A Machine Learning Approach: Enhancing the Predictive Performance of Pharmaceutical Stock Price Movement during COVID

Abstract

Predicting stock price movement direction is a challenging problem influenced by different factors and capricious events. The conventional stock price prediction machine learning models heavily rely on the internal financial features, especially the stock price history. However, there are many outside-of-com-pany features that deeply interact with the companies’ stock price performance, especially during the COVID period. In this study, we selected 9 COVID vaccine companies and collected their relevant features over the past 20 months. We add-ed handcrafted external information, including COVID-related statistics and company-specific vaccine progress information. We implemented, evaluated, and compared several machine learning models, including Multilayer Perceptron Neural Networks with logistic regression and decision trees with boosting and bagging algorithms. The results suggest that the application of feature engineering and data mining techniques can effectively enhance the performance of models predicting stock price movement during the COVID period. The results show that COVID-related handcrafted features help to increase the model prediction accuracy by 7.3% and AUROC by 6.5% on average. Further exploration showed that with data selection the decision tree model with gradient, boosting algorithm achieved 70% in AUROC and 66% in the accuracy.