A Study of High Frequency Stock Price Fluctuation Prediction by Deep Learning Methods Based on Multivariate LSTM

The study of stock return forecasting has been a major challenge for both academics and practitioners, which is disturbed by various factors. Traditional statistical analysis methods are far from being able to capture the important information hidden in stock price fluctuations. In recent years, as deep learning algorithms have been widely used in the field of time series forecasting, more and more scholars have tried to use deep learning algorithms to improve the forecasting accuracy of stock returns. Compared with traditional econometric statistical models, deep learning algorithms have shown great advantages, specifically in that deep learning algorithms can extract volatility series features that are more abstract and have stronger predictive power without doing feature engineering. In this paper, we propose a multivariate LSTM composite model based on LSTM and GARCH class model, combining the advantages of deep learning models and traditional statistical analysis models, in order to improve the accuracy of stock volatility prediction in the context of high frequency data. In this paper, we select intra-day minute series data from 2020–2022 and compare and analyze the forecasting results of hybrid LSTM and GARCH class model, single GARCH class model, single variable LSTM model, and multivariate LSTM model. The experimental results show that the hybrid LSTM and GARCH class model is better than the multivariate LSTM model. The multivariate LSTM model is better than the single LSTM model and the single GARCH class model. In addition, the HG-LSTM model has the highest prediction accuracy.