Enhancing the Accuracy of Stock Prediction by Utilizing Existing Machine Learning Techniques

Abstract

The present volatility of the stock markets makes forecasting stock trends extremely newlinechallenging owing to several socio economic and political factors other than market newlinetrends. While machine learning models can be used to perform regression analysis newlinebased on historical data trends, it becomes extremely challenging to incorporate the newlinevariabilities which are non-numeric in nature. Some of the factors which govern the newlinerise and fall of stock prices are socio economic conditions, trade wars, current newlinepandemic situation and global market slowdown, reliability of a company among newlineothers. Hence, one of the most effective ways to incorporate these trends is analyzing newlinepublic trends pertaining to the same. While public sentiments may not always be newlinecoherent with prevailing market trends, yet they often portray the existential trends newlinein the market and opinions of the public regarding potential purchases of stocks of a newlineparticular company in a given time period. This work presents an approach which is newlinean amalgamation of deep nets with attention, and opinion mining for forecasting newlinestock trends. The attention vector employed as an additional input computed on the newlinemoving average allows for current trend analysis along with opinion mining from newlinepublic datasets encompassing both numeric data trends and non-numeric data newlineparameters pertaining to global influencing features. A naturally pragmatic approach newlineseems to be designing a model trained on historical data, which is able to forecast newlineboth in and out sample. This would render high degree of robustness and practical newlineutility to the developed model. The regression and forecasting accuracy have been newlinecomputed on a diverse set of datasets to validate the performance of the proposed newlineapproach.