Multi-feature stock price prediction by LSTM networks based on VMD and TMFG

Abstract The stock market is characterized by its high nonlinearity and complexity, making traditional methods ineffective in capturing its nonlinear features and complex market dynamics. This paper proposes a novel stock price forecasting model—the Variational Mode Decomposition—Triangulated Maximally Filtered Graph—Long Short-Term Memory (VMD–TMFG–LSTM) combined model—aimed at improving prediction accuracy, stability, and computational efficiency. The proposed model first employs Variational Mode Decomposition (VMD) to decompose the stock price time series into multiple smooth intrinsic mode functions (IMFs), reducing data complexity and mitigating noise interference. Subsequently, the TMFG algorithm is utilized for feature selection, simplifying the input data and accelerating the iterative convergence process. Finally, the filtered features are modeled and predicted using a Long Short-Term Memory (LSTM) network. Experimental results demonstrate that the VMD–TMFG–LSTM model significantly outperforms AutoRegressive Integrated Moving Average (ARIMA), Neural Network (NN), Deep Neural Network (DNN), Convolutional Neural Network (CNN), as well as single LSTM, TMFG–LSTM, and VMD–LSTM models in forecasting the closing prices of multiple stocks. Specifically, for Shanghai International Airport Co., Ltd. (sh600009), the VMD–TMFG–LSTM model achieves a 69.76% reduction in Root Mean Squared Error (RMSE), a 71.41% reduction in Mean Absolute Error (MAE), a 46.28% reduction in runtime, and an improvement of 0.2184 in R-squared (R 2 ), indicating significantly higher prediction accuracy. In conclusion, the combined model proposed in this paper enhances the accuracy, efficiency, and stability of stock price prediction, providing a robust and efficient solution for forecasting stock market trends.