An improved chaotic whale optimization algorithm for MHA-MLP stock trend forecasting

Stock markets, characterized by high noise, non-stationarity, and stochastic fluctuations, pose significant challenges for accurate trend prediction. While deep learning models show promise, their training is often inefficient due to tedious parameter optimization. This paper proposes a novel hybrid prediction model that synergistically integrates an improved chaotic whale optimization algorithm with a multi-head attention mechanism and a multilayer perceptron neural network. The overall architecture consists of three core components: data preprocessing, an attention-enhanced multilayer perceptron module, and an optimization module based on the improved whale optimization algorithm. The attention-enhanced multi-layer perceptron employs residual connections between its attention and hidden layers to mitigate the vanishing gradient problem and effectively capture global temporal dependencies in the input features. The chaotic whale optimization module innovatively utilizes T-distribution wavelet mutation and polynomial differential evolution strategies to automatically optimize the model's architecture, specifically the number of attention heads and hidden layers. Extensive experiments on A-share and U.S. stock market datasets validate the model's efficacy. The results demonstrate that the chaotic whale optimization algorithm not only reduces the number of hyperparameter optimization iterations by 32 % but also improves fitness value convergence accuracy by 19 %, offering a robust and efficient solution for modeling complex financial data.