Research And Analysis of Neural Networks for Financial Time Series Prediction

Abstract

Financial time series prediction is crucial for portfolio optimization, risk management, and financial policy-making. However, traditional technical analysis and statistical models, limited by linear assumptions and data distribution, struggle to handle the market’s nonlinear, high-volatility traits. Neural networks, with their nonlinear mapping and autonomous learning, offer a key solution. This paper reviews the field’s research over the past five years, focusing on four mainstream technologies: Recurrent Neural Network, functional denoising autoencoder, Long Short-Term Memory Neural Network, and hypergraph neural networks. It analyzes each model’s principles, features, pros and cons, and verifies their effectiveness using indicators like IC, Sharpe Ratio, and RMSE on datasets from A-shares, U.S. stocks, and cryptocurrencies. Models integrating spatiotemporal features (e.g., hypergraph neural networks) perform exceptionally well. The paper also notes current challenges, such as single data usage and weak cross-market generalization. Future optimization should involve multi-source data fusion and cross-market transfer learning, providing references for financial practice and advancing neural networks’ practical application.