What question did this study set out to answer?

This analysis aims to understand how different AI and quantitative risk strategies respond during market disruptions.

May 29, 2026Open Access

When Machines Disagree A Comparative Analysis of Four Institutional Risk Architectures Under Regime Breaks (March 2020 and August 2024)

Key Points

This analysis aims to understand how different AI and quantitative risk strategies respond during market disruptions.
Comparative analysis of four institutional risk architectures during March 2020 and August 2024 regime breaks.
Utilized publicly available proxies and ETFs to assess performance and correlations among strategies.
Conducted event-window regressions to evaluate lagged bid-ask spread coefficients.
In March 2020, average correlations among strategies increased significantly from 0.25 to 0.50.
Significant negative lagged bid-ask spread coefficients for AGG (p = 0.006), HYG (p < 0.001), and MTUM (p = 0.005) were identified.
In August 2024, correlations increased by only 19%, and no strategies showed significant negative spread coefficients.

Abstract

When markets break, do architecturally distinct AI and quantitative strategies fail indistinct ways, or do they converge on the same failure mode? This paper examines fourinstitutional risk architectures during two regime breaks: the March 2020 COVID drawdown(a credit-driven liquidity crisis) and the August 2024 yen carry unwind (a rates-driven eventconcentrated in Japan). The four architectures, execution-layer AI (JPMorgan), factor-covariance risk modelling (BlackRock/Aladdin), regime-aware risk parity (Bridgewater), andmulti-factor systematic strategies (Two Sigma), are observed through publicly availableproxies and ETFs. In 2020, average pairwise correlations across strategy proxies nearlydoubled (0.25 to 0.50), and event-window regressions identified significant negative laggedbid-ask spread coefficients for AGG (p = 0.006), HYG (p < 0.001), and MTUM (p = 0.005), apattern consistent with forced selling through the Brunnermeier-Pedersen funding-liquiditychannel. In 2024, correlations rose only 19%, no strategy showed a significant negativespread coefficient, and the Bridgewater risk-parity replicator gained +0.9% versus its -15.7%drawdown in 2020. The pattern is consistent with the credit-liquidity hypothesis: theregression signature appeared during the U.S. credit event and was absent during the non-credit regime break. Two events cannot prove this is a causal mechanism. Architecturaldiversity did not prevent correlated failure in 2020. The evidence is consistent with acommon-mode channel operating at the level of funding markets rather than modelarchitecture, though the analysis does not isolate this channel from alternativeexplanations including leveraged-position deleveraging and dealer balance sheetconstraints. Two events cannot prove causation, and the proxy-based design introducesmagnitude uncertainty, but the findings suggest that funding-channel exposure may be arelevant dimension for systemic risk supervision alongside model similarity. Data pipelineand analysis code are publicly available for reproduction and out-of-sample testing

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper