Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
62,590 characters · 24 sections · 14 citation commands
The Generalized Fisher Transformation: Finite-Sample Properties and Inference
\noindentKeywords: correlation matrices; Fisher transformation; matrix logarithm; finite-sample inference; variance stabilization; realized correlation.
\noindentJEL Classification: C12, C13, C32, C58.
Correlation matrices are central to multivariate analysis in economics and finance, yet statistical inference for these matrices remains notoriously difficult. The elements of the standard sample correlation matrix, $\hat{C}$, are bounded in $[-1,1]$ and exhibit complex dependencies, even when the underlying variables are independent. Fisher:1915 resolved these issues for the bivariate case ($n=2$) with his celebrated $z$-transformation, which stabilizes variance and reduces skewness under appropriate conditions. Hotelling:1953 emphasized that this simultaneous variance stabilization and approximate normalization is a special feature of Fisher's transformation, rather than a generic property of transformations of statistics. However, a multivariate generalization that preserves these desirable properties for correlation matrices of dimension $n>2$ has remained elusive.
In this paper, we analyze the Generalized Fisher Transformation (GFT), $\gamma(C)=\operatorname{vecl}(\log C)$, which was introduced by ArchakovHansen:Correlation. The GFT maps the manifold of positive definite correlation matrices to the Euclidean space $\mathbb{R}^{d}$, where $d=n(n-1)/2$. This is achieved by taking the matrix logarithm of $C$ and vectorizing the below-diagonal elements, as illustrated in the following example for $n=3$: $$C=\left[
\right], \quad \log C = \left[
\right], \quad \gamma = \operatorname{vecl}\log C = \left[
\right].$$ This is a generalization of the Fisher transformation, $\phi=\tfrac{1}{2}\log\tfrac{1+\rho}{1-\rho}$, to correlation matrices, in the sense that $\gamma(C)$ is identical to the Fisher transformation for a $2\times2$ correlation matrix, specifically $$ C=\left[
\right]\quad\Rightarrow\quad\log C=\left[
\right]\quadfor |\rho|<1. $$ For $n \geq 3$, the elements of $\gamma$ do not coincide with the element-wise Fisher transformed correlations. Consequently, $\hat\gamma=\gamma(\hat{C})$ has distinct statistical properties.
The intuition behind the GFT is illustrated in Figure (ref). To generate this figure, we replicated the design used by Fisher:1921 to motivate his original transformation. The design uses just $T=8$ observations and an equicorrelation structure with either $\rho=0$ or $\rho=0.8$. The first two panels in the first row reproduce Fisher:1921. The left panels display the marginal and joint distributions of two standard sample correlation coefficients, $(\hat{\varrho}_{1},\hat{\varrho}_{2})$, which correspond to $(\hat{\rho}_{12},\hat{\rho}_{13})$ in this three-dimensional design. For $\rho=0.8$, the distribution is heavily skewed and exhibits strong dependence, as seen in the lower-left panel. The middle column displays the corresponding Fisher transformed correlations, $(\hat{\phi}_{1}, \hat{\phi}_{2})$. The marginal distributions are close to normal with unit variance, but the strong dependence between the two elements persists for $\rho=0.8$. The right panels display the corresponding GFT coordinates, $(\hat{\gamma}_{1},\hat{\gamma}_{2})$. Not only are the marginal distributions approximately Gaussian, the contours of the bivariate distribution are remarkably spherical for both $\rho=0$ and $\rho=0.8$. Thus, the GFT maps the complex geometry of the correlation cone to coordinates in which the estimates are approximately uncorrelated.
While ArchakovHansen:Correlation established the theoretical bijection and asymptotic results for $\gamma(C)$, the finite-sample statistical properties have largely remained unexplored for general correlation structures and non-Gaussian data. This paper documents three main finite-sample properties of the GFT and shows how they translate into improved inference for correlation matrices.
First, we show that the finite-sample distributions of the elements of $\hat{\gamma}$ are well approximated by their Gaussian asymptotic distributions for elliptically distributed data. This mirrors the behavior of the univariate Fisher transformation and extends an important scalar property to correlation matrices of higher dimension. However, as with the Fisher transformed correlation, the Gaussian approximation can become unreliable in small samples when the data are strongly skewed or otherwise depart from elliptical symmetry.
Second, we document the near orthogonality of the GFT coordinates. The finite-sample covariance matrix $V_{\gamma,T}(C)=\operatorname{var}{\sqrt{T}(\hat{\gamma}-\gamma)}$ is approximately diagonal across a wide range of designs. This contrasts sharply with the strong dependence found in both $\hat\varrho$ and $\hat\phi$, where $\hat\varrho=\varrho(\hat{C})$ and $\hat\phi=\phi(\hat{C})$ denote the vector of sample correlations and the vector of element-wise Fisher transformed sample correlations, respectively.
Third, we document that $V_{\gamma,T}(C)$ is far less sensitive to the true correlation matrix, $C$, than the corresponding covariance matrices $V_{\varrho,T}(C)$ and $V_{\phi,T}(C)$. This near invariance has important implications for inference. Since $V_{\gamma}(C)$ is relatively stable across values of $C$, the plug-in covariance estimator $V_{\gamma}(\hat{C})$ is much less affected by estimation error in $\hat{C}$ than the corresponding plug-in estimators for $\hat\varrho$ and $\hat\phi$. As a result, standardized statistics based on $\hat{\gamma}$ have substantially better finite-sample behavior, and Wald tests based on the GFT converge much faster to their nominal size.
The remainder of the paper is organized as follows. Section (ref) introduces the notation, simulation designs, and empirical data sets. Section (ref) documents the marginal and joint finite-sample properties of $\hat{\varrho}$, $\hat{\phi}$, and $\hat{\gamma}$ using Gaussian simulations, non-Gaussian simulations, and empirical resampling designs. Section (ref) gives theoretical results that explain the covariance stability and weak dependence of the GFT coordinates. Section (ref) studies the implications for inference and shows that GFT-based standardized statistics have substantially better finite-sample behavior. Section (ref) concludes. The Appendix contains proofs, and the Supplement reports additional simulation and empirical results.
We present results for the sample correlation matrix, $\hat{C}$, associated with the sample covariance matrix, $$\hat{\Sigma}=\tfrac{1}{T}\sum_{t=1}^{T}(X_{t}-\bar{X})(X_{t}-\bar{X})^{\prime}, \qquad \bar{X}=\tfrac{1}{T}\sum_{t=1}^{T}X_{t}.$$ Let $\mu=\mathbb{E}[X_{t}]$ and $\Sigma=\operatorname{var}(X_{t})$. In our simulations, we can, without loss of generality, take $\mu=0$ and $\Sigma=C$, because $Y=DX+b$ has the same correlation matrix as $X$ for any non-singular diagonal matrix, $D$, and any vector $b$, and the empirical correlation matrix, $\hat{C}$, retains this invariance. We present results for the case where $X_{t}$ has a multivariate normal distribution and cases where it follows non-Gaussian distributions, including the uniform, multivariate $t$, and Inverse Gaussian distributions. We also simulate by drawing from an empirical distribution of daily industry returns.
As introduced earlier, we let $\varrho=\operatorname{vecl}(C)$ and $\phi=\phi(C)$ denote the vectors of correlations and the corresponding vector of Fisher transformed correlations, respectively. Similarly, $\hat{\varrho}=\varrho(\hat{C})$, $\hat{\phi}=\phi(\hat{C})$, and $\hat{\gamma}=\gamma(\hat{C})$ denote the empirical vectors of correlation measures, where $\hat{C}$ is the sample correlation matrix. Their asymptotic distributions are, under suitable regularity conditions, given by
respectively. General expressions for the asymptotic covariance matrices, $V_{\varrho}(C)=\operatorname{avar}(\hat{\varrho})$, $V_{\phi}(C)=\operatorname{avar}(\hat{\phi})$, and $V_{\gamma}(C)=\operatorname{avar}(\hat{\gamma})$, are given in Appendix (ref), along with simplified expressions for the special case where $X_{t}$ is iid and normally distributed. We use the correlation matrix, $C$, as an argument in the expressions for the asymptotic covariance matrices to make explicit that they depend on the true correlation matrix. The corresponding finite-sample variance-covariance matrices are denoted by $$ V_{\varrho,T}(C)=\operatorname{var}(\sqrt{T}(\hat{\varrho}-\varrho)),\quad V_{\phi,T}(C)=\operatorname{var}(\sqrt{T}(\hat{\phi}-\phi)),\quad V_{\gamma,T}(C)=\operatorname{var}(\sqrt{T}(\hat{\gamma}-\gamma)), $$ respectively, where $T$ is the sample size. Similarly, we define the correlation matrices, $$ R_{\varrho,T}(C)=\operatorname{corr}(\sqrt{T}(\hat{\varrho}-\varrho)),\quad R_{\phi,T}(C)=\operatorname{corr}(\sqrt{T}(\hat{\phi}-\phi)),\quad R_{\gamma,T}(C)=\operatorname{corr}(\sqrt{T}(\hat{\gamma}-\gamma)), $$ and their asymptotic counterparts are denoted by $R_{\varrho}(C)$, $R_{\phi}(C)$, and $R_{\gamma}(C)$, respectively.
Before we present results on the finite-sample properties of $\hat{\varrho}$, $\hat{\phi}$, and $\hat{\gamma}$, it will be useful to illustrate the GFT coordinates with three empirical data sets. These provide realistic correlation structures for the simulation designs in Section (ref).
The three data sets are: (1) the macroeconomic time-series from McCrackenNg:2016 (monthly data from 1953 to 2022); (2) Daily returns for 30 Fama-French industry portfolios downloaded from Kenneth French's website (January 2, 2018 to December 31, 2019); and (3) Realized correlation matrix for 21 assets based on five-minute intraday returns from opening hours 9:30 AM to 4:00 PM during the week, August 22-26, 2016.
We present the empirical correlation matrix for these three data sets in Figure (ref) (left panels) along with the logarithmically transformed correlation matrix (middle column). The right panels are scatter plots of the elements of $\phi(\hat{C})$ plotted against the corresponding elements of $\gamma(\hat{C})$. These reveal a relationship between $\hat{\gamma}$ and $\hat{\phi}$, which also implies a relationship between $\hat{\gamma}$ and $\hat{\varrho}$. Thus, even though the matrix logarithm is a highly nonlinear mapping, where each element of $\log\hat{C}$ depends on all sample correlation coefficients in $\hat{C}$, it is evident that $\hat{\gamma}_{i}$ is primarily influenced by the corresponding sample correlation element, $\hat{\varrho}_{i}$. This is not surprising because the Jacobian, $\partial\varrho/\partial\gamma$, is typically dominated by its diagonal elements.
Detailed information about elements of $\hat{C}$ and $\log\hat{C}$ will be presented in Section (ref) for the industry portfolios along with an extended time series of the realized correlations. Additional empirical results are presented in the Supplementary Material.
The asymptotic distributional theory for $\hat{\gamma}$, $\hat{\varrho}$, and $\hat{\phi}$ is given in ArchakovHansen:Correlation. The goal of this section is to document their finite-sample behavior, with particular emphasis on how the GFT coordinates $\hat{\gamma}$ compare to the conventional coordinates $\hat{\varrho}$ and $\hat{\phi}$.
We first examine variance, skewness, and kurtosis of the marginal distributions, using these moments to assess variance stabilization and the quality of the Gaussian approximation in finite samples. We then study joint behavior through the finite-sample correlation matrices $R_{\varrho,T}(C)$, $R_{\phi,T}(C)$, and $R_{\gamma,T}(C)$ of the corresponding coordinate vectors. The first two are nearly identical, whereas $R_{\gamma,T}(C)$ is typically much closer to the identity matrix.
We first consider finite-sample properties when $\hat{C}$ is computed from $X_{t} \sim iid N_{n}(0,C)$, $t=1,\ldots,T$, where $C$ has the following Toeplitz structure,
and we use $\rho=0.9$ and $T=100$ in this section. Thus, when $n=25$, the correlation coefficients range from about $0.08$ to $0.90$.
The upper panels of Figure (ref) are based on dimension $n=25$ and present the skewness and kurtosis for the marginal distributions of the $d=300$ elements of $\hat{\varrho}$, $\hat{\phi}$, and $\hat{\gamma}$. We have plotted excess kurtosis against skewness, where the color intensity of each dot indicates the variance of the corresponding element. Light colors correspond to small, near-zero variances, and dark colors correspond to variances close to one. For reference, the standard normal distribution, $N(0,1)$, has zero skewness, zero excess kurtosis, and unit variance.
The finite-sample distributions of the sample correlations, $\hat{\varrho}$, are very non-Gaussian, as is evident from their skewness and excess kurtosis. Moreover, there is large variation in their variances because these depend on the population correlations. The elements of $\hat{\phi}$ are the Fisher transformed correlations, and each is well approximated by its asymptotic distribution, which is the standard normal distribution, $N(0,1)$. Moreover, the transformation successfully stabilizes their variances, as expected. The corresponding results for $\hat{\gamma}$ are presented in the upper right panel of Figure (ref). Interestingly, the marginal distributions of the elements of $\hat{\gamma}$ are similar to those of $\hat{\phi}$. Skewness and excess kurtosis are both relatively close to zero, and the variances are approximately stabilized. However, the asymptotic marginal distributions for the elements of $\hat{\gamma}$ are not exactly $N(0,1)$, and the diagonal elements of the asymptotic covariance matrix, $V_{\gamma}(C)$, range from 0.865 to 1.005 in this design.\footnote{The majority of the variances are within a narrow band, with 80% of them falling between 0.911 and 0.966. The asymptotic variances are given by the diagonal elements of $V_{\gamma}(C)$; see ArchakovHansen:Correlation for the expression.}
A surprising property of $\hat{\gamma}$ emerges from the lower panels of Figure (ref). These are based on $n=10$ (i.e., $d=45$) to ensure legibility of the correlation matrices, and display the $45\times45$ correlation matrices, $R_{\varrho,T}(C)$, $R_{\phi,T}(C)$, and $R_{\gamma,T}(C)$, using color codes. The correlation matrices for $\hat{\varrho}$ and $\hat{\phi}$ are virtually identical, which is a consequence of the delta method and the fact that $\hat{\phi}$ is defined by an element-wise transformation of $\hat{\varrho}$. The surprising result in Figure (ref) is the weak dependence between the elements of $\hat{\gamma}$. Not only does the mapping $\gamma(C)$ produce coordinates whose finite-sample distributions are well approximated by normal distributions when the underlying data are normally distributed, the elements are also nearly uncorrelated. The latter is an unexpected property of the GFT coordinates. The weak dependence appears to be a robust property of $\hat{\gamma}$, whereas the near-Gaussian finite-sample distributions of the elements of $\hat{\phi}$ and $\hat{\gamma}$ are more fragile, in the sense that they rely on the underlying data being normally distributed. We next examine whether these findings persist for correlation matrices with more general structures and for correlation matrices estimated from non-Gaussian samples.
The Toeplitz structure in ((ref)) defines a special class of correlation matrices. We now study correlation matrices with more general, arbitrary structures. We generate random correlation matrices using the method developed in ArchakovHansenLuo-RandomCorr:2024. This is straightforward, as it only requires drawing a random vector $\gamma$, followed by evaluating $C(\gamma)$. Specifically, we first draw $\omega$ uniformly on the interval $[0,\tfrac{5}{n}\log n]$, and then draw $\gamma|\omega\sim N_d(\xi\iota,\omega^2I_d)$, where $\xi=\tfrac{\log(1+n)}{n}$.\footnote{The dependence on $n$ is motivated by the fact that an equicorrelation matrix with common correlation coefficient, $\rho$, will translate to a vector, $\gamma$, with identical elements equal to $\tfrac{1}{n}\log(1+n\tfrac{\rho}{1-\rho})$.} This particular design is chosen because it generates a sufficiently interesting range of correlation matrices, including near-singular matrices. Moreover, the design is such that the probability density is positive for any valid non-singular correlation matrix.\footnote{Any distribution for $\gamma$ on $\mathbb{R}^{d}$, with $d=n(n-1)/2$, will translate to a distribution on $\mathcal{C}_{n}$, and any distribution with full support on $\mathbb{R}^{d}$ will induce a distribution with full support on $\mathcal{C}_{n}$.} Three examples of random correlation matrices generated with this design are:
{1.5pt}
We generate 1,000 random correlation matrices, and for each, we generate $N=10{,}000$ estimates of $C$. These are sample correlation matrices based on $X_{t}{\sim}iid N_n(0,C)$ for $t=1,\ldots,T$, with $T=40$, $T=100$, and $T=250$. We present results based on random correlation matrices with dimension $n=25$, which translates to $d=300$ distinct correlation coefficients in $C$, such that the correlation matrix for $\hat{\varrho}$, $R_{\varrho,T}(C)$, has 44,850 unique correlations. The same dimensions apply to the other parametrizations, $\hat{\phi}$ and $\hat{\gamma}$.
While the correlation matrices can take any value in $\mathcal{C}_{n}$, the variation in the estimators' properties across correlation matrices is largely determined by the smallest eigenvalue $\lambda_{\min}(C)=\min\{\lambda\in\sigma(C)\}$. For this reason, we present results that are stratified by $\lambda_{\min}(C)$.
We focus on the GFT coordinates' ability to stabilize variance compared to that of $\hat{\phi}$, and the dependence between elements of the three parametrizations. In the Supplementary Material, we present Q-Q plots showing that the marginal distributions of the elements of $\hat{\phi}$ and $\hat{\gamma}$ are well approximated by their asymptotic normal distributions. We also present results for additional sample sizes and for dimensions $n=5$ and $n=10$.
Next, we consider the finite-sample variance of the elements of $\sqrt{T}(\hat{\varrho}-\varrho)$, $\sqrt{T}(\hat{\phi}-\phi)$, and $\sqrt{T}(\hat{\gamma}-\gamma)$, which are given by the diagonal elements of $V_{\varrho,T}(C)$, $V_{\phi,T}(C)$, and $V_{\gamma,T}(C)$, respectively. There are $d=300$ variances for each of the 1,000 random correlation matrices. We present the smallest and largest finite-sample variances (blue dots) for each correlation matrix, plotted in the top row of Figure (ref) against the smallest eigenvalue of $C$, $\lambda_{\min}(C)$, for the sample size $T=100$. Results for other sample sizes are presented in the Supplementary Material, see Figure (ref). For comparison, we also include their asymptotic variances (red dots), obtained from $V_{\varrho}(C)$, $V_{\phi}(C)$, and $V_{\gamma}(C)$. The finite-sample variances are estimated by simulating 10,000 independent estimates for each of the 1,000 correlation matrices.
The results for $\sqrt{T}(\hat{\varrho}-\varrho)$ and $\sqrt{T}(\hat{\phi}-\phi)$ are as expected. The finite-sample variances for $\hat{\varrho}$ depend on the elements of $C$, and asymptotically they approach $(1-C_{ij}^{2})^{2}$. The Fisher transformation, shown in the middle column of panels, successfully stabilizes the variances to unity, as it is designed to do for normally distributed data. The GFT coordinates do not stabilize the marginal variances to the same degree. There is some dispersion in the variances, especially for near-singular correlation matrices, as seen for small values of $\lambda_{\min}(C)$. For $\hat{\gamma}$, we also note that the finite-sample variances are reasonably close to their asymptotic values when $T=250$ and $n=25$, and for smaller $n$ the finite-sample variances are closer to their asymptotic values. The agreement is also good for smaller sample sizes, including $T=40$ (see Supplementary Material).
Next, we turn to the correlation between the elements within each of the vectors, $\hat{\varrho}$, $\hat{\phi}$, and $\hat{\gamma}$. Rather unexpectedly, we observe substantially less dependence between the elements of $\hat{\gamma}$ than those of the other two vectors. The middle row of panels in Figure (ref) presents the largest and smallest finite-sample correlations between elements of $\hat{\varrho}$, $\hat{\phi}$, and $\hat{\gamma}$ (blue dots) and their corresponding asymptotic population values (red dots) for $T=100$. Note that $n=25$ implies $d=300$ elements in each vector, which translates to $d(d-1)/2=44,850$ distinct correlations. The shaded area in each panel covers the range for 80% of these correlations. The results are plotted against the smallest eigenvalue, $\lambda_{\min}(C)\in[10^{-4},1]$. Results for other sample sizes are presented in Figure (ref) in the Supplementary Material.
Variance stabilization in the multivariate case is better quantified by considering linear combinations of the vectors, e.g., $\operatorname{var}(a^\prime\hat{\phi})$ with $a^{\prime}a=1$. The minimum and maximum variance across normalized linear combinations are given by the smallest and largest eigenvalues of the covariance matrix, respectively. These are presented in the last row of panels in Figure (ref) for $T=100$. Results for other sample sizes are presented in the Supplementary Material, see Figure (ref). The GFT coordinates, $\gamma(C)$, are far better at stabilizing the spectrum of variances, with the range being orders of magnitude smaller than that of $\hat{\varrho}$ and $\hat{\phi}$ (note that a logarithmic scale is used for the $y$-axis in lower panels). Thus, while the Fisher transformation is excellent at standardizing the variance of individual elements of $\hat{\phi}$, it is unable to moderate covariances between elements.
In the Supplementary Material, Figure (ref), we report additional results for the correlations between elements of $\hat{\gamma}$ for the case where $C$ has a Toeplitz structure, while maintaining the Gaussian distribution, $X_{t}\sim iid N_n(0,C)$. Whenever $X_{t}$ is drawn from the Gaussian, we find that the variances of the elements in $\hat{\gamma}$ have similar values and are approximately equal to one. This appears to hold for all types of correlation matrices with eigenvalues bounded away from zero. Moreover, as $n$ increases, the results for Toeplitz correlation matrices become very similar to those obtained with random correlation matrices. These findings suggest that the weak dependence and approximate variance stabilization of $\hat{\gamma}$ are not artifacts of a particular correlation design, but rather robust finite-sample features of the GFT coordinates.
We next examine whether the weak dependence between elements of $\hat{\gamma}$ persists when the data are not Gaussian. The marginal Gaussian approximation for transformed correlations is known to be more fragile in this case. For example, skewness and kurtosis of the Fisher transformed correlations can deviate substantially from the standard normal, as illustrated in Figure (ref) in the Supplementary Material.
Figure (ref) reports results for samples drawn from uniform, Student's $t$, and Inverse Gaussian marginal distributions, using the Toeplitz correlation matrix with $n=25$ and $\rho=0.9$. Each panel shows the distribution of the 44,850 off-diagonal elements of $R_{\phi,T}(C)$ and $R_{\gamma,T}(C)$ for $T=40$, $T=100$, and $T=250$. The corresponding results for $R_{\varrho,T}(C)$ are essentially indistinguishable from those for $R_{\phi,T}(C)$. The correlations between elements of $\hat{\gamma}$ remain concentrated near zero in all three non-Gaussian designs, whereas the correlations between elements of $\hat{\phi}$ are far more dispersed. Thus, the weak dependence of the GFT coordinates appears to be much more robust than the marginal Gaussian approximation.
We next study the properties of $\hat{\gamma}$ using empirical structures from the three data sets used in Figure (ref): the FRED-MD macroeconomic data set from McCrackenNg:2016, the 30 Fama-French industry portfolios, and high-frequency returns for 21 assets, from which we compute realized correlations using five-minute returns.
We generate artificial samples by resampling, with replacement, from the empirical distributions and then apply the correlation estimators to these samples. This preserves empirically relevant correlation structures while also introducing the non-Gaussian features present in the data. Thus, the exercise complements the controlled Gaussian and non-Gaussian simulations above.
To conserve space, we present the results for two of the data sets here, the 30 Fama-French industry portfolios and the high-frequency return data. The analogous results for a subset of the macro variables are presented in the Supplementary Material.
Our data for the 30 industry portfolios consist of 503 daily returns from January 2, 2018 to December 31, 2019. The sample correlation matrix is shown below the diagonal in Table (ref), and the corresponding logarithmic correlation matrix is shown above the diagonal. The smallest eigenvalue is $\lambda_{\min}=0.0704$, so this is a realistic correlation structure that is moderately close to singularity.
We construct artificial samples in two ways. First, we simulate Gaussian return vectors with correlation matrix equal to the empirical correlation matrix, $X_{t}\sim iid N(0,\hat{C})$. Second, we resample return vectors from the empirical distribution, with replacement. The first design isolates the role of the empirical correlation structure, while the second also incorporates the non-Gaussian features of the observed returns. For each sample, we estimate its correlation matrix and compute $\hat{\varrho}$, $\hat{\phi}$, and $\hat{\gamma}$. Figure (ref) presents results for $T=100$; additional results for $T=40$ and $T=250$ are reported in the Supplementary Material.
The top panels of Figure (ref) show the variance, skewness, and kurtosis for the 435 elements of $\sqrt{T}(\hat{\varrho}-\varrho)$, $\sqrt{T}(\hat{\phi}-\phi)$, and $\sqrt{T}(\hat{\gamma}-\gamma)$ under the Gaussian design. The bottom panels show the corresponding results based on empirical resampling. The contrast is clear: under Gaussian sampling, the finite-sample distributions of $\hat{\phi}$ and $\hat{\gamma}$ are close to their asymptotic normal distributions, whereas empirical resampling produces substantial deviations from normality. The GFT coordinates nevertheless retain better finite-sample behavior than the sample correlations.
Figure (ref) provides a bivariate view of the same phenomenon. It shows the joint finite-sample distributions of selected elements involving Food Products, Apparel, and Automobiles and Trucks, computed from empirical resamples with $T=40$, $100$, and $250$. The left panels are for $\hat{\varrho}$, while the middle and right panels show the corresponding elements of $\hat{\phi}$ and $\hat{\gamma}$. The distributions for $\hat{\varrho}$ and $\hat{\phi}$ remain visibly non-elliptical, and even bimodal, whereas the corresponding distribution for $\hat{\gamma}$ is much closer to elliptical, even for $T=40$.
This subsection reports an empirical analysis of realized correlation matrices constructed directly from high-frequency TAQ data. We use the same $n=21$ assets as in Figure (ref), but extend the sample from the illustrative week of August 22--26, 2016 to the full period January 1, 2005 to December 31, 2020.
For each ofthe $K=4,027$ trading days, we compute a realized correlation matrix, $\hat{C}_{t}$, from 78 five-minute intraday returns, yielding daily time series of $\hat{\varrho}_{t}$, $\hat{\phi}_{t}$, and $\hat{\gamma}_{t}$. Intraday returns depart from the iid benchmark underlying the closed-form asymptotic variances, since serial correlation and microstructure noise alter the sampling distribution of $\hat C_t$; we therefore read the high-frequency evidence as documenting the robustness of the GFT coordinates under weak dependence rather than as a test of the iid asymptotic variances, and the qualitative findings are unchanged.
We sort days according to the average realized variance across the 21 assets. Figure (ref) presents histograms of realized correlations for the 400 lowest-volatility days and the 400 highest-volatility days. Each histogram is based on $210\times400=84,000$ realized correlations, since $d=210$ in this application. Correlations are higher on high-volatility days, as expected, but the shape of the distribution also changes. In particular, the high-volatility distribution is clearly left-skewed, showing that the distribution of realized correlations depends strongly on the volatility regime.
The regime dependence becomes even clearer in the bivariate case. Figure (ref) compares joint distributions for two representative coordinate pairs on the low- and high-volatility days. The low- and high-volatility regimes occupy distinct regions of the joint distribution, especially for the Fisher-transformed coordinates. Thus, the apparent bimodality in the pooled distribution of $\hat\phi$ is largely a mixture-of-regimes effect rather than evidence of bimodality within each volatility regime. The corresponding GFT coordinates display less regime separation and weaker dependence.
We revisit ABDE:2001, who found that the unconditional empirical distribution of realized correlations, $\hat{\varrho}_{i,t}$, for stock returns is approximately normally distributed.\footnote{This empirical result was also demonstrated for exchange rate data; see ABDL:2001.} Our focus is on the distributional properties of the measurement errors in the realized measures. Since the latent paths $\varrho_t$, $\phi_t$, and $\gamma_t$ are unobserved, we approximate them using a standard linear Gaussian filter, obtaining smoothed paths $\varrho_t^f$, $\phi_t^f$, and $\gamma_t^f$. We then define standardized residuals by $\hat{\varepsilon}(\hat{\varrho})=\sqrt{T}(\hat{\varrho}_{t}-\varrho_{t}^{f})$, $\hat{\varepsilon}(\hat{\phi})=\sqrt{T}(\hat{\phi}_{t}-\phi_{t}^{f})$, and $\hat{\varepsilon}(\hat{\gamma})=\sqrt{T}(\hat{\gamma}_{t}-\gamma_{t}^{f})$, and use these as proxies for the scaled measurement errors.
The upper panels of Figure (ref) show the skewness and excess kurtosis of these residuals, computed separately for the low- and high-volatility subsamples. Residuals based on realized correlations have substantial negative skewness and positive excess kurtosis, especially on high-volatility days. The residuals based on $\hat{\phi}$ and $\hat{\gamma}$ are much closer to Gaussian and display fewer systematic differences across volatility regimes.
The lower panels of Figure (ref) show the distributions of correlations between residual elements. Dependence among the residuals is substantial for $\hat{\varepsilon}(\hat{\varrho})$ and $\hat{\varepsilon}(\hat{\phi})$, especially in the high-volatility subsample. In contrast, the residuals based on $\hat{\gamma}$ are only weakly dependent: their average correlation is close to zero, and the distribution is similar across the low- and high-volatility subsamples.
Overall, the empirical results agree with the simulation evidence. Elements of the log-transformed realized correlation matrix are approximately uncorrelated, and their measurement errors are closer to Gaussian than those of realized correlations. Thus, the finite-sample properties of $\hat{\gamma}$ appear to persist in a realistic data environment with non-Gaussian intraday returns and latent correlations that are likely to vary within the trading day HansenLuo-RobustCorr:2023.
Section (ref) documents that the GFT coordinates have much weaker finite-sample dependence than the raw correlations and the element-wise Fisher transformed correlations. We now give two theoretical results that explain this pattern. The first shows that the asymptotic covariance matrix $V_\gamma(C)$ is stable because the matrix logarithm maps perturbations through a well-conditioned linear operator except near the boundary of the correlation cone. The second shows that, locally around $C=I_n$, the GFT cancels the first-order dependence among overlapping sample correlations. Together, these results explain why $V_\gamma(C)$ is both comparatively stable in $C$ and nearly diagonal in empirically relevant designs.
The stability of $V_\gamma(\cdot)$ admits a clean explanation under elliptical sampling. Normalizing the sample covariance to a correlation matrix removes the radial component of the data, so the kurtosis enters the GFT covariance only through a scalar; the remaining sensitivity is geometric and is governed by the conditioning of $A_C$, which deteriorates only as the eigenvalue spread of $C$ grows.
For an elliptical distribution we write $X_t\sim E(0,\Sigma,\kappa)$, where $\kappa$ is the scalar kurtosis parameter appearing in the asymptotic covariance of $\hat\Sigma$, as made explicit in the proof. The Gaussian case corresponds to $\kappa=0$.
The conditioning of $A_C$ inflates $\|\Pi_C\|_2$ only as $C$ approaches singularity, which is exactly the regime in which $\lambda_{\min}(C)$ was found to drive finite-sample behavior in Section (ref). The bound thus provides an analytic account of both the invariance of $V_\gamma(\cdot)$ and its single point of fragility.
The spectral bound explains why the GFT covariance remains stable when $C$ is away from the boundary of the correlation cone. We next isolate a more local mechanism. Around $C=I_n$, the raw correlations and the element-wise Fisher transformed correlations acquire first-order dependence through overlapping index pairs, whereas the GFT cancels these first-order terms.
For a symmetric zero-diagonal matrix $\Delta$, let $B_\Delta$ denote the $d\times d$ matrix with zero diagonal whose first-order off-diagonal entries are determined by triangles in the lower-triangular indexing convention: $$ [B_\Delta]_{(ij),(ik)}=\Delta_{jk},\qquad [B_\Delta]_{(ij),(jk)}=\Delta_{ik},\qquad [B_\Delta]_{(ik),(jk)}=\Delta_{ij}, $$ for distinct indices $i,j,k$, with entries involving disjoint pairs set equal to zero.
The local diagonalness of $V_\gamma(C)$ is therefore stronger than the corresponding local diagonalness of $V_\varrho(C)$ and $V_\phi(C)$. All three covariance matrices are diagonal at $C=I_n$, but the raw and Fisher coordinates acquire first-order off-diagonal terms as soon as the correlation matrix is perturbed away from the identity. These terms arise from triangles: for example, the asymptotic covariance between $\hat\varrho_{ij}$ and $\hat\varrho_{ik}$ is proportional to $C_{jk}$ to first order. The Fisher transform does not remove this effect, because its derivative is equal to one at the origin and its first nonlinear correction is cubic.
The GFT behaves differently. The matrix logarithm has the expansion $\log(I_n+\Delta)=\Delta-\frac{1}{2}\Delta^2+O(\|\Delta\|_2^3)$, so each GFT coordinate subtracts, to second order, the indirect two-step correlation paths running through the other variables. This is the local cancellation mechanism behind the weak dependence seen in the simulations: the raw and Fisher coordinates inherit first-order triangle dependence, whereas the GFT cancels it to first order.
The finite-sample regularities documented in Section (ref) and the covariance-stability results in Section (ref) have a single inferential payoff: in the GFT coordinates the estimation error is approximately Gaussian, weakly dependent, and governed by a covariance matrix that varies little with the unknown $C$. Standardizing with a plug-in estimate of this covariance is therefore reliable in samples where the analogous operation is much less stable for $\hat\varrho$ and $\hat\phi$. This section examines the consequences for standardized statistics, plug-in covariance estimation, and Wald tests.
Inference about $C$ can be based on the limit distributions in ((ref)), with the asymptotic covariances estimated by the plug-in estimators $V_\varrho(\hat C)$, $V_\phi(\hat C)$, and $V_\gamma(\hat C)$. Define the standardized statistics
where $A^{1/2}=Q\Lambda^{1/2}Q^\prime $ for the eigendecomposition $A=Q\Lambda Q^\prime $. If $\hat C\xrightarrow{p}C$ and the relevant covariance is continuous at $C$, then each statistic converges to $N_d(0,I_d)$. In finite samples, however, the quality of the $N_d(0,I_d)$ approximation is controlled not by the limit but by two distinct sources of error: the non-normality of the underlying coordinate vector, and the error in the plug-in covariance, $V(\hat C)-V(C)$. The GFT improves both, and for a common reason.
The decisive quantity is the sensitivity of $V(\cdot)$ to its argument. Whitening in ((ref)) requires $V(\hat C)$ to be a good estimate of $V(C)$, yet $V(\hat C)$ inherits the full estimation error in $\hat C$. When $V(\cdot)$ varies steeply with $C$, as $V_\varrho(\cdot)$ and $V_\phi(\cdot)$ do, a moderate error in $\hat C$ produces a large error in the whitening matrix, and the standardized statistic is poorly behaved even when its components are individually close to normal. The weak dependence documented in Section (ref) and the stability results in Section (ref) imply that the GFT whitening matrix is much less sensitive to the plug-in step: $V_\gamma(\hat C)\approx V_\gamma(C)$ over a wide range of designs, so the standardization behaves much closer to the infeasible standardization based on the true covariance matrix.
We first assess the standardized statistics directly. We draw $10{,}000$ random $5\times5$ correlation matrices using the design of Section (ref); for each we simulate a sample, compute $\hat C$, and form $Z_{\varrho,T}$, $Z_{\phi,T}$, and $Z_{\gamma,T}$ (each of dimension $d=10$). Figure (ref) reports the pooled Q--Q plots against the standard normal for $T=40,100,250$.
The contrast is stark, and it separates the two error sources. The poor behavior of $Z_{\varrho,T}$ is expected from the non-normality of $\hat\varrho$ documented earlier. The behavior of $Z_{\phi,T}$ is more instructive: even though the marginal elements of $\hat\phi$ are close to normal, the dependence among them is strong and highly sensitive to $C$, so the plug-in matrix $V_\phi(\hat C)$ is a poor estimate of $V_\phi(C)$ and the whitening in ((ref)) degrades the joint distribution. $Z_{\gamma,T}$ is close to standard normal already at $T=40$ because the GFT coordinates are both nearly Gaussian and nearly orthogonal, and because $V_\gamma(\hat C)$ is much closer to $V_\gamma(C)$ than the corresponding plug-in covariance matrices for $\hat\varrho$ and $\hat\phi$, as documented in Section (ref) and explained by Section (ref).
We quantify the plug-in error directly with the Stein loss
computed for the plug-in estimators of $V_\varrho(C)$, $V_\phi(C)$, and $V_\gamma(C)$ across $1{,}000$ random correlation matrices ($10{,}000$ samples each) and plotted against $\lambda_{\min}(C)$ in Figure (ref).
The reduction is large and systematic. On the $\log_{10}$ scale, $L\big(V_\gamma(\hat C),V_\gamma(C)\big)$ is typically $2$ to $5$ units below the corresponding losses for $\hat\varrho$ and $\hat\phi$, a reduction in plug-in covariance loss by factors of roughly $10^2$ to $10^5$. The gap widens as $\lambda_{\min}(C)\to0$, consistent with Theorem (ref): it is precisely near singularity that $\|\Pi_C\|_2$, and hence the curvature of $V(\cdot)$ that the plug-in must contend with, grows.
Finally, Figure (ref) reports the size of Wald tests of $H_0\!: C=C_0$ at the nominal $5\%$ level, using $W_{\varrho,T}=Z_{\varrho,T}^\prime Z_{\varrho,T}$, $W_{\phi,T}=Z_{\phi,T}^\prime Z_{\phi,T}$, and $W_{\gamma,T}=Z_{\gamma,T}^\prime Z_{\gamma,T}$ against the $\chi^2_d$ critical value. The upper panels use an equicorrelation matrix with all correlations $0.8$; the lower panels use the Toeplitz structure ((ref)) with $\rho=0.8$.
All three tests are oversized when $d/T$ is moderate or large, and all approach the nominal level as $T\to\infty$, since the standardized statistics share the same limit. Convergence is far faster for $W_{\gamma,T}$: in these designs $W_{\varrho,T}$ and $W_{\phi,T}$ require roughly five times as many observations to reach a comparable proximity to nominal size. The improvement traces back to the same source as in Figures (ref) and (ref): a whitening matrix that is reliable because $V_\gamma(\cdot)$ is nearly invariant to $C$. The practical implication is that GFT-based Wald inference is usable at sample sizes where the conventional and element-wise Fisher tests are not.
A non-singular correlation matrix, $C$, can be parametrized by $\gamma(C)=\operatorname{vecl}\log C$. We have investigated $\hat{\gamma}=\gamma(\hat{C})$, where $\hat{C}$ is the sample correlation matrix, with emphasis on its finite-sample distribution and dependence structure. When $\hat{C}$ is computed from independent Gaussian data, the marginal distributions of the elements of $\hat{\gamma}$ are well approximated by their asymptotic normal distributions. Since $\gamma(C)$ coincides with Fisher's transformation when $C$ is a $2\times2$ correlation matrix, this extends a familiar property of the scalar Fisher transformation to higher-dimensional correlation matrices. As with the scalar Fisher transformation, however, the marginal Gaussian approximation can deteriorate when the data are strongly non-Gaussian.
The more striking finding is the joint behavior of $\hat{\gamma}$. The elements of $\hat{\gamma}$ are nearly uncorrelated in finite samples, and the covariance matrix $V_{\gamma,T}(C)$ is far more stable across values of $C$ than the corresponding covariance matrices for $\hat{\varrho}$ and $\hat{\phi}$. This weak dependence appears to be a robust feature of the GFT coordinates. It is present for Gaussian samples, persists across a range of non-Gaussian designs, and is also evident in simulations based on empirical distributions of macroeconomic variables, daily industry returns, and high-frequency financial data.
These properties have direct implications for inference. Because $V_{\gamma}(C)$ is relatively insensitive to the true correlation matrix, the plug-in estimator $V_{\gamma}(\hat{C})$ is much less affected by estimation error in $\hat{C}$ than the analogous plug-in covariance matrices for $\hat{\varrho}$ and $\hat{\phi}$. As a result, standardized statistics based on $V_{\gamma}^{-1/2}(\hat{C})\sqrt{T}(\hat{\gamma}-\gamma)$ have substantially better finite-sample behavior than those based on the sample correlations or the element-wise Fisher transformed correlations. This is reflected in both the Q-Q plots and the Wald-test simulations, where the GFT-based statistic converges much more rapidly to its asymptotic reference distribution.
Our analysis assumes iid observations, which underlies both the closed-form asymptotic covariances and the bootstrap resampling in the empirical designs. Many target applications, such as realized correlations and multivariate volatility, involve serial dependence and, at high frequency, microstructure noise. Under weak dependence the asymptotic covariance of $\hat{C}$ acquires the usual long-run (HAC) form, and feasible inference would replace the iid plug-in with a corresponding long-run covariance estimator. The properties that motivate the GFT, however, namely the near-orthogonality and near-invariance of the coordinate system and the scalar role of higher-order moments in the spectral bound, are properties of the transformation rather than of the sampling scheme, and the high-frequency evidence in Section (ref) suggests they persist under realistic dependence. A formal treatment of dependent data is left for future work.
The approximate orthogonality and covariance stability of $\hat{\gamma}$ also suggest that the GFT coordinates may be useful for regularization of large correlation and covariance matrices. A systematic treatment of this possibility, including near-singular and high-dimensional cases, is left for future research.
We thank participants at the 2022 Vienna--Copenhagen Conference on Financial Econometrics, the inaugural Virtual Time Series Seminar (VTSS), the 2023 Aarhus Workshop in Econometrics II, the 2024 SoFiE Conference, and seminars at University of Freiburg, the University of Chicago Booth School of Business, and Toronto Metropolitan University for helpful comments and discussions. The second author thanks York University for its hospitality during a research visit.
The authors received no financial support for this research. The authors declare no conflicts of interest.
In preparing this manuscript, the authors used ChatGPT (GPT-5.5 Thinking, OpenAI) and Claude (Opus 4.8, Anthropic) for language editing, notational consistency checks, and proofreading. The authors reviewed and edited all output and take full responsibility for the content of the manuscript, including the accuracy of the results, references, and conclusions.
The FRED-MD macroeconomic data are publicly available from the Federal Reserve Bank of St. Louis. The Fama-French 30 industry portfolio returns are publicly available from Kenneth French's Data Library. The high-frequency intraday return data used in Section (ref) are based on NYSE TAQ data, which are proprietary and cannot be redistributed by the authors. Researchers with TAQ access may reconstruct the data from the securities and date range described in the paper. Simulation code and replication scripts are available from the authors upon request.