Diversification is one of the oldest and most widely taught principles in asset management, stemming from a central idea that combining imperfectly correlated assets can reduce portfolio risk without sacrificing expected return. In practice, this principle has traditionally been implemented by summarizing how asset returns move together on average—using measures including covariance matrices, correlation estimates and factor models.
These representations are powerful, but they quantify how assets move in relation to each other without explaining the “why” or how their underlying structure links movements. For example, a portfolio diversified across technology, industrials, energy, and consumer sectors may appear well balanced based on traditional risk metrics but still highly exposed to one underlying economic driver. As a result, portfolios that seem diversified on paper could become highly correlated in stress periods—when diversification matters most. This poses a recurring challenge in portfolio construction.
We believe that network-based representations, which visually illustrate relationships and correlations among different assets, are a complementary way to think about diversification. Instead of viewing correlation as a table of numbers, this approach treats assets as nodes and relationships as connections, making it easier to see clusters of tightly connected assets, the pathways through which shocks are transmitted, and concentrations of exposure. This has the potential to enhance the understanding, practice and application of diversification and portfolio analysis.
Through a conceptual framework and practical applications, we show how graph-based representations connect naturally to traditional covariance-based approaches. This may offer more insight into how assets move in relation to each other, how risks concentrate, and spillover effects that are central to today’s asset management. This “market as a graph” concept has developed over decades, spanning portfolio theory and systemic risk research. Recent advances in graph-based learning and neural networks like graph neural networks (GNNs) treat the market explicitly as a graph that can be used for forecasting and portfolio decision-making.
We also reframe diversification through a network lens and discuss how graph-based representations address the shortcomings of classical techniques. We highlight methods to construct market graphs, describe how GNNs may improve investment insights and provide practical examples that enable portfolio managers to enhance decision-making.
