When to Use

• When allocating capital across multi-strategy portfolios and the allocation assumes diversification survives a crash.
• When auditing a newly onboarded sub-strategy whose full-sample correlation to the book looks benign.
• During stress testing, when you need the joint-crash probability of a pair rather than its average comovement.
• When calibrating portfolio-level risk limits that are only binding in extreme regimes.

When NOT to Use

• As a capital control on its own. The output is evidence for an allocation committee, not a limit engine. Enforce with correlation-aware-exposure-limits and multi-strategy-capital-allocation-limits.
• On short histories. At $\alpha = 0.10$ an independent pair puts only about $\alpha^2 n$ observations in the joint tail — roughly 5 for $n = 500$. Below min_tail_observations the engine returns is_determinate=False; that is a statement of ignorance, not a clean bill of health.
• For upper-tail or asymmetry questions. This module measures the lower tail only. Correlation asymmetry between the two tails is a separate estimation problem.
• As a substitute for full-sample correlation monitoring. Use cross-strategy-correlation-monitoring for rolling $\rho$ and diversification ratios; this skill answers a narrower, harder question.
• To read lower_tail_matrix into an optimizer. Every pair is estimated on its own overlap and its own tail subsample, so the matrix is not guaranteed positive semi-definite.

Prerequisites

• Overlapping daily return series per pair, sharing one timestamp index, with at least min_observations (default 20) aligned non-null rows — and realistically far more, since the joint tail is what must be populated, not the sample.
• No non-finite values and no zero-variance (flat, stale or idle) series; the engine rejects both rather than imputing them.
• Python 3.10+ with numpy and pandas.

Workflow

1. Align and validate. Join the pair on its index and drop non-overlapping rows. Equal series lengths do not imply a shared index — alignment happens before any sufficiency check, and the count of dropped rows is logged. Reject ±inf, non-numeric values, duplicate index labels and zero-variance series outright.
2. Compute the unconditional correlation. This is context, not the comparison baseline (see step 5).
3. Compute the joint-tail exceedance correlation. Take the marginal $\alpha$-quantiles $q_A, q_B$ and correlate the observations where both $R_A \le q_A$ and $R_B \le q_B$ (the intersection, per Longin–Solnik and Ang–Chen). If fewer than min_tail_observations survive, or the tail slice is flat, stop: report is_determinate=False and NaN. Do not substitute a number.
4. Compute empirical tail dependence. $\hat\chi(\alpha) = \mathbb{P}(R_B \le q_B \mid R_A \le q_A)$. Read it against the independence baseline of $\alpha$ — under independence this statistic equals $\alpha$, not zero.
5. Benchmark against a Gaussian copula. Simulate the same estimator on bivariate normals drawn at the pair's own $\rho$, $n$ and $\alpha$. Report the excess over that benchmark and a one-sided p-value. This is the decision variable; the raw $\rho_{\text{tail}} - \rho_{\text{uncond}}$ delta is reported for continuity only and is dominated by selection bias.
6. Flag breakdown. Warn when the exceedance correlation reaches breakdown_threshold in absolute level, or when its excess over the Gaussian benchmark reaches breakdown_excess_threshold at a p-value no greater than breakdown_max_pvalue. Detection is one-sided: negative tail comovement is a diversification benefit, not a breach.

Common Pitfalls

• Comparing tail correlation to full-sample correlation. Conditioning a sample on the size of its own variables changes the correlation of the retained subsample even when the true correlation is constant (Boyer, Gibson & Loretan 1997; Forbes & Rigobon 2002). A negative "delta" is the expected result for a perfectly well-behaved Gaussian pair, not a finding. Compare against a simulated null instead.
• Conditioning on the union of the two tails. Selecting rows where $R_A \le q_A$ or $R_B \le q_B$ retains an L-shaped region in which low-$A$ days pair with typical $B$ and vice versa. This manufactures strong negative correlation: a bivariate normal with true $\rho = 0.6$ scores about $-0.19$. Version 1.0.0 of this skill did exactly that and could therefore almost never fire.
• Reading a thin joint tail as diversification. The dangerous failure is a reassuring number computed from four observations. Treat is_determinate=False as "unmeasured", and require the joint tail to be populated before signing off on an allocation.
• Mistaking $\hat\chi(\alpha)$ for the copula coefficient $\lambda_L$. $\lambda_L$ is the limit as $u \to 0^+$; $\hat\chi(0.10)$ is a finite-level estimate whose independence baseline is $0.10$. A Gaussian pair with $\rho = 0.6$ scores $\hat\chi(0.10) \approx 0.39$ while being asymptotically tail independent ($\lambda_L = 0$). Judging "severe coupling" without that baseline flags ordinary correlation as tail risk.
• Assuming Gaussian joint distributions in the risk model itself. A bivariate normal with $|\rho| < 1$ has $\lambda_L = 0$ at any correlation (Embrechts, McNeil & Straumann 2002), so a Gaussian portfolio model structurally cannot produce joint crashes. That is precisely why this skill measures the excess over it.
• Injecting identical crash values into test fixtures. A block of constant crash returns gives the tail slice zero variance and an undefined correlation. The engine returns NaN rather than a spurious $\pm 1$.

Verification

Run the test suite:

python -m unittest discover -s skills/tail-correlation-between-strategies-under-stress/scripts

Expected statistical behavior (reproduced by the suite): a Clayton pair with $\lambda_L \approx 0.71$ is flagged; independent and Gaussian pairs at $\rho$ up to $0.8$ are not; a 30-observation sample returns is_determinate=False.

Related Skills

• cross-strategy-correlation-monitoring
• correlation-aware-exposure-limits
• portfolio-stress-test-including-liquidity-crunch-scenarios
• stress-testing-against-historical-crash-scenarios