Delta-Method Uncertainty for the Coefficient of Variation

Why Quantify Uncertainty in the Coefficient of Variation?

A fitted severity model produces more than estimates of its parameters. It also produces summaries, such as the coefficient of variation, that describe the modeled loss distribution. Because the parameters are estimated from data, the fitted coefficient of variation is uncertain too. The delta method converts an estimated covariance matrix for the parameters into an approximate variance for this derived quantity.

The coefficient of variation (CV) is the standard deviation divided by the mean. It measures dispersion relative to the scale of the mean, so it is dimensionless. For a fitted loss model, its uncertainty can help distinguish a stable estimate from one that is highly sensitive to sampling variation. This tutorial develops the calculation for a single fitted CV; it does not treat uncertainty in multiple model functions jointly.

Definition: If a loss $X$ has $0<E[X]<\infty$ and $0<\operatorname{Var}(X)<\infty$, its coefficient of variation is $$ c=\frac{\sqrt{\operatorname{Var}(X)}}{E[X]}. $$ For a parametric model with parameter vector $\boldsymbol{\theta}$, write this quantity as $c(\boldsymbol{\theta})$. Given an estimator $\hat{\boldsymbol{\theta}}$, the fitted coefficient of variation is $c(\hat{\boldsymbol{\theta}})$.

The Delta-Method Calculation

Suppose the parameter vector has $p$ components and the estimated covariance matrix of $\hat{\boldsymbol{\theta}}$ is $\widehat{\Sigma}$. The gradient of the CV is the column vector of its partial derivatives, evaluated at the estimated parameters:

\widehat{\boldsymbol{g}} = \nabla c(\hat{\boldsymbol{\theta}}) = \begin{pmatrix} \frac{\partial c}{\partial \theta_1}\\ \vdots\\ \frac{\partial c}{\partial \theta_p} \end{pmatrix}_{\boldsymbol{\theta}=\hat{\boldsymbol{\theta}}}.

The delta method approximates the variance of the fitted CV by

\widehat{\operatorname{Var}}\!\left[c(\hat{\boldsymbol{\theta}})\right] = \widehat{\boldsymbol{g}}^{\,T} \widehat{\Sigma} \widehat{\boldsymbol{g}}, \qquad \widehat{\operatorname{SE}}\!\left[c(\hat{\boldsymbol{\theta}})\right] = \sqrt{\widehat{\boldsymbol{g}}^{\,T} \widehat{\Sigma} \widehat{\boldsymbol{g}}}.

Here $\widehat{\Sigma}$ is a covariance matrix of parameter estimates, not a matrix of standard errors. It may come from an appropriate estimated information matrix or another estimator-variance calculation. The delta method is an approximation: it is most useful when the parameter estimator is reasonably concentrated and $c$ is differentiable near the fitted parameter values.

Interpretation: The gradient describes how sensitive the CV is to small changes in each parameter. The covariance matrix describes the uncertainty and dependence among the parameter estimates. The quadratic form combines both; off-diagonal covariances matter when the CV depends on more than one parameter.

Why the Formula Works

The justification is a first-order Taylor approximation around the true parameter vector $\boldsymbol{\theta}$. For a nearby estimate $\hat{\boldsymbol{\theta}}$,

c(\hat{\boldsymbol{\theta}}) \approx c(\boldsymbol{\theta}) + \nabla c(\boldsymbol{\theta})^T (\hat{\boldsymbol{\theta}}-\boldsymbol{\theta}).

The first term is constant. For any fixed vector $\boldsymbol{a}$, the variance of a linear combination $\boldsymbol{a}^T\hat{\boldsymbol{\theta}}$ is $\boldsymbol{a}^T\operatorname{Cov}(\hat{\boldsymbol{\theta}})\boldsymbol{a}$. Applying that fact to the linear term in the approximation gives the delta-method variance:

\operatorname{Var}\!\left[c(\hat{\boldsymbol{\theta}})\right] \approx \nabla c(\boldsymbol{\theta})^T \operatorname{Cov}(\hat{\boldsymbol{\theta}}) \nabla c(\boldsymbol{\theta}).

In practice, the unknown parameter vector and its covariance are replaced with their estimates. This argument also explains the limitation: if the function is strongly curved over the range of plausible estimates, or the estimate is near a boundary where the function is not smooth, a first-order approximation may not describe the uncertainty well.

Worked Calculations for Common Severity Models

For many two-parameter severity models, the CV depends on the shape parameter but not on the scale parameter. Multiplying every loss by a positive constant multiplies both the mean and standard deviation by that constant, leaving their ratio unchanged. The examples make this cancellation explicit and show how to carry the parameter covariance through the calculation.

Worked Example: Gamma Severity and Shape Uncertainty

A hypothetical portfolio is fitted with a Gamma severity model parameterized by shape $k$ and scale $\theta$, so that $E[X]=k\theta$ and $\operatorname{Var}(X)=k\theta^2$. The estimates are $\hat{k}=4$ and $\hat{\theta}=\$200$. The estimated covariance matrix, in the parameter order $(k,\theta)$, is

\widehat{\Sigma} = \begin{pmatrix} 0.16 & 1.2\\ 1.2 & 100 \end{pmatrix}.

There is no interest-rate or mortality assumption in this severity-model calculation. Find the fitted CV, its delta-method standard error, and a two-sided approximate 95% confidence interval.

Method. From the Gamma mean and variance, \[ c(k,\theta)=\frac{\sqrt{k\theta^2}}{k\theta} =\frac{1}{\sqrt{k}}, \] where $k,\theta>0$. The scale cancels. The fitted CV is $c(4,\ 200)=1/\sqrt{4}=0.5$.

Differentiate with respect to both parameters: \[ \frac{\partial c}{\partial k}=-\frac{1}{2k^{3/2}}, \qquad \frac{\partial c}{\partial \theta}=0. \] At $k=4$, the gradient is $(-1/16,\ 0)^T=(-0.0625,\ 0)^T$. Thus the estimated variance is \[ (-0.0625,\ 0) \begin{pmatrix}0.16&1.2\\1.2&100\end{pmatrix} \begin{pmatrix}-0.0625\\0\end{pmatrix} =0.0625^2(0.16)=0.000625. \] The standard error is $\sqrt{0.000625}=0.025$.

Using the normal critical value $1.96$, the approximate interval is \[ 0.5\pm1.96(0.025)=0.5\pm0.049 =[0.451,\ 0.549]. \] Rounded to three decimal places, the fitted CV is $0.500$, its standard error is $0.025$, and the interval is $[0.451,\ 0.549]$. Check: the estimated variance is positive, and the interval is centered on the fitted CV.

Common mistake avoided: the supplied $0.16$ is the estimated variance of $\hat{k}$, not its standard error. Using it as a standard error would give $0.0625(0.16)=0.010$, which is not the delta-method standard error. The correct calculation uses the variance formula and then takes its square root.

Worked Example: Lognormal Severity With Correlated Estimates

A hypothetical claim severity follows a lognormal model with $\ln X\sim N(\mu,\sigma^2)$. The fitted parameters are $\hat{\mu}=6.0$ and $\hat{\sigma}=0.5$. The estimated covariance matrix, in the order $(\mu,\sigma)$, is

\widehat{\Sigma} = \begin{pmatrix} 0.01 & -0.001\\ -0.001 & 0.0009 \end{pmatrix}.

There is no interest-rate or mortality assumption. Find the fitted CV and its delta-method standard error. Also give a two-sided approximate 95% interval.

Method. For this parameterization, $E[X]=e^{\mu+\sigma^2/2}$ and $\operatorname{Var}(X)=(e^{\sigma^2}-1)e^{2\mu+\sigma^2}$. Substitution into the definition gives \[ c(\mu,\sigma)=\sqrt{e^{\sigma^2}-1}. \] The parameter $\mu$ cancels, so $\partial c/\partial\mu=0$. Differentiating with respect to $\sigma>0$ gives \[ \frac{\partial c}{\partial\sigma} =\frac{\sigma e^{\sigma^2}}{\sqrt{e^{\sigma^2}-1}}. \]

At $\hat{\sigma}=0.5$, the fitted CV is \[ \sqrt{e^{0.25}-1}=\sqrt{1.2840254-1} =\sqrt{0.2840254}\approx0.532940. \] The gradient’s $\sigma$-component is \[ \frac{0.5e^{0.25}}{\sqrt{e^{0.25}-1}} \approx\frac{0.642013}{0.532940} \approx1.20466. \] Since the $\mu$-component is zero, the estimated variance is \[ (1.20466)^2(0.0009)\approx0.001306, \] and the standard error is $\sqrt{0.001306}\approx0.03614$. The off-diagonal covariance makes no contribution here because the gradient component for $\mu$ is zero.

The approximate 95% interval is \[ 0.532940\pm1.96(0.03614) \approx0.532940\pm0.07083 =[0.46211,\ 0.60377]. \] Rounded to three decimal places, the fitted CV is $0.533$, the standard error is $0.036$, and the interval is $[0.462,\ 0.604]$. Check: the interval lies above zero and is centered, up to rounding, at $0.533$.

Common mistake avoided: treating $\mu$ as if it affected the CV would add an unnecessary gradient component. A lognormal location shift changes the mean and standard deviation by the same multiplicative factor, so it does not change their ratio.

Worked Example: Lomax Severity and the Finite-Variance Condition

A hypothetical portfolio is fitted with a Lomax (Pareto Type II) loss model with shape $\alpha$ and scale $\theta$, where $\alpha>2$ and $\theta>0$. Its mean and variance are \[ E[X]=\frac{\theta}{\alpha-1}, \qquad \operatorname{Var}(X)=\frac{\alpha\theta^2}{(\alpha-1)^2(\alpha-2)}. \] The estimates are $\hat{\alpha}=5$ and $\hat{\theta}=\$1{,}000$, with estimated covariance matrix

\widehat{\Sigma} = \begin{pmatrix} 0.25 & 0.5\\ 0.5 & 40000 \end{pmatrix}.

There is no interest-rate or mortality assumption. Calculate the fitted CV and its delta-method standard error, then give an approximate 95% interval.

Method. Dividing the standard deviation by the mean and simplifying gives \[ c(\alpha,\theta) =\sqrt{\frac{\alpha}{\alpha-2}}. \] The scale cancels, and the condition $\alpha>2$ is essential: without it, the variance and hence the CV are not finite. Differentiation yields \[ \frac{\partial c}{\partial\alpha} =-\frac{c(\alpha,\theta)}{\alpha(\alpha-2)}, \qquad \frac{\partial c}{\partial\theta}=0. \]

At $\hat{\alpha}=5$, the fitted CV is $\sqrt{5/3}\approx1.290994$. The shape derivative is \[ -\frac{1.290994}{5(5-2)} =-\frac{1.290994}{15} \approx-0.0860663. \] Therefore, \[ \widehat{\operatorname{Var}}[c(\hat{\alpha},\hat{\theta})] =(-0.0860663)^2(0.25) \approx0.00185185, \qquad \widehat{\operatorname{SE}}\approx0.043033. \] Using $1.96$, the approximate interval is \[ 1.290994\pm1.96(0.043033) \approx1.290994\pm0.084345 =[1.20665,\ 1.37534]. \] Rounded to three decimal places, the fitted CV is $1.291$, the standard error is $0.043$, and the interval is $[1.207,\ 1.375]$. Check: the estimate $\hat{\alpha}=5$ satisfies the finite-variance condition $\alpha>2$.

Common mistake avoided: using a formula for the CV without checking the moment conditions can produce a numerical expression even where the variance does not exist. For this Lomax model, the CV calculation requires $\alpha>2$.

Practical Interpretation and Exam Traps

The delta-method standard error is measured on the CV scale, just like the CV itself. A larger standard error indicates greater estimated sampling uncertainty, not greater dispersion of individual losses. The CV describes the fitted distribution; its standard error describes uncertainty in estimating that summary.

A common approximation is the Wald interval $c(\hat{\boldsymbol{\theta}})\pm z\,\widehat{\operatorname{SE}}$, where $z$ is the selected standard normal critical value. It can extend below zero even though a CV cannot be negative. Such an interval is then a warning that a symmetric normal approximation may be poor, especially when the estimate is near a boundary or uncertainty is large. Do not interpret the interval as an exact probability statement about a fixed parameter.

When a CV depends on several parameters, expand the quadratic form rather than discarding the covariance terms. For two parameters, with gradient components $g_1,g_2$, parameter variances $V_1,V_2$, and covariance $C_{12}$, the approximation is

\widehat{\operatorname{Var}}[c(\hat{\boldsymbol{\theta}})] =g_1^2V_1+g_2^2V_2+2g_1g_2C_{12}.

The covariance term can increase or decrease the result depending on the signs of the derivatives and the covariance. It vanishes in the examples above because the CV does not depend on the scale or location parameter, not because covariance is generally irrelevant. Before calculating, check the parameterization, moment conditions, gradient, and order of the entries in the covariance matrix.

Write the model-specific CV.
Start from the fitted model’s mean and variance, and simplify their ratio before differentiating.

Check existence and parameter conditions.
Confirm that the mean is positive and finite, the variance is finite, and the function is differentiable at the fitted values.

Evaluate the gradient and covariance matrix.
Use the stated parameter order and retain covariance terms whenever the corresponding gradient components are nonzero.

Calculate and interpret uncertainty.
Compute the quadratic form, take its square root for the standard error, and assess whether a normal-based interval is plausible.

Key takeaway: Approximate uncertainty in a fitted coefficient of variation by multiplying its parameter gradient on both sides of the estimated parameter covariance matrix. The result is specific to the chosen model, parameterization, and validity of its moments.

Check Your Understanding

For each problem, assume the supplied parameter covariance is appropriate for the fitted model. Round CVs and standard errors to four decimal places when needed.

A Gamma severity model has fitted shape $k=9$ and scale $\theta=\$80$. The estimated variance of $\hat{k}$ is $0.09$. Find the fitted CV and its delta-method standard error. (Answer: CV $=0.3333$; standard error $=0.0056$.)
A lognormal severity model has $\hat{\sigma}=0.4$ and estimated variance of $\hat{\sigma}$ equal to $0.0004$. Find the fitted CV and its delta-method standard error. (Answer: CV $\approx0.4165$; standard error $\approx0.0225$.)
A Lomax severity model has $\hat{\alpha}=4$, with estimated variance of $\hat{\alpha}$ equal to $0.16$. Find the fitted CV and its delta-method standard error, and state whether the fitted shape satisfies the finite-variance condition. (Answer: CV $=\sqrt{2}\approx1.4142$; standard error $\approx0.0707$; yes, $4>2$.)
A Gamma model has $\hat{k}=4$ and estimated variance of $\hat{k}$ equal to $0.04$. Calculate the delta-method standard error and the approximate 95% Wald interval for its fitted CV. (Answer: standard error $=0.0125$; interval $[0.4755,\ 0.5245]$.)
A Lomax model is fitted with $\hat{\alpha}=1.8$. Can the finite-variance CV and its delta-method standard error be calculated from the ordinary Lomax variance formula? Explain the issue. (Answer: No. Since $\hat{\alpha}\leq2$, the fitted variance is not finite, so the CV is undefined and the stated delta-method calculation does not apply.)

What You'll Learn