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

Abstract

Annual reports are one of the most important sources of information for financial decisions. They contain forward-looking statements (FLS), which describe future trends and expectations. Thus, several studies deal with the automated identification of FLS, where the latest ones involve a combination of a rule-based approach and machine learning classification. In this paper, we extend this research with state-of-the-art NLP methods. We use DistilBERT for FLS identification and determine their sentiment with FinBERT. The result is processed by a Random Forest model for stock price growth prediction of different periods. Our evaluation shows that DestilBERT achieves higher accuracies on FLS identification than existing methods. For short-term stock price rate prediction, the extracted FLS information together with historical stock data outperforms the sole use of historical stock data. For mid-term prediction, using FLS alone with DestilBERT shows the best result. Finally, in the long-term, FLS provide no benefit.