News Sentiment and Company Reports Impact on Stock Returns

Abstract

Stock market returns prediction is a complicated topic with ever-growing interest within the past 20 years. Recent studies have been constantly trying to find more useful predictors, which quite often were found to be extracted from text sources such as social networks or discussion forums. In this study, we propose to take a closer look at the information contained in Securities and Exchange Committee (SEC) reports in the form of 10-K and 10-Q. In particular, we focus on analyzing the Management Analysis and Discussion (MD&A) section.To perform our analysis, we collected and processed almost 300,000 reports following the best practices of distributed computing on PySpark. It is worth noting that, unlike many other research studies in the same field, we consider not only S&P 500 companies but a broader range of companies as well. Based on the MD&A sections, we created a Bag-of-Words (BOW) corpus, and then utilized the TF-IDF technique to determine the weight of each word.To evaluate the sentiment of the text, we utilized an exhaustive dictionary of words allocated to different emotional categories. This allowed us to compute sentiment scores for each category. Our findings indicate that certain emotions have a significant correlation with stock returns in the following quarter and can improve the performance of price movement direction classification models. Interestingly, commonly extracted Positive and Negative sentiment scores were found to have little importance, while the Weak and Understatement categories emerged as the top features.