Research and Application of LSTM Neural Network Algorithm

We developed a predictive model using Long Short-Term Memory (LSTM) neural networks aimed at forecasting a specific benchmark index. The focus was on refining constituent data based on predefined criteria and normalizing and scaling daily return factor data. Multi-indicator data was processed through Principal Component Analysis (PCA) and TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) methods to identify critical components affecting predictions. The suitability of the data for analysis was confirmed through a Kaiser-Meyer-Olkin (KMO) measure of 0.714 and Bartlett's test of sphericity with a p-value less than 0.05. This approach offers insights into improving forecast accuracy and handling dynamic changes in the underlying dataset.