Transformer-based Summarization and Sentiment Analysis of SEC 10-K Annual Reports for Company Performance Prediction

Abstract

Annual reports published by companies contain important insights regarding their performance and are often analyzed in a manual, subjective manner. We address this point by combining the streams of research on text summarization and topic modelling with the one on sentiment analysis. Our approach consists of the steps of text summarization using BERTSUMEXT, topic modelling with LDA, sentiment analysis with FinBERT, and performance prediction with Decision Trees and Random Forest. The result provides decision makers with an interpretable and condensed representation of the content of annual reports, together with its relationship to future company performance. We evaluate our approach on 10-K reports, demonstrating both its interpretability for analysts and explanatory power regarding future company performance.

Extraction of Forward-looking Financial Information for Stock Price Prediction from Annual Reports Using NLP Techniques

“…Items 1A, 3 and 7A deal with specific risk factors the company is (or might be) facing and therefore also potentially contain important FLS. Among other things, 10-K reports have been analyzed to predict stock prices (Hsieh & Hristova, 2022) or for the report's effects on brand value (Huang, Liu, & Xie, 2020).…”

“…Thus, for each sentence, each sentiment class, each report and each FLS dataset, we obtain the corresponding score from the model. To aggregate that at a report level, we calculate the score mean for each sentiment class over all FLS in the report (Hsieh & Hristova, 2022) resulting in mean_positive, mean_neutral and mean_negative. In order to be able to represent the report with a single sentiment value, we define the variable sentiment_score as mean_positive-mean_negative.…”

“…In particular, they contain both quantitative financial metrics on company performance and the corresponding textual analysis of those numbers from the management's perspective. These reports are mostly manually read to extract important information for decision-support (Hsieh & Hristova, 2022).…”

“…This is an annual report required by the U.S. Securities and Exchange Commission (SEC) for all publicly traded companies. As a result, researchers recognized the need for approaches to automatically analyze annual reports, thus increasing efficiency and objectivity (Hsieh & Hristova, 2022;Lewis & Young, 2019).…”

“…sentiment, cf. Krinitz and Neumann (2021)) are extracted from the resulting output and 3) those are related to company performance, such as the prediction of IPO valuation (Tao, Deokar, & Deshmukh, 2018), stock price (Hsieh & Hristova, 2022;Kraus & Feuerriegel, 2017) or period returns (Loughran & McDonald, 2011).…”