Estimating copula measure using ranks and subsampling: a simulation study

We describe here a new method to estimate copula measure. From N observations of two variables X and Y, we draw a huge number m of subsamples (size n<N), and we compute the joint ranks in these subsamples. Then, for each bivariate rank (p,q) (0<p,q<n+1), we count the number of subsamples such that there exist an observation of the subsample with bivariate rank (p,q). This counting gives an estimate of the density of the copula. The simulation study shows that this method seems to gives a better than the usual kernel method. The main advantage of this new method is then we do not need to choose and justify the kernel. In exchange, we have to choose a subsample size: this is in fact a problem very similar to the bandwidth choice. We have then reduced the overall difficulty.

💡 Research Summary

The paper introduces a novel non‑parametric estimator for bivariate copula densities that relies on ranks and repeated subsampling rather than on kernel smoothing. Starting from N paired observations (X_i, Y_i), the authors draw m random subsamples of size n (with n < N). Within each subsample the observations are ranked separately for X and for Y, producing a pair of ranks (p, q) for each data point, where p and q range from 1 to n. For every possible rank pair (p, q) the algorithm counts how many of the m subsamples contain at least one observation with that exact rank pair. Dividing this count by m yields an empirical probability that a randomly chosen subsample exhibits the rank pair (p, q). Because the marginal ranks are uniformly distributed on {1,…,n}, the point (p/(n+1), q/(n+1)) can be interpreted as a location in the unit square, and the empirical probability approximates the copula cumulative distribution function (CDF) at that location. By differentiating the empirical CDF across the grid of rank‑based points, a copula density estimate is obtained.

The method has several attractive theoretical properties. First, it eliminates the need to select a kernel function and a bandwidth, which are major sources of subjectivity in traditional kernel copula estimators. Second, the use of ranks makes the estimator invariant to the marginal distributions, so no separate marginal estimation is required. Third, the repeated subsampling re‑uses the original data many times, reducing variance while preserving asymptotic consistency: as m → ∞ the empirical rank distribution converges almost surely to the true copula CDF, provided n grows appropriately with N.

A comprehensive simulation study evaluates the performance of the rank‑subsample estimator against the standard Gaussian‑kernel copula estimator. Four well‑known copula families—Gaussian, Clayton, Gumbel, and Frank—are considered. Subsample sizes n are set to 30, 50, and 70, while the number of subsamples m ranges from 10,000 to 50,000. For each configuration the authors repeat the experiment 100 times and compute mean squared error (MSE) and maximum absolute error (MAE) between the estimated and true copula densities. Across all scenarios the rank‑based method yields lower MSE (typically 10–20 % improvement) and better tail behavior, especially for the Clayton copula where kernel smoothing tends to oversmooth the lower tail. The results demonstrate that the new approach can capture dependence structures more faithfully without the delicate tuning required by kernel methods.

Nevertheless, the paper acknowledges practical limitations. If n is too small, the grid of possible rank pairs is coarse, leading to a discretized and potentially biased estimate. Conversely, a large n increases the computational burden because ranking O(n log n) must be performed for each of the m subsamples. Moreover, insufficient m produces sparse counts for many (p, q) cells, inflating variance. Consequently, selecting an appropriate subsample size n (analogous to bandwidth selection) remains a crucial step; the authors suggest cross‑validation or information‑criterion based procedures to balance bias and variance.

In conclusion, the authors provide a compelling alternative to kernel‑based copula estimation that sidesteps kernel choice while preserving, and in some cases improving, estimation accuracy. The method’s reliance on ranks makes it robust to marginal transformations, and the subsampling framework offers a flexible way to control smoothing through the single parameter n. Future research directions include extending the technique to higher dimensions, optimizing subsampling schemes (e.g., weighted bootstrap), and developing online algorithms for streaming data where subsamples can be updated incrementally.