Beyond Euclidean: a Metric-optimized Type-1 Fuzzy Svr Functions Architecture for Time Series Forecasting
Year: 2026
Type: article
Abstract: This paper introduces a distance-aware forecasting framework based on type-1 fuzzy support vector regression functions, designed to examine the impact of distance metric selection on fuzzy time series prediction performance. The proposed approach extends fuzzification mechanisms by incorporating nine alternative distance functions into a fuzzy clustering-driven regression structure. These distance measures determine the formation of fuzzy partitions and directly influence the weighting scheme used in cluster-specific regression models. A unified optimization strategy is employed to jointly tune both fuzzification and regression components using a grid search procedure. This integrated optimization allows the framework to dynamically adapt to varying data distributions, temporal dependencies, and structural characteristics of time series data. The proposed method is validated through extensive experiments conducted on four real-world datasets from production and financial domains. The empirical findings reveal that distance metric choice significantly affects forecasting accuracy. In many experimental scenarios, non-Euclidean distance measures outperform classical Euclidean-based metrics. Distance functions emphasizing directional similarity, scale normalization, or maximum deviation frequently produce superior results. Chebyshev, standardized Euclidean, and Cosine distances achieve the lowest prediction errors across several datasets and evaluation periods. The results indicate that no single distance metric is universally optimal, underscoring the importance of metric flexibility and data-aware selection. The study demonstrates that integrating distance metric optimization into type-1 fuzzy support vector regression function-based architectures leads to consistent improvements in forecasting accuracy. The proposed architecture offers a generalizable solution for time series forecasting.
Authors: İbrahim Cimşit, Ali Zafer Dalar, Ufuk Yolcu
Institutions: Giresun University, Marmara University
Language: Turkish
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7155690277