Deep Learning Model for Fusing Spatial and Temporal Data for Stock Market Prediction

One of the most significant challenges in stock market forecasting is that the majority of stock price analysis and prediction models based on quantitative data rely on stock trends as their primary metric. These metrics may exhibit good intra-day performance, but have scalability issues when applied to inter-day trading, limiting their accuracy for real-time prediction. The systems that use news and social media data have limited correlation capabilities due to inefficiencies in trend analysis. Fusion of these data sources is expected to improve stock market prediction accuracy. This paper proposes a novel augmented analysis model for fusing spatial and temporal stock trends with global–local market movements via incremental learning. The novel multiparametric augmentation model is based on hybrid of machine and deep neural architectures like support vector machine, recurrent neural network and convolutional neural network. The model integrated five heterogeneous data sources. The model initially identifies global news trends, and correlates them with temporal and spatial stock values. This correlation is further improved by evaluation of local news trends with respect to stock specific geographies and events. This assists in identification of spatial and temporal factors that drive a particular stock’s value, and improve the efficiency of trend analysis. The estimated trend is combined with an incremental learning model, that estimates intra-day stock values with respect to incremental value variance. The proposed model has been tested on numerous local and international stocks from Ten different sectors namely IT, healthcare, energy, communication services, financial, industrial, real estate, consumer discretionary, consumer staples and others over a period of 300 days. The highest accuracy of 95% is observed in terms of stock trend prediction and 99% for stock value close price prediction, with average accuracy of 97.36%.