Stock index trend prediction based on TabNet feature selection and long short-term memory.

Area of Science:

• * Computational Finance
• * Machine Learning in Finance
• * Quantitative Trading

Background:

• * Stock market prediction is complex due to numerous interrelated factors affecting individual stocks.
• * Existing models often lack comprehensive factor libraries and efficient feature selection.
• * Accurate stock index trend prediction requires robust methodologies and extensive data.

Purpose of the Study:

• * To propose a novel predictive model, TabLSTM, for stock index trend forecasting.
• * To develop a comprehensive factor database incorporating macro, micro, and technical indicators.
• * To demonstrate the efficacy of combining advanced machine learning techniques for improved prediction accuracy.

Main Methods:

• * Construction of a factor database including macro, micro, and technical indicators.
• * Utilization of TabNet for calculating factor importance and performing feature selection.
• * Application of Long Short-Term Memory Neural Network (LSTM) for trend prediction using selected optimal factors.

Main Results:

• * The TabLSTM model, integrating TabNet for feature selection and LSTM for prediction, demonstrated superior performance compared to existing methods.
• * Empirical validation confirmed the necessity and effectiveness of constructing a comprehensive factor database.
• * The model successfully predicted stock index trends across different financial markets.

Conclusions:

• * The TabLSTM model offers a feasible and effective approach for stock index trend prediction.
• * The study provides a complete factor database and a comprehensive architecture valuable for quantitative investments.
• * The findings offer significant references for future research in stock forecasting and algorithmic trading.