Scholarlink

<p>Stock price forecasting, as a core topic in financial time series analysis, is crucial for enhancing the accuracy of quantitative investment decisions. To address key challenges including high noise levels in financial market data, the limitations of single models in capturing temporal features, and the reliance on empirical parameter tuning for hyperparameter optimization, this paper proposes a Feature-enhanced LSTM-Optuna-Augmented Model (F-LOAM). The core innovations of this study are as follows: (1) proposing an enhanced Variational Modal Decomposition (VMD) denoising algorithm, which introduces the Spearman Rank Correlation Coefficient (SRCC) to establish an adaptive modal selection mechanism, thereby improving the signal-to-noise ratio by 23.6%; (2) designing a cascaded Long Short-Term Memory (LSTM)-Light Gradient Boosting Machine (LightGBM) architecture that first extracts high-order feature representations from time series data using LSTM, then utilizes LightGBM’s strengths in large-scale data processing for precise predictions, while optimizing both models’ hyperparameters through the Optuna framework. In comprehensive evaluations on the self-built MFD12 dataset and public datasets including ACL18 and KDD17, F-LOAM achieves superior performance. Furthermore, the prediction results are converted into executable trading strategies, and backtesting on live markets yield an annualised return of 20.27% and a Sharpe ratio of 1.308, demonstrating outperforming existing methods and providing an innovative solution for intelligent investment decision-making.</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

F-LOAM: an efficient hybrid model for stock price prediction based on SVMD denoising

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月21日

查看全文
  • Liefa Liao,
  • Jingjing Lin

摘要

Stock price forecasting, as a core topic in financial time series analysis, is crucial for enhancing the accuracy of quantitative investment decisions. To address key challenges including high noise levels in financial market data, the limitations of single models in capturing temporal features, and the reliance on empirical parameter tuning for hyperparameter optimization, this paper proposes a Feature-enhanced LSTM-Optuna-Augmented Model (F-LOAM). The core innovations of this study are as follows: (1) proposing an enhanced Variational Modal Decomposition (VMD) denoising algorithm, which introduces the Spearman Rank Correlation Coefficient (SRCC) to establish an adaptive modal selection mechanism, thereby improving the signal-to-noise ratio by 23.6%; (2) designing a cascaded Long Short-Term Memory (LSTM)-Light Gradient Boosting Machine (LightGBM) architecture that first extracts high-order feature representations from time series data using LSTM, then utilizes LightGBM’s strengths in large-scale data processing for precise predictions, while optimizing both models’ hyperparameters through the Optuna framework. In comprehensive evaluations on the self-built MFD12 dataset and public datasets including ACL18 and KDD17, F-LOAM achieves superior performance. Furthermore, the prediction results are converted into executable trading strategies, and backtesting on live markets yield an annualised return of 20.27% and a Sharpe ratio of 1.308, demonstrating outperforming existing methods and providing an innovative solution for intelligent investment decision-making.