Towards a generic framework for short term firm-specific stock forecasting

This paper investigates the predictive power of technical analysis, sentiment analysis and stock market analysis coupled with a robust learning engine in predicting stock trends in the short term for specific companies. Using large and varied datasets stretching over a duration of ten years, we set out to train, test and validate our system in order to either contradict or confirm efficient market hypothesis. Our results reveal a significant improvement over the efficient market hypothesis for majority companies and thus strongly challenge it. Technical parameters and algorithms operating upon them are shown to have a significant impact upon the end-predictive power of the system, thus bolstering claims of their efficacy. Moreover, sentiment analysis results also show a strong correlation with future market trends. Lastly, the superiority of supervised non-shallow learning architectures is illustrated via a comparison of results obtained through a myriad of optimization and clustering algorithms.