Robust time series forecasting using a novel fuzzy regression approach based on kernel functions

In recent years, the use of fuzzy regression approaches in time series forecasting has increased notably. However, the influence of outliers in time series persists as a significant challenge. In this study, we propose a novel robust fuzzy regression functions approach, which can effectively address the issue of outliers. The proposed method incorporates robust techniques at both the clustering and inference stages. In particular, the fuzzy c-medoids clustering algorithm is employed in the initial stage, while a robust estimator based on kernel functions is utilised in the latter stage. To assess the forecasting performance of the proposed method, two financial time series datasets are considered, including Shanghai Stock Exchange Composite index time series and Taiwan Stock Exchange time series. Furthermore, to evaluate the robustness of the proposed method against outliers, four scenarios of contaminated data are examined. The experimental results demonstrate that the proposed method outperforms several popular methods in the majority of cases for both the original and contaminated datasets.