An optimized LSTM network for improving arbitrage spread forecasting using ant colony cross-searching in the K-fold...

Area of Science:

• Quantitative Finance
• Machine Learning in Trading

Background:

• Arbitrage spread prediction is vital for algorithmic trading, but complex models often yield unstable returns.
• Current methods struggle with precision and efficient model development for arbitrage signal identification.

Purpose of the Study:

• To develop an optimized Long Short-Term Memory (LSTM) network for accurate and stable arbitrage spread prediction.
• To enhance the efficiency and adaptability of model development in financial forecasting.

Main Methods:

• Introduction of a K-fold cross-search (KCS) heuristic algorithm, enhancing ant colony optimization with iterative search space updates.
• Optimization of LSTM hyperparameters using a modified fitness function for automated data set adaptation.
• Validation of the KCS-LSTM network on real-world rebar and hot-rolled coil spread data.

Main Results:

• The KCS-LSTM network demonstrated superior performance compared to common models, with significant improvements in symmetric mean absolute percentage error (sMAPE).
• Performance gains ranged from 12.6% to 72.4% in prediction accuracy.
• The model proved competitive against more complex neural network architectures.

Conclusions:

• The KCS-LSTM network offers a robust and efficient solution for arbitrage spread prediction in algorithmic trading.
• The proposed method enhances prediction accuracy and return stability, simplifying model development.
• This approach presents a competitive alternative to existing complex models for financial market forecasting.