Comparative Analysis of Mainstream Stock Prediction Models

Abstract

In response to the challenges in stock price prediction caused by the high noise and non-stationary characteristics of the financial market, this paper systematically studies the technical evolution path of deep learning hybrid models. By deconstructing the collaborative mechanisms of four architectures: Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN)-LSTM, Gate Recurrent Unit (GRU), and Temporal Convolutional Network (TCN)-Transformer, and using various evaluation conditions such as Mean Absolute Error and Root Mean Square Error to analyze the advantages of the models in stock prediction, the research further reveals the insufficiency of cross-market adaptability and the bottleneck of extreme event prediction. It proposes a phased technical path - in the near term, build a federated learning data framework to improve the utilization rate of small samples, and in the medium and long term, integrate causal inference and quantum acceleration mechanisms to drive the prediction model towards an industrial-level decision-making system. The results show that the Mean absolute error (MAE) index of TRAformer is 66.8% lower than that of the ordinary LSTM model, and the composite model supported by the ARO algorithm also performs better than the ordinary LSTM model.