Targeted Sentiments and Hierarchical Events based Learning Model for Stock Market Forecasting from Multi-Source Data

Abstract

Stock market price movement forecast from multi-source data has gained massive interest in recent years. Studies were focussed on extracting the events and sentiments from different source data and employ them in learning the stock price movement patterns. This approach provided accurate and highly reliable forecasting as it involves multiple stock price indicators. However, some aspects of sentiment analysis and event extraction increase the training time and computation complexity in big data stock analysis. To overcome these issues, the hierarchical event extraction and the target dependent sentiment analysis are performed in this paper to improve the learning rate stock price movement patterns. In this paper, the events are hierarchically extracted from news articles using Deep Restricted Boltzmann Machine (DRBM). The target based sentiments from the tweets are detected using Improved Extreme Learning machine (IELM) whose parameters are optimally selected using Spotted Hyena Optimizer (SHO). The stock indicators obtained from these two processes are used in the learning process performed using Tolerant Flexible Multi-Agent Deep Reinforcement Learning (TFMA-DRL) model for analysing the stock patterns and forecasting the future stock trends. The forecasting results obtained by using the TFMA-DRL model by combining the stock indicators of targeted sentiments and hierarchical events are trustworthy and reliable. Evaluations are performed using three datasets collected for 12 months period from three sources of Twitter, Market News and Stock exchange. Results highlighted that the proposed stock forecasting model achieved 90% accuracy with minimum training time.