Forecasting Stock Price Changes Using Natural Language Processing

Abstract

All investors attempt to predict stock market returns when they make investment decisions. However, making such predictions is not a trivial task. As a result, many strategies have been proposed by researchers as potential ways to predict stock returns. More recently, data analytics - in general - and natural language processing, in particular, have been identified as viable options. In this project, we investigate the use of natural language processing to forecast stock price changes. Specifically, we analyze firms’ 10-K and 10-Q reports to identify sentiment. Using the computed sentiment scores, we develop models to predict the direction of stock price movements both in the short run and in the long run. Our first step in developing these models is to investigate some sentiment scoring methods and apply them to the Loughran-McDonald dictionary. Next, we use the model word2vec to extend the usage of the Loughran-McDonald dictionary and then apply the sentiment metrics. Additionally, we apply the proposed sentiment metrics to FinBERT, which learns contextual relations between words. Finally, we build supervised machine learning algorithms that use the proposed sentiments as inputs to forecast price changes. We train our algorithms on 10-K and 10-Q reports of 48 companies in the S&P 500 from 2013 to 2017. Finally, we test our models on the corresponding reports from 2018 to 2019 and conclude that predictive signals can be extracted from 10-K and 10-Q reports.