Research on stock prediction based on CED-PSO-StockNet time series model

Area of Science:

• Quantitative Finance
• Machine Learning
• Time Series Analysis

Background:

• Stock market prediction faces challenges due to high-noise environments, leading to low accuracy.
• Existing models struggle to effectively handle noise and extract relevant features from complex time series data.

Purpose of the Study:

• To introduce an innovative time series model, CED-PSO-StockNet, for enhanced stock prediction accuracy.
• To address the limitations of existing models in noisy financial data environments.

Main Methods:

• Data decomposition using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and frequency estimation via the extreme point method.
• An Encoder-Decoder framework with an attention mechanism for predicting reconstructed components.
• Parameter optimization using an Improved Particle Swarm Optimization (IPSO) algorithm.

Main Results:

• The CED-PSO-StockNet model demonstrated significant improvements in stock prediction accuracy.
• Achieved a 45.59% improvement in R² metric compared to the standalone LSTM model on the Pudong Bank dataset.
• Validation on the Ping An Bank dataset confirmed the model's generalization capability and significant advantages.

Conclusions:

• The CED-PSO-StockNet model effectively enhances stock prediction accuracy in high-noise environments.
• The combination of CEEMDAN, an attention-based Encoder-Decoder, and IPSO optimization offers a robust solution for financial time series analysis.
• The proposed model shows significant potential for practical applications in stock market forecasting.