Correlate normal scores of marginal ranks, reproduce a four-row example, and distinguish monotone invariance from a Gaussian copula assumption.
Gaussian rank correlation
Gaussian rank correlation is the Pearson correlation of normal scores obtained from each variable’s marginal ranks. For a quant comparing return or signal series, it measures association after removing the original marginal scales. Making an already-largest observation more extreme leaves its rank unchanged, so its magnitude cannot directly dominate the calculation as it can with raw Pearson correlation.
The word Gaussian describes the score transformation. It does not certify that the original observations, or their joint distribution, are Gaussian. The normal-score definition is developed in Boudt, Cornelissen and Croux, §1.
A convention you can reproduce
Start with n≥2 matched, finite pairs (xi,yi). Remove any incomplete pair before ranking either column. Assign ranks separately within the retained sample, with rank 1 for the smallest value. Here ties receive the average of their occupied ranks; tied values receive identical scores, without random tie-breaking.
This article uses midpoint plotting positions and the standard normal distribution quantile function Φ−1:
The smallest and largest possible plotting positions are 1/(2n) and 1−1/(2n), so every score is finite. Using Ri/n instead would put the maximum at 1 and produce an infinite normal quantile. Other interior conventions exist: the Boudt–Cornelissen–Croux manuscript uses Ri/(n+1). These are different finite-sample conventions; the worked number below specifically uses the midpoint rule.
A common divisor n or n−1 in the covariance and both variances cancels. If either column is constant, the denominator vanishes and the coefficient is undefined. With ties, retain the centring and actual score variances: do not assume the score mean is zero or substitute an untied-sample normalising constant. Extensive ties also mean the transformed marginal sample no longer has the untied normal-score pattern.
Four rows: 0.85748 rather than 0.8
Consider four fictional paired observations, specified by their ranks. No market calibration or sampling model is attached to this arithmetic example.
Row
RX
RY
z
w
A
1
1
−1.15035
−1.15035
B
2
3
−0.31864
0.31864
C
3
2
0.31864
−0.31864
D
4
4
1.15035
1.15035
The plotting positions for ranks 1 through 4 are 0.125,0.375,0.625,0.875. Let a=Φ−1(0.875) and b=Φ−1(0.625). Both score means are zero, so the calculation simplifies to
ρG=2a2+2b22a2−2b2≈0.85748376.
Spearman’s rank correlation instead applies Pearson correlation directly to the two rank columns and gives 0.8. The two middle rows disagree in their ordering, while the extreme rows agree. Normal quantiles space the outer scores farther apart than the inner ones, giving those agreeing extreme rows more weight in this example. The difference is a choice of score system, not a discovery of additional observations. All calculations use unrounded scores.
Four synthetic rows have X ranks 1, 2, 3, 4 and Y ranks 1, 3, 2, 4. Midpoint normal scores are −1.15035, −0.31864, 0.31864 and 1.15035. Pearson correlation of the scores is 0.85748; Spearman correlation of the ranks is 0.80000.
Original synthetic four-row calculation. Normal-score spacing changes the coefficient even though both methods use the same marginal ranks.
What survives the transformation
A strictly increasing transformation applied separately to either column preserves its ranks, ties and coefficient. This is why a change of marginal scale can leave Gaussian rank correlation unchanged. It does not protect against observations being mispaired, missing data being selected differently, or contaminated observations changing ranks. The estimator’s population influence function is unbounded; rank-based construction should not be read as immunity to outliers. See Boudt, Cornelissen and Croux, §3.
For continuous population margins, the corresponding normal-score dependence functional is
ρG=Corr(Φ−1(FX(X)),Φ−1(FY(Y))).
Under an assumed Gaussian copula model, the two population scores are jointly normal and this functional equals their latent Gaussian correlation parameter. Arbitrary continuous margins can then coexist with that Gaussian dependence model. The separation between margins and a chosen copula family is explicit in Hoff, Niu and Wellner, §1.
Rank invariance itself does not validate that family. A zero normal-score correlation does not establish independence for an arbitrary copula, and a four-row sample coefficient cannot establish a population parameter or joint-tail behaviour. No significance test or p-value is implied by this example.
A matrix built by applying the same transformation columnwise on one common complete-case sample, then taking its ordinary Pearson correlation matrix, is positive semidefinite. That computational advantage does not make it a covariance matrix of the original returns: marginal scales and a model connecting score dependence to return covariance require separate treatment.
To revisit the centring step and the distinction between zero covariance and independence, read Covariance and Correlation (public preview; full lesson requires membership).
Sources
Boudt, K., Cornelissen, J. and Croux, C. (2012), The Gaussian rank correlation estimator: robustness properties, Statistics and Computing 22, 471–483. Author manuscript: §1, definition and R/(n+1) convention; §3, sensitivity and population functional.
Hoff, P. D., Niu, X. and Wellner, J. A. (2014), Information bounds for Gaussian copulas, Bernoulli 20(2), 604–622, §1: continuous margins, Gaussian copula models and rank invariance.
