Research on Crude Oil Price Prediction Based on Improved Long Short-Term Memory Neural Network

Abstract

The prediction of financial time series has always been challenging. This study aims to improve the accuracy and robustness of crude oil price prediction by employing various analysis methods and models. Firstly, we introduce CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) and Hilbert transformation, using these methods for multiscale decomposition of crude oil price time series. The decomposed components are then forecasted using LSTM and LSTM models with attention mechanism, demonstrating superiority in multiscale features. In the reconstruction phase, we employ an intelligent reconstruction method, achieving more accurate restoration of crude oil price fluctuations compared to simple summation reconstruction. Overall, the composite model constructed in this study exhibits superior predictive performance across multiple metrics such as MSE, RMSE, MAE, and R². Compared to other traditional methods and single models, our model demonstrates stronger adaptability and predictive accuracy in complex market environments. This study introduces the ideas of multiscale features and intelligent reconstruction to the field of crude oil price prediction, providing a new approach to improving prediction accuracy. In the current uncertain and volatile crude oil market, our model not only has better fitting capabilities but also responds more flexibly to complex market scenarios, offering a reliable reference for relevant decision-making.