Fund managers at investment trust management companies make investment policy decisions by referring to the results of research on candidate companies for investment compiled by analysts. However, when there are many candidate companies, it is necessary to refer to a considerable number of reports, which is considered dificult to read carefully. Therefore, technology is required to (1) accurately determine the business sentiment of the companies concerned and (2) extract important information for investment decisions from the contents of the reports. In this study, we developed a machine learning model that predicts the rating, which is a rating index for investment decisions, in order to support the work of fund managers, especially for the requirement (1). There are two types of ratings that afiect investment decisions: outperform and underperform. Since the number of cases that fall into these two categories is small compared to other ratings, we attempted to expand the data of documents that give the same rating. As an existing data expansion method, there is a method to expand data by synonyms that can be obtained from WordNet. In this study, we propose a method to expand data based on the frequency of occurrence in financial documents. As a result of experiments, we verified the efiectiveness of the proposed data expansion method.
