Scholarlink

<p>Crude oil prices are affected by supply-demand conditions, financial markets, and macroeconomic shocks, making interval forecasting more informative than point forecasting under uncertainty. To improve the robustness of interval crude oil price forecasting, this study proposes a hybrid multi-factor and multi-scale framework for West Texas Intermediate (WTI). First, improved interval grey relational analysis (IGRA) is used to select key factors. Second, bidirectional empirical mode decomposition (BEMD) and sample entropy reconstruction are employed to extract high-frequency, medium-frequency, low-frequency, and trend components. Third, an interval k-sigma rule combined with K-nearest neighbors (KNN) smoothing is applied to detect and repair abnormal observations. Finally, Transformer models are used for higher-frequency components, while a back-propagation neural network (BPNN) is adopted for the trend component, and all component forecasts are aggregated into the final interval prediction. Empirical results show that the proposed framework outperforms single-model, single-factor, and non-decomposition benchmarks in terms of IMAPE, IRMSE, IARV, UI and CR. Additional robustness tests further confirm its stability and effectiveness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid Multi-factor and Multi-scale Framework for Interval Crude Oil Price Forecasting With Robust Feature Selection, Decomposition-Reconstruction and Outlier Detection Strategy

  • 期刊论文

  • 发表日期：

    2026年7月31日

查看全文
  • Yi Xiao,
  • Lijie Han,
  • Chen He,
  • Yi Hu,
  • Ming Yi

摘要

Crude oil prices are affected by supply-demand conditions, financial markets, and macroeconomic shocks, making interval forecasting more informative than point forecasting under uncertainty. To improve the robustness of interval crude oil price forecasting, this study proposes a hybrid multi-factor and multi-scale framework for West Texas Intermediate (WTI). First, improved interval grey relational analysis (IGRA) is used to select key factors. Second, bidirectional empirical mode decomposition (BEMD) and sample entropy reconstruction are employed to extract high-frequency, medium-frequency, low-frequency, and trend components. Third, an interval k-sigma rule combined with K-nearest neighbors (KNN) smoothing is applied to detect and repair abnormal observations. Finally, Transformer models are used for higher-frequency components, while a back-propagation neural network (BPNN) is adopted for the trend component, and all component forecasts are aggregated into the final interval prediction. Empirical results show that the proposed framework outperforms single-model, single-factor, and non-decomposition benchmarks in terms of IMAPE, IRMSE, IARV, UI and CR. Additional robustness tests further confirm its stability and effectiveness.