Stock Market Forecasting Based on Spatiotemporal Deep Learning.

Area of Science:

• Financial forecasting
• Machine learning for finance
• Time series analysis

Background:

• Traditional models like LSTM and Transformers struggle to incorporate spatial information in stock price prediction.
• Accurate stock price forecasting is crucial for investment strategies and market analysis.

Purpose of the Study:

• To introduce the Spacetimeformer model, a novel approach for stock price prediction.
• To evaluate the effectiveness of incorporating a time-space mechanism in financial forecasting.
• To compare Spacetimeformer's performance against existing LSTM and Transformer models.

Main Methods:

• Developed the Spacetimeformer model, integrating a time-space mechanism into the Transformer architecture.
• Utilized ten-minute stock price data for Taiwan 50 Index constituents and intraday data from the Taiwan Stock Exchange.
• Trained the model using multi-time-step stock price data and daily moving windows.

Main Results:

• The Spacetimeformer model demonstrated superior performance in stock price prediction compared to LSTM and Transformer models.
• The model successfully captured essential stock price trend changes and provided stable predictions.
• The time-space mechanism was shown to be significant and valuable for enhancing prediction accuracy.

Conclusions:

• The Spacetimeformer model offers a significant advancement in stock price forecasting by effectively integrating spatial and temporal data.
• The proposed method provides a valuable tool for investors seeking to improve their return on investment.
• The study highlights the importance of the space-time mechanism in developing more accurate predictive models for financial markets.