A layered stock prediction model based on novel feature selection and model parameter optimization

As one of the core subjects in the financial field, stock market forecasting has been attracting extensive attention from the academic and industry, and has promoted the development of many complex forecasting models. However, due to the inherent complexity and volatility of the stock market, existing models still have significant limitations in terms of forecasting accuracy and stability. For example, overly verbose features can lead to data redundancy; Inadequate optimization of model parameters may prevent the full capture of market dynamics, thus limiting the further improvement of forecasting performance. To address these challenges, this paper proposes a bidirectional gated cyclic unit layering model (HFSLS-PSO-BIGRU) based on local shuffling technology and particle swarm optimization. The innovative framework effectively perturbs the data set through a local shuffling method to accurately assess feature importance and achieve optimal feature subset selection. This strategy significantly reduces data redundancy and enhances the generalization ability of the model. In addition, by using PSO in BIGRU architecture for global parameter tuning, the optimal parameter configuration is ensured, which significantly improves the prediction effect. At the same time, the reweighting mechanism and hierarchical loss calculation strategy are implemented in the training process, so that we can optimize the overall model performance. Based on the data of Shanghai Stock Exchange Index and four individual stocks, the empirical study shows that the HFSLS-PSO-BIGRU model proposed in this paper shows significant performance advantages compared with six benchmark models in MSE, RMSE, MAE and \(R^{2}\) evaluation indicators. In addition, the importance and necessity of each component within the framework is further validated through systematic ablation experiments, providing an innovative solution for stock market forecasting.