Improving stock price forecasting based on recurrent neural networks

Abstract

The article is devoted to the study of the application of machine learning for forecasting the stock price in the stock markets in the context of the growing capabilities of artificial intelligence. The article examines the issues and relevance of stock price forecasting, as well as describes current approaches to such forecasting as fundamental and technical analysis. Within technical analysis, widely used methods and algorithms of statistics, pattern recognition, sentiment analysis and machine learning are considered. The article focuses on the advantages of using recurrent neural networks as one of the types of artificial neural networks that work with sequential data or time series. Traditional neural networks treat data inputs and outputs as independent of each other, while recurrent neural networks take into account previous elements in the sequence to generate the output, which allows for better prediction of price trends. The article describes the main architectures of recurrent neural networks that exist at the moment and their features, including bidirectional recurrent neural networks, closed recurrent block, and long short-term memory. The last architecture - long short-term memory - was considered in the most detail. This architecture is one of the most difficult to learn and configure, at the same time, it is the most effective for complex tasks such as forecasting the dynamic course of shares. One of its key advantages is solving the problem of gradient damping, that is, the ability to store the context within which price data is analyzed, which makes it possible to detect long-term historical trends and predict future values with high accuracy.