Fund Similarity: A Use of Bipartite Graphs

Fund similarity is important for investors when constructing diversified portfolios. Because mutual funds do not always adhere closely to their stated investment policies, investors may unintentionally hold funds with overlapping exposures, leading to reduced diversification and instead causing “diworsification”, which is an investment term for when too much complexity leads to worse results. As a result, various quantitative methods have been proposed in the literature to investigate fund similarity, primarily using portfolio holdings. Recently, machine learning tools such as clustering and graph theory have been introduced to capture fund similarity. This paper builds on this literature by applying bipartite graphs and node2vec embeddings to a comprehensive dataset that covers 3874 funds over a nearly 6-year period. Our empirical evidence suggests that, bipartiteness is not preserved for non-index (active) funds. Furthermore, while graph embeddings yield higher similarity scores than holding-based measures, they do not necessarily outperform holding-based similarity in explaining returns. These findings suggest that graph-based embeddings capture structural relationships among funds that are distinct from direct portfolio overlap but are not sufficient substitutes when similarity is evaluated solely through returns. As a result, we recommend a more comprehensive similarity measure that includes important risk metrics such as volatility risk, liquidity risk, and systemic risk.