Building an Ensemble of Time Series Models Using Empirical Risk Space

This paper presents a novel approach to building ensembles of time series models within the framework of empirical risk space. By conducting an in-depth analysis of the errors made by individual experts, particularly in relation to specific features of the data, the proposed method effectively optimizes expert weights through a sophisticated aggregation algorithm. This approach not only incorporates the concept of expert specialization but also meticulously considers the feature- specific manifestations of errors to accurately identify and exclude experts exhibiting consistent erroneous behavior. Experimental results demonstrate significant improvements in prediction accuracy when compared to traditional ensemble methods. These findings contribute to the advancement of ensemble modeling techniques and underscore the critical importance of feature-specific error analysis in the construction of robust time series ensembles.