Correlation Under Stress: Why Your Hedge Stopped Hedging at 14:30

Cross-asset correlations are the foundation of every multi-leg hedge in institutional FX. A risk team that does not understand the conditional structure of those correlations is a risk team that will, eventually, watch a hedge stop hedging at exactly the moment the hedge was supposed to matter. This post is about why.

The calibration problem

The standard practice for estimating cross-asset correlations is to compute a rolling Pearson correlation over a window of historical returns — typically 60 to 250 days. The resulting number is the average correlation across that window, weighted by however many observations the window contains.

The implicit assumption is that the average correlation is the relevant statistic for hedging purposes. It is not. The relevant statistic is the conditional correlation in the regime in which the hedge will be tested — which, for most institutional desks, is a stress regime that constitutes 10-30% of trading hours and dominates the loss tail.

A calibration that averages across calm and stress regimes under-weights stress, because there are fewer stress observations. The resulting average correlation is therefore biased toward the calm-regime value. The bias is precisely in the direction that makes the hedge look more reliable than it actually is when called upon.

What happens to correlations under stress

Three things, in roughly this order:

1. Inter-asset correlations migrate toward ±1

In a global risk-on or risk-off move, assets that are normally weakly correlated converge. EUR/USD, gold, and the Nasdaq-100 spend most of their lives caring about different things; in a stress event they all care about the same thing, and their conditional correlation jumps. This is the classic 'correlation goes to 1 in a crisis' observation, and it is the easiest part of the problem to anticipate.

2. Intra-asset basis relationships break

The relationship between an FX pair and its forward curve, or between a spot and its equivalent NDF, has a strong arbitrage anchor in normal conditions and a much weaker one under stress. Funding constraints widen, balance-sheet capacity contracts, and the basis that a desk was using as a cheap hedge becomes both expensive and unstable. This is the part of the problem that surprises desks most, because the basis hedge was — yesterday — the obvious answer.

The mechanics of why the basis breaks down — and what it means for the wholesale spread you pay on the residual — are part of the broader topic covered in The Balance-Sheet Cost of Leverage.

3. Lead-lag structures invert

In calm regimes, a fast asset tends to lead a slow asset by a predictable interval. In stress, the lead-lag structure inverts on a per-event basis: which asset moves first depends on which channel the stress arrives through, and the channel is generally not the one the previous month's flow data would predict. A hedge that worked because of an implicit lead-lag assumption is at best decorrelated, at worst inverted, in a regime change.

Calibrating for the regime you care about

There is no single right answer to the calibration problem, but there are several wrong answers it pays to avoid:

• Equal-weighted Pearson on a year of daily returns. The dominant signal is the calm regime, the relevant signal is the stress regime, and the hedge will be confidently miscalibrated.
• Exponentially-weighted Pearson with a short half-life. Tracks current correlation but is procyclical — it adapts to the new regime exactly when the new regime has already done the damage.
• A copula fit on the same equal-weighted dataset. The tail dependence is mechanically what the body dependence said it would be; nothing has been learned about the tail.

Two approaches that work better in practice:

• Regime-conditional estimation: tag your historical days as calm, normal, or stress using a regime-identification model (a simple volatility-percentile threshold works); estimate correlations within each regime; use the regime-specific correlation in the corresponding regime.
• Tail-conditional estimation: compute the correlation on the days where one asset moved more than its 95th-percentile daily range. The resulting number is a direct estimate of correlation in the conditional event you actually care about.

What Drovix does

Drovix's risk engine maintains both regime-conditional and tail-conditional cross-asset correlation matrices, updated daily on a rolling 3-year window of intraday data. The wholesale-book pricing for multi-leg structures uses the tail-conditional matrix; the streaming-book pricing for single-leg flow uses the regime-conditional matrix at the live regime tag. The choice of matrix is exposed to the counterparty in the trade ticket, so a risk officer reading a hedge ticket can see exactly which estimate is being used.

The point is not that our matrices are the best possible. The point is that the choice is explicit and reviewable, rather than buried behind a single 'correlation' number that drifts silently. A counterparty whose hedge breaks because the correlation regime shifted has at least a documented record of the assumption that did not hold; that record is the precondition for either improving the next hedge or pricing the next residual correctly.

Operational implications

• Backtest your hedges separately on calm-regime, normal-regime, and stress-regime sub-samples. If the calm and stress performance differ by more than a factor of 2, your sizing should reflect the worse number, not the average.
• Define an exit trigger that fires when the live regime tag flips. The hedge that worked yesterday may not be the hedge that works tomorrow, and the alpha of acting early on the regime change is large relative to its operational cost.
• Demand from your counterparty an explicit statement of which correlation regime is priced into the spread you are paying. Refusal to provide it is a signal.

Where to go next

→ Signed Flow and Information Half-Life — the time-domain analog of this regime question: your flow's predictive content also changes with regime.

→ B-Book to Wholesale — how a regulated retail broker's residual hedge is priced when its statistical character is correctly estimated.