Multi-strategy modified sparrow search algorithm for hyperparameter optimization in arbitrage prediction models.

Area of Science:

• Financial Engineering
• Computational Intelligence
• Machine Learning

Background:

• Deep learning models face challenges in predicting arbitrage spreads due to data's non-linear characteristics.
• Optimizing network structure and hyperparameters is crucial for improving model performance.
• Swarm intelligence algorithms offer effective solutions for complex optimization problems in financial modeling.

Purpose of the Study:

• To develop an advanced arbitrage spread prediction model by integrating a modified swarm intelligence algorithm with a deep learning network.
• To enhance the spatial exploration capabilities of the Sparrow Search Algorithm (SSA) for better optimization.
• To evaluate the proposed model's effectiveness using real-world financial futures data.

Main Methods:

• Implementation of the Multi-Strategy Modified Sparrow Search Algorithm (MSMSSA) to optimize the Long Short-Term Memory (LSTM) network.
• Incorporation of good point set theory, proportion-adaptive strategy, and improved location updates within the MSMSSA.
• Validation of the MSMSSA-LSTM model using rebar and hot coil futures spread data from the Chinese futures market.

Main Results:

• The MSMSSA-LSTM model demonstrated substantial reductions in prediction errors.
• Mean Absolute Percentage Error (MAPE) decreased by up to 58.5%.
• Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) decreased by up to 65.2% and 67.6%, respectively, compared to classical models.

Conclusions:

• The MSMSSA-LSTM model achieves high accuracy in predicting arbitrage spreads.
• The enhanced optimization strategy significantly boosts the performance of LSTM networks for financial forecasting.
• The model provides a valuable tool for investors in the futures market.