Evaluating Knowledge Graph Sources for Non-personalized Financial Asset Recommendation: 10K Reports vs. Wikidata

Lubingzhi Guo, Javier Sanz-Cruzado, Richard McCreadie

Abstract: Financial asset recommender (FAR) systems suggest investment assets to customers based on past market information. Many of these models choose those securities which they estimate to be more profitable for customers. Financial knowledge graphs (KGs) – data structures containing information about assets and their relations to other involved entities (companies, people) – have been one of the data sources exploited to drive asset selection. Although the construction of knowledge graphs from different sources (news, reports) has previously been investigated, there has been limited analysis of the effect these construction strategies have for FAR. In this work, we compare two different knowledge graphs representing U.S. stocks under a unified FAR framework: a knowledge graph crawled from a general knowledge base, Wikidata, and a knowledge graph built by extracting entities and relations from 10K financial reports using the GoLLIE open information extraction model. We show that integrating these KGs in FAR can lead up to 10.7% improvements in monthly ROI. However, the nature of these graphs makes algorithms prone to bias the recommendations towards different asset types. Therefore, we further propose and evaluate an adaptive graph selection strategy, which dynamically chooses the suitable graph prediction model—trained on either the 10K Graph or the Wikidata Graph—for each asset. The findings indicate that stock-level and sector-level selection strategies respond differently to the length of the recency window, reflecting, respectively, a preference for short-term responsiveness and long-term stability.

External IDs:dblp:conf/keir/GuoSM25