Quantitative models are one of the core challenges for investors in stock dynamic prediction. The original LSTM(long short-term memory) stock prediction model was affected by noise in the input data, which interfered with the prediction effect. In this paper, there are 259 indicators that affect stock prices. Firstly, the input data was reduced in dimensionality using dimensionality reduction methods to preserve key information, and then input into LSTM to form an improved prediction model, namely PCA-LSTM model, ISOMAP-LSTM model, and PCA-ISOMAP-LSTM model. Through empirical comparison, compared with the original LSTM prediction model and the attention mechanism model MHA-LSTM, the PCA-LSTM model and ISOMAP-LSTM model reduce training time. The average absolute error (MAE), average relative error (MAPE), and root mean square error (RMSE) in the prediction error evaluation indicators are significantly reduced, and the average rise and fall accuracy (ARRF) is significantly improved. However, the PCA-ISOMAP-LSTM model has an increase in error rate and a certain decrease in accuracy. The Diebold Mariano test also showed that the PCA-LSTM model and ISOMAP-LSTM model have stronger stock prediction abilities than the original LSTM model and MHA-LSTM model, while the PCA-ISOMAP-LSTM model and MHA-LSTM model have weaker prediction abilities than the original LSTM model. The difference in prediction accuracy between the PCA-LSTM and ISOMAP-LSTM models is not significant, and both can be used as a new technical support for quantitative stock investment.

dimensionality reduction  /  principal component analysis  /  ISOMAP(isometric mapping)  /  LSTM(long short-term memory)  /  stock prediction  /  attention mechanism