SVNexus · Dependence
Tail dependence
Also called: tail dependence coefficient, extreme co-movement, joint tail risk
Tail dependence measures the chance that two assets suffer extreme moves together, which ordinary correlation can understate badly.
Definition
Tail dependence describes how likely one asset is to have an extreme move given that another has one. The lower tail dependence coefficient measures joint crashes; the upper one measures joint spikes.
It is a property of the copula, the part of a joint distribution that describes dependence separately from each asset's own distribution.
Intuition
Two assets can have a modest correlation on ordinary days and still fall together in a crisis. Correlation averages over all days, while diversification fails exactly on the worst ones. Tail dependence focuses on those days.
A Gaussian model, which underlies many risk systems, has zero tail dependence for any correlation below one, meaning joint extremes become independent in the limit. That is one reason such models understated joint losses in past crises.
How it's calculated
With U and V the probability-integral transforms (ranks scaled to 0–1) of the two return series and C their copula:
λL = limu→0⁺ P(U ≤ u | V ≤ u) = limu→0⁺ C(u, u) / u
λU = limu→1⁻ P(U > u | V > u) = limu→1⁻ [1 − 2u + C(u, u)] / (1 − u)
Parametric values for common copulas: Gaussian, λ = 0; Student-t, positive in both tails and larger for fewer degrees of freedom; Clayton, λL = 2−1/θ; Gumbel, λU = 2 − 21/θ. Non-parametric estimates use the empirical frequency of joint exceedances below a low quantile, such as the 5th percentile.
Example: if two assets each have 50 days below their own 5th percentile in a 1,000-day sample and 20 of those days coincide, the empirical conditional probability is 20 / 50 = 0.4. Under independence it would be about 0.05, and under a Gaussian model with moderate correlation it would sit well below 0.4 and shrink further at more extreme quantiles.
How to read it
Values near zero mean extremes rarely coincide; values of 0.3 or more mean a large share of one asset's worst days are also bad days for the other. Compare lower and upper tails: equity pairs typically show more dependence in crashes than in rallies.
Estimates rely on few observations, so confidence intervals are wide. Look at stability across windows and quantile choices.
Common mistakes
- Using correlation as a stand-in for crash co-movement.
- Fitting a Gaussian copula and concluding tails are independent.
- Estimating from a short calm sample with too few extreme days.
- Ignoring time variation; tail dependence often rises in stressed regimes.
In Saqrid
SVNexus can build its network from tail-dependence links as well as return correlation, DCC correlation, transfer entropy or implied-vol spillover.
A pair drill-down shows tail risk, lead-lag, regime sync, wavelet coherence and correlation dynamics.
Open SVNexus.
Go deeper
Saqrid Institute lesson: Tail Dependence & Joint Extremes. Lesson text is open to read without an account.