Dynamically forming a group of human forecasters and machine forecaster for forecasting economic indicators

Abstract

How can human forecasts and a machine forecast be combined in inflation forecast tasks? A machine-learning-based forecaster makes a forecast based on a statistical model constructed from past time-series data, while humans take varied information such as economic policies into account. Combination methods for different forecasts have been studied such as ensemble and consensus methods. These methods, however, always use the same manner of combination regardless of the situation (input), which makes it difficult to use the advantages of different types of forecasters. To overcome this drawback, we propose an ensemble method for estimating the expected error of a machine forecast and dynamically determining the optimal number of humans included in the ensemble. We evaluated the proposed method by using the seven datasets on U.S. inflation and confirmed that it attained the highest forecast accuracy for four datasets and the same accuracy as the highest one of traditional methods for two datasets.