Study of Mainstream Methods in Stock Price Prediction: Performance Characteristics, Comparative Analysis and Fusion Innovation Directions
TL;DR
A primary insight is that future research should move towards integrating these complementary methodologies and leveraging multi-modal data sources, which holds the potential for developing more robust and economically interpretable predictive frameworks.
Abstract
Stock price prediction is a typical but difficult problem in quantitative finance that can help investors make decisions and manage risks. However, due to the high volatility of the financial market and other reasons, it cannot be known in advance. The three main methodological paradigms in the current studies are introduced systematically here. The first is feature selection and extraction to improve the quality of the model's input data from raw data. The second type is a neural network-based approach; RNNs, LSTMs and Transformers are used to learn complex, non-linear temporal dependencies in the historical data. The third kind is a graph-structure-based method that builds a model of the connections among all stocks to study cross-asset spillover effects. The review shows that each method has distinct strengths and is suited to different market conditions, with no single approach dominating all scenarios. A primary insight is that future research should move towards integrating these complementary methodologies and leveraging multi-modal data sources, which holds the potential for developing more robust and economically interpretable predictive frameworks.
