Why Copula Goodness-of-Fit Needs Its Own Diagnostics
A fitted dependence model can look plausible in a likelihood calculation and still miss important features of the observed joint behavior. For example, it may understate joint lower-tail observations or place too much probability near the center. Copula goodness-of-fit diagnostics ask whether the dependence structure implied by the fitted copula is compatible with the observed ranks, while separating that question from the marginal distributions.
For continuous random variables \(X_1,\ldots,X_d\) with marginal distribution functions \(F_1,\ldots,F_d\), the probability integral transforms \(U_j=F_j(X_j)\) have uniform marginal distributions. Their joint distribution is the copula. In practice, the margins may be estimated, so rank-based pseudo-observations are commonly used to examine dependence without requiring a fully specified marginal model.
The denominator \(n+1\) keeps pseudo-observations strictly inside the unit interval, which is useful for copula likelihood calculations. A different common empirical convention uses ranks divided by \(n\); whichever convention is chosen, it must be used consistently in fitting and diagnostics. With ties, the rank convention must also be stated—for example, average ranks or randomized ranks—because ties can change the empirical comparison.
Empirical Versus Fitted Copula
After estimating a copula parameter \(\hat\theta\), compare the empirical copula \(C_n\) with the fitted copula \(C_{\hat\theta}\). At a point \(\mathbf{u}\), the discrepancy is \(C_n(\mathbf{u})-C_{\hat\theta}(\mathbf{u})\). A positive discrepancy means the sample has more observations in the lower rectangle \([0,u_1]\times\cdots\times[0,u_d]\) than the fitted model predicts; a negative discrepancy means it has fewer.
A plot over a grid of threshold values can reveal systematic departures. In two dimensions, one can display the difference \(C_n(u,v)-C_{\hat\theta}(u,v)\) as a surface or heat map. Patterns concentrated near small \(u,v\) may signal a mismatch in joint lower-tail behavior; patterns near large \(u,v\) can indicate a mismatch in joint upper-tail behavior. A single point is only a local diagnostic: it cannot summarize fit over the whole unit square.
A formal discrepancy statistic can aggregate differences across a grid or across sample pseudo-observations. For example, a Cramér–von Mises-type statistic has the general form $$ T_n=\sum_{k=1}^K w_k\{C_n(\mathbf{u}_k)-C_{\hat\theta}(\mathbf{u}_k)\}^2, $$ where \(\mathbf{u}_1,\ldots,\mathbf{u}_K\) are chosen evaluation points and \(w_k\ge 0\) are specified weights. Larger values indicate greater discrepancy under this definition. The statistic alone does not supply a universal cutoff: its null distribution depends on the sample size, the copula family, the parameter-estimation method, and the chosen grid or weights.
Worked Example: Reading a Local Empirical-Copula Discrepancy
A hypothetical insurer has four paired loss observations. Their ranks in the first and second margins are \((1,1),(2,3),(3,2),(4,4)\). The fitted dependence model is a Clayton copula with parameter \(\hat\theta=1\). Compare the empirical and fitted copulas at \((u,v)=(0.5,0.5)\).
Method. Convert the ranks to pseudo-observations using \(r/(n+1)=r/5\). The pairs are \((0.2,0.2),(0.4,0.6),(0.6,0.4),(0.8,0.8)\).
The empirical copula at \((0.5,0.5)\) counts pairs with both components at most \(0.5\). Only \((0.2,0.2)\) qualifies, so $$ C_n(0.5,0.5)=\frac{1}{4}=0.25. $$
For a Clayton copula with \(\theta=1\), \(C(u,v)=(u^{-1}+v^{-1}-1)^{-1}\). Thus $$ C_{\hat\theta}(0.5,0.5)=(2+2-1)^{-1}=\frac{1}{3}\approx 0.3333. $$
The signed discrepancy is \(0.25-0.3333\approx-0.0833\). At this threshold, the sample has less joint lower-rectangle probability than the fitted model. The squared discrepancy is approximately \((-0.0833)^2=0.0069\), rounding only this final squared value to four decimals. This calculation is a local diagnostic, not a goodness-of-fit conclusion by itself.
Common mistake avoided: Counting observations with either component below the threshold would give a different quantity. The empirical copula counts observations for which both component conditions hold.
Copula Probability Integral Transforms
A marginal probability integral transform checks a marginal model, not the copula. Even if two variables have strong dependence, each correctly transformed continuous margin is uniform. Therefore, separately plotting or testing \(F_1(X_1)\) and \(F_2(X_2)\) for uniformity cannot establish that their joint dependence is correctly modeled.
For a bivariate copula \(C\), a dependence diagnostic is the Rosenblatt transform. For continuous distributions, define $$ W_1=U_1,\qquad W_2=F_{U_2\mid U_1}(U_2\mid U_1) =\frac{\partial C(U_1,U_2)}{\partial u_1}. $$ The derivative is the conditional distribution function of \(U_2\) given \(U_1=u_1\), evaluated at \(u_2\). Under a correctly specified, known, continuous copula, \(W_1\) and \(W_2\) are independent uniform random variables. Consequently, plots and tests can examine both the uniformity of transformed values and their remaining dependence. If parameters are estimated from the same sample, that estimation affects diagnostic behavior and should be reflected in formal calibration.
Worked Example: Calculating Rosenblatt Values for a Clayton Copula
For two hypothetical loss observations, the margin-transformed pairs are \((u,v)=(0.5,0.25)\) and \((0.25,0.5)\). Use a fitted Clayton copula with \(\hat\theta=1\) to calculate their Rosenblatt transforms.
Method. For \(\theta=1\), \(C(u,v)=(u^{-1}+v^{-1}-1)^{-1}\). Differentiating with respect to \(u\) gives $$ \frac{\partial C(u,v)}{\partial u} =u^{-2}(u^{-1}+v^{-1}-1)^{-2}. $$ For the first pair, \(W_1=0.5\), and $$ W_2=\frac{(0.5)^{-2}}{\{(0.5)^{-1}+(0.25)^{-1}-1\}^2} =\frac{4}{(2+4-1)^2} =\frac{4}{25}=0.16. $$ For the second pair, \(W_1=0.25\), and $$ W_2=\frac{(0.25)^{-2}}{\{(0.25)^{-1}+(0.5)^{-1}-1\}^2} =\frac{16}{(4+2-1)^2} =\frac{16}{25}=0.64. $$
The transformed pairs are \((0.5,0.16)\) and \((0.25,0.64)\). These two values illustrate the calculation, but a sample of two cannot provide a useful assessment of uniformity or independence. With a larger sample, inspect the transformed values graphically and use a suitable calibrated test if a formal conclusion is needed.
Common mistake avoided: Using the marginal transform \(F_2(X_2)\) alone as the second Rosenblatt component would simply return \(U_2\), not its conditional distribution given \(U_1\). That misses the dependence structure the diagnostic is intended to assess.
Parametric-Bootstrap Goodness-of-Fit
A parametric bootstrap approximates the null distribution of a goodness-of-fit statistic under the fitted copula. Its key requirement is to repeat the fitting procedure in every bootstrap sample. Comparing the observed statistic with simulated statistics from a fixed, known parameter would generally omit the variability introduced by estimating the parameter from data.
Estimate the copula parameter using the same method used for the analysis, such as pseudo-likelihood based on rank pseudo-observations.
Compute a specified discrepancy \(T_{\mathrm{obs}}\), such as a weighted sum of squared empirical-versus-fitted copula differences.
For each bootstrap replicate, simulate a sample of size \(n\) from the fitted null model, carry out the same data transformation and estimation procedure, and calculate its statistic \(T_b^*\).
With \(B\) replicates, use \(\hat p=\{1+\sum_{b=1}^B I(T_b^*\ge T_{\mathrm{obs}})\}/(B+1)\). The added 1 avoids a reported zero p-value when the number of replicates is finite.
If margins have been estimated as part of the procedure, a bootstrap intended to assess the full fitted model should reproduce that estimation too. If the diagnostic is specifically conditional on fixed, known margins, the bootstrap should instead preserve that assumption. The simulation and refitting steps must match the null hypothesis and the analysis pipeline; otherwise the reference distribution may not answer the intended question.
Worked Example: Interpreting a Bootstrap Goodness-of-Fit Test
A hypothetical risk analyst compares a fitted copula with empirical values at two grid points. The observed signed differences are \(0.08\) and \(-0.06\), and the chosen statistic is the sum of their squared differences with equal weights. Four parametric-bootstrap replicates, each simulated from the fitted copula and refitted, produce statistics \(0.004, 0.009, 0.013,\) and \(0.021\). Calculate the observed statistic and the bootstrap p-value.
Method. Square each observed discrepancy and add: $$ T_{\mathrm{obs}}=(0.08)^2+(-0.06)^2=0.0064+0.0036=0.0100. $$ The bootstrap statistics at least as large as \(0.0100\) are \(0.013\) and \(0.021\), so the exceedance count is 2. With \(B=4\), $$ \hat p=\frac{1+2}{4+1}=\frac{3}{5}=0.6000. $$
The estimated p-value is \(0.6000\), rounded to four decimal places. In this small illustration, the observed discrepancy is not unusual relative to the bootstrap statistics. Four replicates are far too few for a precise p-value in practical work; many more would ordinarily be used. A large p-value is not proof that the copula is correct. It means this test, with this sample and statistic, has not found strong evidence against the fitted model.
Common mistake avoided: Counting only bootstrap statistics strictly greater than the observed value, or dividing the exceedance count by \(B\) without the finite-simulation adjustment, changes the stated calculation. Here the specified rule counts values greater than or equal to \(T_{\mathrm{obs}}\) and uses the plus-one estimate.
Interpreting Diagnostics in Risk Work
No single diagnostic detects every form of misspecification. Empirical-versus-fitted comparisons can show where the model departs from observed joint probabilities, while Rosenblatt values can reveal non-uniformity or remaining dependence after conditioning. A global bootstrap statistic summarizes a chosen set of discrepancies, but its result depends on that choice. Use graphical and numerical checks together, and interpret results in light of the risk feature being modeled.
Tail diagnostics require particular care. With a small sample, few observations may fall in an extreme joint tail, so empirical copula values there are noisy. A visually large tail discrepancy may be driven by one observation; conversely, limited data may make a consequential departure difficult to detect. Report which region or statistic was examined and avoid treating a non-rejection as confirmation of tail behavior.
These methods provide a disciplined way to assess a fitted dependence structure before using it in aggregate-risk calculations. Their purpose is diagnostic: they identify potential incompatibility between data and model, but do not automatically select a replacement family or guarantee that future dependence will match historical observations. This treatment is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.
Check Your Understanding
Work each problem before checking its final answer.
- A sample of five paired observations has rank pairs \((1,2),(2,1),(3,4),(4,3),(5,5)\). Using pseudo-observations \(r/(n+1)\), calculate the empirical copula at \((0.5,0.5)\). (Answer: \(1/5=0.2000\))
- At a chosen grid point, an empirical copula is \(0.42\) and the fitted copula is \(0.35\). State the signed discrepancy and interpret its direction. (Answer: \(0.07\); the sample has more joint lower-rectangle probability than the fitted model predicts.)
- For a Clayton copula with \(\theta=1\), calculate \(W_2\) for \((u,v)=(0.5,0.5)\) using \(W_2=u^{-2}(u^{-1}+v^{-1}-1)^{-2}\). (Answer: \(4/9\approx0.4444\))
- A goodness-of-fit statistic is \(0.018\). Of 99 bootstrap statistics, 7 are at least \(0.018\). Calculate the plus-one bootstrap p-value to four decimals. (Answer: \((1+7)/(99+1)=0.0800\))
- An analyst reports that both marginal PIT samples appear uniform and concludes that the fitted copula is adequate. Identify the logical gap and name a dependence diagnostic that addresses it. (Answer: Uniform marginal PITs do not assess the joint dependence; compare empirical and fitted copulas or examine bivariate Rosenblatt transforms for uniformity and independence.)