Research Information System
Engineering Applications of Artificial Intelligence, vol.148, 2025 (SCI-Expanded)
Time series forecasting has progressed from relying solely on traditional linear or non-linear methods to hybrid approaches that integrate complementary strengths. Traditional models often excel in specific scenarios but struggle with the complexities of real-world data, where linear trends and non-linear dynamics coexist. To address these limitations, this study introduces hybrid forecasting methods that combine autoregressive techniques with type-1 fuzzy functions. A second version of the model integrates a subsampling bootstrap approach, enabling statistical inference and confidence interval estimation within a fuzzy framework. The proposed methods were tested on datasets from four international stock exchanges, each encompassing five distinct years (2010–2014). Forecasting was conducted across three different test set sizes (10, 20, and 40 observations), resulting in a total of 60 experimental scenarios. The hybrid models consistently outperformed benchmark methods in 75% of cases, demonstrating their ability to capture both linear trends and non-linear complexities effectively. The bootstrap-enhanced version further improved reliability, delivering sharper probabilistic forecasts and robust confidence intervals, particularly in volatile financial environments. To ensure precise and adaptive parameter tuning, artificial intelligence-based optimization techniques, such as particle swarm optimization, were employed. This integration of artificial intelligence not only enhanced forecasting accuracy but also demonstrated the transformative potential of hybrid models for real-world applications. These findings provide a valuable tool for decision-making in dynamic and uncertain domains, such as financial market analysis.