Scholarlink

<p>Time series prediction is critical in various domains, yet challenges such as non-stationarity, high volatility, and abrupt fluctuations often degrade model performance. Advanced architectures that excel in capturing complex temporal dependencies are usually hindered by significant computational requirements and challenges of extensive fine-tuning, limiting their practicality in resource-constrained environments. This study proposes a Recursive EMD method as a solution to those challenges, which decomposes raw time series data into intrinsic mode functions, offering a simplified and structured representation of complex temporal patterns. The proposed approach enhances time series forecasting by iteratively decomposing rapidly fluctuating signals into structured subcomponents. Evaluations across diverse datasets demonstrate significant performance gains; for instance, the method facilitated reductions in Mean Absolute Error (MAE) of up to 58% for LSTM and 50% for GRU models on the Traffic dataset. Furthermore, it achieved a 40-65% improvement on highly volatile, non-stationary series such as GBPUSD. By refining the input signal, this method improves the predictive performance and robustness of streamlined sequence learners and lightweight transformer-based models, establishing the potential of decomposition-based techniques to enhance forecasting accuracy while avoiding the computational burden of more complex models.</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

Enhancing time series prediction accuracy with recursive EMD

  • 期刊论文

  • 发表日期：

    2026年4月27日

查看全文
  • Gökberk Açıkgöz,
  • Kerem Zengin,
  • Ümit Çavuş Büyükşahin,
  • Alya Su Bolelli,
  • Şeyda Ertekin

摘要

Time series prediction is critical in various domains, yet challenges such as non-stationarity, high volatility, and abrupt fluctuations often degrade model performance. Advanced architectures that excel in capturing complex temporal dependencies are usually hindered by significant computational requirements and challenges of extensive fine-tuning, limiting their practicality in resource-constrained environments. This study proposes a Recursive EMD method as a solution to those challenges, which decomposes raw time series data into intrinsic mode functions, offering a simplified and structured representation of complex temporal patterns. The proposed approach enhances time series forecasting by iteratively decomposing rapidly fluctuating signals into structured subcomponents. Evaluations across diverse datasets demonstrate significant performance gains; for instance, the method facilitated reductions in Mean Absolute Error (MAE) of up to 58% for LSTM and 50% for GRU models on the Traffic dataset. Furthermore, it achieved a 40-65% improvement on highly volatile, non-stationary series such as GBPUSD. By refining the input signal, this method improves the predictive performance and robustness of streamlined sequence learners and lightweight transformer-based models, establishing the potential of decomposition-based techniques to enhance forecasting accuracy while avoiding the computational burden of more complex models.