Deep Learning techniques for stock market forecasting: Recent trends and challenges

Abstract

Stock market forecasting has been a very intensive area of research in recent years due to the highly uncertain and volatile nature of stock data which makes this task challenging. By accurately predicting a particular stock's price investors can gain maximum profit out of their investment. With the great success of Deep Learning methods in various domains, it has attracted the research community to apply these models for financial domain also. These DL methods have been proven to achieve better accuracy and predictions compared to econometric and traditional ML methods. This work reviews recent papers according to various Deep Learning models which included: Artificial Neural Networks, Convolution Neural Networks, Sequence to Sequence models, Generative Adversarial Networks, Graph Neural Networks and Transformers applied for stock market forecasting. Furthermore this work also reviews datasets, features, evaluation parameters and results of various methods. From the analysis done on various DL models we found that Graph Neural Networks and Transformer models have potential to interpret dynamic and non-linear patterns of financial time series data with greater accuracy. In addition to this, correlation among various stock indices and investors sentiment along with historical data has great influence on the prediction accuracy. We also identified the benchmark datasets for stock market forecasting based on market capitalization value of an economy. The aim of this paper is to provide insight into most recent work done in the finance domain and identify future directions for more accurate predictions.