A Review of Stock Forecasting Methods Based on Improved LSTM Series Models
TL;DR
This paper systematically reviews the research process for stock prediction based on improved LSTMs, elaborates on the different improved LSTMs and their drawbacks, and summarizes the principles, advantages, and disadvantages of the four models: PCA-LSTM, CNN-LSTM-Attention, BERT-Bi-LSTM, and Event LSTM-Attention.
Abstract
The trend of stock prices not only immediately affects policymakers' decisions and income, but also reflects macroeconomic conditions and market status. In the Chinese capital market, scientific and reliable methods for stock price prediction are highly relevant to practice. Financial time series exhibit nonlinearity, high noise, and strong randomness. And it is affected by sentiment and emergency events—the traditional and fundamental models. LSTMs have not met the demand for predictive performance. This paper systematically reviews the research process for stock prediction based on improved LSTMs, elaborates on the different improved LSTMs and their drawbacks, and summarizes the principles, advantages, and disadvantages of the four models: PCA-LSTM, CNN-LSTM-Attention, BERT-Bi-LSTM, and Event LSTM-Attention. This paper compares the distinctions among this model's Data Adaptation, Feature Mining, and Market Response, and outlines future development directions for Multimodal Fusion, Interpretability Enhancement, and Extreme Market Condition Modeling. This paper provides theoretical foundations and methodological support for model selection and innovation research in stock price prediction.