Are Negative Weights in Combining Forecasts So Bad?

Abstract

Purpose of the study. In this paper, we consider the problem of negativity of weight coefficients when combining forecasts. Combining forecasts as a method has long ago proved itself in practice as a good way to improve forecast accuracy. However, in the literature little attention is paid to the issue of negative weights during aggregation, although the cases of obtaining such weights in practice are quite common. The reasons why this may happen are not considered or analyzed. Often, when obtaining weights less than zero, such weights are reset to zero, thus excluding the information contained in the particular forecasting method from the combination, which may reduce the accuracy of the combined forecast. In this regard, it is important to understand why when combining forecasts, negative weight can be obtained and determine options for how to avoid such situations in combining without losing accuracy.

Materials and methods. It is proposed to consider various approaches to eliminate excluded weights when combining forecasts, including truncation of weight coefficients or imposing restrictions on them, including the option of sequential combining of forecasts.

Results. The result is a list of reasons why negative weights can be obtained when combining forecasts, what risks they have and how to avoid them.

Conclusion. Based on the results obtained, it can be concluded that the negative weights themselves when combining forecasts can be triggers for identifying problems when combining. However, it is dangerous to retain them, as they can lead to uncertain prediction results and degrade the accuracy of the resulting combined forecast. The proposed methods of work allow you to bypass the negativity of the weights without a strong deterioration in forecasting.