A network model finds different market hubs—and limits to its crisis signals

Efficiency as a network property

A market’s price history is only part of the picture in a framework proposed by Hitesh Kumar Sahu, Avishek Bhandari and Ipsita Parida of the Indian Institute of Technology Bhubaneswar. Their Network Adaptive Market Hypothesis combines patterns within each market’s returns with estimated information flows between markets. The work was posted as a preprint on November 6, 2026; its findings should not be treated as established by peer review.

The researchers extend the adaptive-markets view, in which efficiency changes rather than remaining a permanent market characteristic. Their model assigns each asset a local statistical-efficiency score, then adds contributions transmitted through a directed network, including indirect paths through other markets.

That score is a proxy, not a direct measurement of whether prices accurately reflect underlying economic value. The paper explicitly acknowledges that an unpredictable noise process could score highly without conveying useful fundamental information. The network-adjusted score can also exceed one: it is not a probability or a bounded percentage of informational efficiency.

Building the connections

The study uses 17 G20 equity indices and five commodity futures—gold, silver, natural gas, crude oil and copper—from January 2001 to July 2026. Its common trading calendar contains 3,361 dates. The main network analysis uses 25 overlapping windows, each containing 252 common-calendar observations and spanning roughly two calendar years.

Within each window, the authors estimate a Hurst exponent, a statistic describing persistence or reversal in returns. Values near 0.5 provide their random-walk benchmark. They also estimate transfer entropy: how much a source market’s past helps predict another market’s next return beyond that target’s own past. These are statistical predictive relationships, not a structural demonstration of economic causation. The model weights them using the markets’ Hurst estimates.

A crucial qualification concerns which connections survive. The headline network retains the strongest 30% of positive, bias-corrected transfer-entropy estimates. Under the stricter procedure intended to control false discoveries across all directed market pairs, no links survive in any window, even with 999 surrogate series per pair. The authors attribute this largely to limited statistical resolution from the surrogate counts. Nevertheless, the headline network is selected by magnitude, not validated link by link under that stricter criterion.

Japan’s hub role is conditional

In the magnitude-selected network, Japan leads eigenvector centrality in 11 of 25 windows. This measure rewards a market for receiving connections from other highly central markets. The United States leads only once by that measure, but ranks first in total outgoing connection weight in six windows—all classified as crisis windows. The authors describe this contrast as Japan functioning as a hub and the United States as a broadcaster.

The distinction is not universal. When links are selected using uncorrected statistical significance instead, Japan leads in only four windows and Australia in five. Changing the weighting formula preserves Japan’s prominence much better than changing the link-selection rule. The result therefore describes a particular estimated network, not an unconditional ranking of global financial importance.

Crises register differently across measures

The authors examine windows associated with the global financial crisis, European sovereign-debt crisis, COVID-19 and Russia–Ukraine dislocations. Their spectral radius—a matrix statistic governing repeated amplification in the model—is not detectably higher during those crisis windows. Its crisis average is 0.106, compared with 0.108 in calm windows.

That does not establish that crises lack financial contagion or amplification. It means this network statistic does not identify a consistent crisis signature at the chosen resolution. A one-year-window check finds the largest spectral radius during the global financial crisis, but still no statistically distinguishable overall crisis increase.

A separate analysis of weekly returns tells a different story. Using a model that allocates forecast-error variance among market shocks, the researchers find higher connectedness during crises. The increase comes through contributions across efficiency groups, while within-group contributions remain essentially unchanged. Connectedness is more closely associated with lower network modularity—less separation into communities—than with the spectral radius.

Similar efficiency does not mean similar connections

The researchers also divide markets into five statistical-efficiency tiers. The middle tier has the largest own-market contribution under both the variance accounting and the network-efficiency accounting. Independence from other markets therefore does not increase steadily with their efficiency ranking.

But efficiency tiers organise connectedness only marginally better than random groups of the same sizes. Geography and the distinction between equities and commodities perform much more strongly.

Important limits remain: tier assignments depend on the clustering method, overlapping windows weaken the independence of crisis comparisons, and the permutation-test simulations do not cover heavy tails or structural breaks. The study offers a framework for comparing different aspects of market networks—not a single validated gauge of efficiency or systemic danger.