Prediction and Analysis of the Performance of CSI 300 Index Constituent Stocks Based on Integrated Algorithms

This paper aims to explore the effectiveness of stock price prediction models based on ensemble algorithms by predicting the trends of the CSI 300 Index components. The study utilizes 20 feature variables from 242 trading days of the CSI 300 components in 2024, totaling 72,504 data points. Regression interpolation, DBSCAN and wavelet thresholding methods are employed to enhance data stability and quality. During the model construction phase, this study transforms the stock price prediction problem into a binary classification problem, using the direction of price changes in the denoised closing prices as classification labels. Two ensemble learning algorithmsRandom Forest and XGBoostare employed for modeling. Random Forest achieves predictions by constructing multiple decision tree models, with parameters set to n_estimators=500 and max_features=4. XGBoost further optimizes the boosting tree model and determines the optimal parameters through grid search. Both models use a 2:1 ratio to divide the training set and validation set, respectively, for model training and evaluation. XGBoost achieves a prediction accuracy of 55.76%, slightly higher than Random Forest's 55.46%.