ST Stock Price Forecasting with the PSO-KAN-Transformer Hybrid Model

The rapid advancement of big data and Internet of Things (IoT) technologies has expanded the application scope of high-dimensional time series data. To address the limited interpretability of Multilayer Perceptron models in ST stock price prediction, this study enhances the internal structure of the Transformer by replacing the MLP layer with a KAN layer, thereby improving the interpretability of the Transformer in stock price forecasting. For experimental analysis, empirical research is conducted on a dataset including the SSE 50 Index, SZSE 100 Index, and two ST stocks. Experimental results show that the PSO-KAN-Transformer hybrid model achieves superior performance in time series prediction, reducing the mean square error by an average of 18.85% compared with the original Transformer. This indicates that the proposed model has improved accuracy and interpretability.