In calm markets, assets are driven by their own fundamentals and correlations are moderate. In a liquidity event, everything is driven by one factor - the need to raise cash - and pairwise correlations converge upward.
The numbers are consistent across episodes. Equity sector pairs that sit near 0.4 in normal conditions routinely print above 0.8 during sharp selloffs; so-called uncorrelated assets are repeatedly found to be correlated when it matters. This is why stress-testing should be run with crisis correlations rather than the trailing average, and why value-at-risk estimated on calm data understates crisis losses so badly.
The implication for a trading book is blunt: your true worst day is closer to the sum of your open risks than to any diversified estimate. Size max-open-risk as though correlations are 1, and treat any diversification benefit as a bonus rather than a budget.
Related: correlation-matrix, max-open-risk, stress-testing, fat-tails