Combining bag-of-words and sentiment features of annual reports to predict abnormal stock returns

Abstract

Automated textual analysis of firm-related documents has become an important decision support tool for stock market investors. Previous studies tended to adopt either dictionary-based or machine learning approach. Nevertheless, little is known about their concurrent use. Here we use the combination of financial indicators, readability, sentiment categories and bag-of-words (BoW) to increase prediction accuracy. This paper aims to extract both sentiment and BoW information from the annual reports of U.S. firms. The sentiment analysis is based on two commonly used dictionaries, namely a general dictionary Diction 7.0 and a finance-specific dictionary proposed by Loughran and McDonald [1]. The BoW are selected according to their tf-idf. We combine these features with financial indicators to predict abnormal stock returns using a multi-layer perceptron neural network with dropout regularization and rectified linear units. We show that this method performs similarly as Naïve Bayes and outperforms other machine learning algorithms (Support Vector Machine, C4.5 decision tree, and k-nearest neighbour classifier) in predicting positive/negative abnormal stock returns in terms of ROC. We also show that the quality of the prediction significantly increased when using the correlation-based feature selection of BoW. This prediction performance is robust to industry categorization and event window.

Sustainability and Corporate Social Responsibility in the Text of Annual Reports—The Case of the IT Services Industry

“…Another limitation of this study is the use of a dictionary-based approach. Although we addressed the problem of context-specific nature of sustainable development vocabulary, machine learning approaches, such as deep neural networks, may provide more accurate evaluations of sentiment [91]. However, it must be noted that those approaches are difficult to interpret.…”

Integrating Sentiment Analysis and Topic Detection in Financial News for Stock Movement Prediction

“…These include 354 positive and 2,329 negative words (available at https://sraf.nd.edu/textual-analysis/resources/). In agreement with previous studies [7,22,25], we calculated the raw frequencies of positive (POS) and negative (NEG) words in the company-related financial news. The problem of negations of positive words was addressed by using collocation analysis with negation words in QDA Miner 5.…”

“…More recently, topic detection has also attracted attention in related literature because topics discussed in the financial news or social media may carry additional important information for investors [6]. However, the advantages of both textual analysis approaches have been studied only for corporate annual reports [7] and social media, namely message board [6]. These studies have shown that combining sentiment analysis and topic detection improves the accuracy of financial prediction models.…”

“…Therefore, here we propose a system integrating sentiment analysis and topic detection. In conformity with the state-of-the-art literature [7], we employ a deep neural network to train this system to predict stock movements. We demonstrate that this approach is more effective not only than traditional support vector machine (SVM) method but also than sentiment analysis and topic detection used separately.…”

“…For topic detection, we employ latent dirichlet allocation (LDA), a generative probabilistic model that is used to detect a mixture of hidden topics in texts [6]. To perform the one-day-ahead prediction of stock movements (up or down), we use a deep feedforward neural network with dropout regularization and rectified linear units [7], hereinafter referred to as the DNN. This method overcomes the limitations of traditional machine learning methods used for the stock movement prediction, such as problem with overfitting and optimization convergence to a poor local minimum.…”