With the development of the social economy, research on stock market prediction is in full swing. However, the fluctuations in stock price and returns are influenced by many factors, including political policies, market environment, investor psychology, and so on. The traditional analysis method, based on subjective experience, requires significant time and effort, and its prediction accuracy is often poor. Now, the application of machine learning algorithms to predict stock returns has become a hot topic among scholars. This article comprehensively analyzes the advantages and disadvantages of support vector machine (SVM), tree-based algorithms, and neural network algorithms in processing tabular data and time series data. It proposes a hybrid model based on the deep neural network (DNN) and TabNet models, combining the strengths of the DNN and tree-based models. In the model training stage, two neural networks are established to accept the inputs of ID features and numerical features, respectively, and multiple fully connected layers are used to complete the construction of the DNN model. The TabNet is implemented based on the attention transformer and feature transformer, and the prediction results of the two models are fused. The proposed model has a best Pearson correlation coefficient (PCC) value and a lowest root mean square error (RMSE) value at the same time, because the hybrid algorithm performs particularly well on large data sets with the least feature engineering and has strong interpretability, such as quantifying the contribution of different features in the model, it has certain theoretical significance and wide application value.
Full Text Finder