A new approach to time series modeling with fuzzy measures and the Choquet integral

In this paper, we propose a new approach to modeling of time series using nonmonotonic fuzzy measures and the Choquet integral, which removes the assumption of linear independency between time series data. As an example, we propose a subset interactive AR model. The modeling consists of a structure identification problem and a parameter identification problem. For structure identification which involves a large number of possible alternative models, we develop a method that uses genetic algorithms to select a subset of regression variables. We apply this model to well-known time series data (i.e., the sunspot numbers), and analyze the quality of the forecast. The results show that the proposed approach is useful for time series analysis with nonlinear features.<
>