A Multi-Market Data-Driven Stock Price Prediction System: Model Optimization and Empirical Study

Abstract

As global financial markets become increasingly interconnected, relying solely on historical data from a single market is insufficient for achieving high prediction accuracy in stock price forecasting. It is essential to consider the impact of cross-market transactions on China’s stock market fluctuations. This study investigates this impact by constructing a comprehensive dataset that includes stock market, bond market, and foreign exchange market data from multiple countries. A deep learning algorithm, combined with the grid search method, was employed to compare different structural hyperparameters of four models: LSTM, GRU, MLP, and Transformer. The results indicate that the GRU model outperformed the others in terms of prediction accuracy. Consequently, a GRU-Arbitrage trading system based on optimal structural hyperparameters was developed. This system effectively integrates international and domestic information for stock price prediction, providing investors with a practical and accurate tool. The research findings show that the mean squared error (MSE) of stock prices predicted using international market transaction information for the Chinese stock market is 0.0069, followed by 0.0035 for the bond market and 0.0102 for the foreign exchange market. These results verify the significant impact of cross-market data on predicting China’s stock market fluctuations.

Keywords

• Stock index prediction
• cross-market
• hyperparameter tuning
• GRU trading system