Peer Firm Identification Using Word Embeddings

In the task of peer firm identification, researchers have relied on existing industry classification system regardless of their critical limitations. In the existing industry classification system, a company should be categorized into one group regardless of the number of products and services it offers. Furthermore, it is not possible to measure the similarity of companies belonging to the same group. The systems are revised manually, rendering it difficult for them to keep up with the fast-changing industry landscape. In this paper, we propose a novel peer firm identification method based on Word 2Vec. By computing the cosine similarity of word embedding vectors trained on a 10-year corpus of financial news articles, we developed a method that produces peer firms with their numeric similarity scores. Our approach allows us to observe chronological changes in the peer firms by having firm words that appear in news articles from different periods in the same vector space. Last but not least, our Word 2Vec-based method produced more economically homogeneous groups of peer firms compared to the existing classification systems.