the most weight when combined with older forecasts.