Accounting choices reshape which Chinese firms look systemically central

A new study in Applied Network Science asks a deceptively simple question about financial risk research: when scientists build networks of firms to find the most systemically important ones, does the answer depend on how those networks are measured? For listed companies in mainland China and Hong Kong between 2006 and 2023, the researchers find that it does, and sometimes dramatically.

Most studies of financial networks infer links between firms either from market co-movement, where share prices move together, or from observed exposures, such as direct lending relationships. The team behind the new paper instead starts one step earlier, with a measurement question. They build networks from accounting data alone, connecting firms whose financial profiles look similar, and then check whether the same firms remain central when similarity is defined in different accounting spaces.

The method works like this. Each year, the researchers standardize a set of accounting ratios for every listed firm, so that each company is described by a vector of standardized financial characteristics. They then compute Euclidean distances between firms in these ratio spaces and transform those distances into continuous similarity weights, producing a weighted, undirected network for each year. Firms with similar balance-sheet profiles end up strongly connected; firms with dissimilar profiles are weakly connected or not at all.

Crucially, the design compares several representations built from the same underlying data. A GLOBAL layer uses the full vector of ratios. A RISK layer captures balance-sheet vulnerability, and a RETURN layer captures profitability. The authors also construct a family of single-ratio networks, each based on one financial ratio alone. Within each representation, weighted eigenvector centrality measures which firms are recursively prominent, meaning they are central and connected to other central firms.

The composite layers turn out to differ systematically in how tightly connected they are. Average network density, the share of possible connections that carry meaningful weight, is 0.286 in the GLOBAL layer, 0.384 in the RISK layer, and 0.457 in the RETURN layer. In other words, firms resemble each other most in profitability terms and least when the full financial profile is considered.

But cohesion and the concentration of centrality do not move together. GLOBAL is the least dense of the three composite layers, yet it concentrates the most centrality at the top: the ten most central firms hold 11.07 percent of total centrality there, compared with 9.72 percent in RISK and 9.07 percent in RETURN. A sparser network, counterintuitively, can funnel prominence into fewer hands.

The identity of the core firms is even more representation-dependent. On average across years, the top-ten sets of the GLOBAL and RISK networks share 5.50 firms. GLOBAL and RETURN share only 1.78, and RISK and RETURN share just 1.33. A firm that looks systemically prominent through a broad financial lens may nearly vanish from the core when the lens narrows to profitability, or the reverse.

The single-ratio networks show still greater topological dispersion, and the study documents a notable divergence after 2017: RISK and RETURN densities frequently move in opposite directions from one year to the next. That pattern suggests that balance-sheet vulnerability and profitability can follow different cross-sectional configurations across the Chinese listed universe, rather than rising and falling together.

The authors frame the contribution as a result about measurement rather than about any single ranking of dangerous firms. Graph cohesion, centrality concentration, and core membership are distinct properties, and each depends on the accounting representation used to construct the network.

The paper is also explicit about what these networks cannot do. They capture common financial-profile proximity, not bilateral exposures between firms, and they do not identify causal spillovers or the contribution of any firm to systemic loss. The findings are descriptive measurements of network structure, not estimates of how distress would actually propagate through the Chinese financial system.

Subject of Research: Multilayer financial similarity networks and firm centrality among listed Chinese firms

Article Title: Multilayer financial similarity networks and systemic firm centrality in China

Article References: Yu, Y., Rufé, M. C., Pidelasserra, J. M., & El Zein, S. A. (2026). Multilayer financial similarity networks and systemic firm centrality in China. Applied Network Science. https://doi.org/10.1007/s41109-026-00833-z

Image Credits: AI Generated

DOI: 10.1007/s41109-026-00833-z

Keywords: financial networks, systemic risk, eigenvector centrality, accounting ratios, China, Hong Kong, network density, multilayer networks, measurement, listed firms, Multilayer, financial