Download PDF
Download page Copula | Copula Selection.
Copula | Copula Selection
Neptune fits 6 copulas (Gaussian, Student's t, Gumbel, Clayton, Frank, and Joe) to each copula scenario. The Akaike Information Criterion (AIC) is computed and tabulated for each copula model. The Akaike Information Criterion is computed using the following equation (Burnham and Anderson 1998):
AIC_c = \frac{2Nk}{N-k-1} - 2ln(\hat{L})
where k is the number of parameters estimated by the probability model, N is the number of data points, and \hat{L} is the maximized value of the likelihood function of the probability model i.e. \hat{L} = p(x|\hat{\theta}, M) where \hat{\theta} are the parameter values that maximize the likelihood function and M is the probability model.
The above form of AIC is used for small samples sizes (N/k < 40). As the sample size increases with respect to the number of parameters, this equation converges to the standard form of the AIC:
AIC = 2k - 2ln(\hat{L})
If the user entered two copula scenarios on the Copula Scenarios tab, the Next and Previous chips in the upper right of the tab (boxed in red) can be used to toggle between the copula model fits for each scenario. The name of the copula scenario is shown in the upper left.

In addition, the samples and probability isolines are plotted for each copula model. The user must select a copula model for each copula scenario. By default, HEC-Neptune populates the Selected Copula with the copula with the lowest AIC. The user can change this selection.

The dependence characteristics of the copula models are described in the table below. Kendall's τ in the third column is the value implied by each copula's fitted dependence parameter; the empirical τ̂ reported for the data pair is a separate rank statistic (Kendall's tau-b, tie-corrected), so the two need not agree exactly. The final column gives k, the number of estimated parameters used in the AIC equations above.
| Copula | Dependence feature | Kendall's τ | τ attainable | Lower tail λL | Upper tail λU | k |
|---|---|---|---|---|---|---|
| Gaussian | Symmetric; no tail dependence | τ = (2/π)·asin(ρ) | −1 < τ < 1 | 0 | 0 | 1 |
| Student's t | Symmetric upper and lower tail dependence | τ = (2/π)·asin(ρ) | −1 < τ < 1 | λU | 2·Tν+1(−√[(ν+1)(1−ρ)/(1+ρ)]) | 2 |
| Gumbel | Upper tail | τ = 1 − 1/δ | 0 ≤ τ < 1 | 0 | 2 − 21/δ | 1 |
| Clayton | Lower tail | τ = δ/(δ + 2) | 0 < τ < 1 | 2−1/δ | 0 | 1 |
| Frank | Symmetric; no tail dependence | τ = 1 + 4·[D₁(δ) − 1]/δ | −1 < τ < 1 | 0 | 0 | 1 |
| Joe | Upper tail | τ = 1 + [2/(2 − δ)]·{ψ(2) − ψ[2/δ + 1]} | 0 ≤ τ < 1 | 0 | 2 − 21/δ | 1 |
Notes
- ρ is the dependence parameter of the elliptical copulas (Gaussian, Student's t), and δ is the dependence parameter of the Archimedean copulas (Gumbel, Clayton, Frank, Joe). Kendall's τ is the only column whose values are comparable across rows. The parameter values are not: a Joe δ of 2.2 and a Frank δ of 4.3 can describe data pairs with the same strength of association, because the two families scale their parameters differently.
- Only the Gaussian, Student's t and Frank copulas can represent negatively associated variables (τ < 0). The Gumbel, Clayton and Joe forms are restricted to positive dependence, so for a pair such as peak flow against antecedent precipitation those three are not available choices regardless of their AIC.
- ψ is the digamma function. Joe's expression has a removable singularity at δ = 2, where τ = 1 − ψ₁(2) = 2 − π²/6 ≈ 0.355, with ψ₁ the trigamma function.
- D₁ is Debye's function of order 1, D₁(x) = (1/x)·∫₀ˣ t/(eᵗ − 1) dt.
- Tν+1 is the cumulative distribution function of Student's t with ν + 1 degrees of freedom.
- λL and λU are the lower and upper tail-dependence coefficients: the probability that one variable is extreme given that the other is extreme in the same tail. A nonzero λU is the reason a joint analysis can differ from designing each variable separately; it raises the likelihood that both are extreme together.
- τ depends on ρ alone for both elliptical copulas, and the degrees-of-freedom parameter ν does not enter it. A Gaussian and a Student's t copula with the same τ describe equally strong association, so choosing between them rests on the AIC ranking rather than on τ. For the same reason τ-inversion cannot determine ν, and the Student's t copula is fitted by maximum likelihood.