Similarity Analysis of Knowledge Graph-based Company Embedding for Stocks Portfolio

The stocks are affected by various factors, and studies have shown that companies in the same industrial chain or with the same attributes are more closely related, and their share prices are influenced by each other. In the financial field, this kind of stock prices correlation can provide good investment suggestion. The previous method is that investors analyze the company's industrial chain relationship, industry and trading information, and then define the degree of relevance between companies. However, this approach is subjective and lack of statistical evidence. In this paper, we propose to analyze the correlation between companies by knowledge graph technology, the main algorithm is based on graph embedding. In order to analyze the production relationship between companies, we construct a company graph established from the logic of industrial chain and supplement nodes attributes with companies' information. Then, we train and calculate the vector representation of all companies by graph embedding model. Finally, we define the cosine distance of the vector as the similarity of the companies. The embedding vector of companies obtained by the industrial chain graph has achieved good results in link prediction task, visualization and practical application of stock portfolio, and the graph with weight has better effect than that without weight., the back test results show that the stock portfolio has obtained considerable extra-return.