Avoiding Time Series Prediction Disbelief with Ensemble Classifiers in Multi-class Problem Spaces

Time series data is everywhere: it comes e.g. from IoT devices, financial transactions as well as medical and scientific observations. Time series analysis provides powerful tools and methodologies for modeling many kinds of related processes. Predictions based on such models often are of great value for many applications. But even the most accurate prediction will be useless if potential users will not want to accept and further use it. The article presents the problem of prediction disbelief and its relation with acceptance tests of predictions during lifecycle of time series analysis. The main contribution of the paper is classification and modeling of possible types of organization of acceptance tests of the outcomes of forecasting tools. This is done in the form of ensembles of classifiers working contextually in multi-class problem spaces. This allows to formulate, analyze and select the best methods of avoiding influence of prediction disbelief problem during time series analysis lifecycle.