Stock Price Prediction Model Based on Feature Engineering and Optimized Transformer: An Empirical Study on Improving the Precision of Investment Decision-making through Time Series Analysis

Abstract

Stock price prediction has always been a topic of great concern in the financial field, and it has important practical significance for investment decision-making and market supervision.The CEEMDAN-Lasso-BO-Transformer model constructed in this paper obtains features through Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), screens key features using the Least absolute shrinkage and selection operator (Lasso), builds the model with Transformer, and tunes hyperparameters through Bayesian Optimization (BO), thus finally completing feature engineering and model optimization.Experiments show that the CEEMDAN-Lasso-BO-Transformer model constructed this time has a fitting degree as high as 98.7%, and its indicators such as MAE, MAPE, and RMSE are the best. It successfully predicts the trend of SSE stock prices from April to June 2025. Simulated trading shows that the Sharpe ratio of this model reaches 0.75, with the most stable profit, which can provide more accurate decision-making support for investors.