Deriving Marginal Contributions from a Capital Function

Why Derivatives Describe Marginal Capital

An aggregate capital model maps a portfolio’s exposures to a capital amount. To understand how a small change in one component affects that amount, hold the other components fixed and differentiate the capital function with respect to the exposure of interest. This gives a local sensitivity: the rate at which modeled capital changes as that exposure changes near the portfolio being evaluated.

That sensitivity is useful for comparing how strongly different components affect aggregate capital at a particular portfolio. It is not, by itself, the exact capital change for a finite exposure adjustment. The distinction matters: a derivative is a local rate, while a finite change can also reflect curvature and interactions in the capital function.

Definition: Let $C(x_1,\ldots,x_n)$ be a differentiable aggregate capital function, where $x_i$ is the exposure measure for component $i$. At a specified portfolio $x=(x_1,\ldots,x_n)$, the marginal capital contribution of component $i$ is the partial derivative $$ MC_i(x)=\frac{\partial C}{\partial x_i}(x). $$ It measures the local change in aggregate capital per unit increase in $x_i$, with all other coordinates held fixed.

The units follow directly from the definition. If $C$ is measured in millions of dollars and $x_i$ is measured in exposure units, then $\partial C/\partial x_i$ is measured in millions of dollars per exposure unit. The value is specific to the point where it is evaluated; for a nonlinear capital function, the same component can have different marginal contributions at different portfolio sizes.

Partial Derivatives and the Total Differential

For a differentiable function of several components, the first-order change in capital for a small vector of exposure changes is given by the total differential. If each exposure changes by $dx_i$, then

dC=\sum_{i=1}^{n}\frac{\partial C}{\partial x_i}\,dx_i.

This expression combines the component sensitivities at the current portfolio with the proposed exposure changes. It gives a first-order approximation to the change in capital. In general, for a small finite change $h$,

C(x+h)-C(x)=\nabla C(x)^\mathsf{T}h+o(\lVert h\rVert),

where the remainder divided by $\lVert h\rVert$ tends to zero as $h$ tends to zero. Thus, the derivative-based approximation becomes relatively accurate for sufficiently small changes, but it need not equal the exact change for a nonzero adjustment.

Specify the capital function and portfolio.
Write the units of the capital output and each exposure, then record the point at which sensitivities are required.

Differentiate with respect to each component.
For a partial derivative, vary the selected component and hold the other exposure coordinates fixed.

Evaluate and interpret.
Substitute the portfolio values into the derivatives, preserve the units, and describe each result as a local change in aggregate capital.

Use the total differential for a small portfolio change.
Multiply each marginal contribution by its corresponding exposure change and add the resulting first-order effects.

Why the Chain Rule Is Needed

Capital functions are sometimes written in terms of intermediate quantities rather than directly in terms of business-unit exposures. Suppose $C$ depends on $z_1,\ldots,z_m$, and each $z_j$ depends on exposures $x_1,\ldots,x_n$. A change in $x_i$ can affect capital through every intermediate quantity that depends on it. The multivariable chain rule is

\frac{\partial C}{\partial x_i} = \sum_{j=1}^{m} \frac{\partial C}{\partial z_j} \frac{\partial z_j}{\partial x_i}.

To see why, change $x_i$ by a small amount $dx_i$. Each intermediate quantity changes locally by $dz_j=(\partial z_j/\partial x_i)dx_i$. The resulting first-order capital change is $dC=\sum_j(\partial C/\partial z_j)dz_j$. Substituting the expressions for $dz_j$ and dividing by $dx_i$ gives the stated derivative. This argument requires the relevant functions to be differentiable at the point under consideration.

Worked Example: Direct Marginal Contributions and a Small Change

A hypothetical insurer uses the capital function $C(x,y)=4x^2+3xy+2y^2$, with $C$ measured in millions of dollars and $x,y$ measured in exposure units. At $(x,y)=(1,2)$, calculate both marginal contributions and estimate the effect of increasing $x$ by $0.01$, leaving $y$ unchanged. Compare the estimate with the exact change.

Method. Differentiate the capital function with respect to each exposure, holding the other one fixed:

\frac{\partial C}{\partial x}=8x+3y, \qquad \frac{\partial C}{\partial y}=3x+4y.

At $x=1,y=2$, the marginal contributions are

\frac{\partial C}{\partial x}=8(1)+3(2)=14, \qquad \frac{\partial C}{\partial y}=3(1)+4(2)=11.

Thus, the local sensitivities are $14$ and $11$ million dollars per exposure unit. For $dx=0.01$ and $dy=0$, the total differential estimates the capital change as

dC\approx 14(0.01)+11(0)=0.14

million dollars. To check against the exact change, first calculate $C(1,2)=4+6+8=18$ million dollars. At the new point $(1.01,2)$,

$$ C(1.01,2)=4(1.01)^2+3(1.01)(2)+2(2)^2 =4.0804+6.06+8=18.1404. $$

The exact increase is $18.1404-18=0.1404$ million dollars. The derivative estimate is $0.14$ million dollars, close to the exact amount for this small change. The $0.0004$ million difference is the higher-order effect. The common mistake is to treat the derivative estimate as an exact finite change; it is a local approximation unless the function is linear over the change.

Worked Example: Applying the Chain Rule to Exposure Transformations

A hypothetical model defines intermediate quantities $a=2s+t$ and $b=s-t$, with aggregate capital $C=10+a^2+2b^2$ in millions of dollars. The coordinates $s,t$ are exposure units. At $(s,t)=(1,0)$, find both marginal contributions and estimate the capital change for $ds=0.02$ and $dt=-0.01$.

Method. First evaluate the intermediate quantities: $a=2(1)+0=2$ and $b=1-0=1$. Differentiate capital with respect to those quantities:

\frac{\partial C}{\partial a}=2a, \qquad \frac{\partial C}{\partial b}=4b.

The exposure derivatives of the intermediate quantities are $a_s=2,\ a_t=1,\ b_s=1,$ and $b_t=-1$. Applying the chain rule gives

\frac{\partial C}{\partial s} =2a(2)+4b(1)=4a+4b, \qquad \frac{\partial C}{\partial t} =2a(1)+4b(-1)=2a-4b.

At $a=2,b=1$, this yields

\frac{\partial C}{\partial s}=4(2)+4(1)=12, \qquad \frac{\partial C}{\partial t}=2(2)-4(1)=0.

The marginal contributions are $12$ and $0$ million dollars per exposure unit. The first-order change for the specified adjustment is

dC\approx 12(0.02)+0(-0.01)=0.24

million dollars. As a check, the initial capital is $10+2^2+2(1^2)=16$ million dollars. At $s=1.02,t=-0.01$, the intermediate values are $a=2.03$ and $b=1.03$, so

$$ C=10+(2.03)^2+2(1.03)^2 =10+4.1209+2.1218=16.2427. $$

The exact increase is $16.2427-16=0.2427$ million dollars, consistent with the $0.24$ million first-order estimate. A common chain-rule error is to differentiate $C$ with respect to $t$ as $2a-4b$, then substitute incorrect values or signs. Here it is $2(2)-4(1)=0$, not $12$: $t$ increases $a$ but decreases $b$, and those effects cancel at this particular portfolio.

Interpreting Marginal Contributions Carefully

A positive partial derivative means that a small increase in that exposure, with the others fixed, increases modeled capital locally. A negative derivative means that a small increase locally reduces modeled capital. A zero derivative means there is no first-order change in that coordinate at the evaluated point; it does not establish that larger changes have no effect.

The size of a derivative is also influenced by the units used for the exposure. If the same exposure is recorded in hundreds rather than individual units, the numerical derivative changes accordingly. Comparisons across components therefore require attention to both the exposure definitions and their units.

Worked Example: Several Components Changing Together

A hypothetical portfolio has capital function $C(x,y,z)=5+(2x+y)^2+3yz+2z^2$, measured in millions of dollars. The exposure variables are measured in units. At $(x,y,z)=(1,2,1)$, calculate the three marginal contributions and estimate the effect of $dx=-0.02,\ dy=0.03,\ dz=0.01$.

Method. Differentiate with respect to each coordinate:

\frac{\partial C}{\partial x}=4(2x+y),\qquad \frac{\partial C}{\partial y}=2(2x+y)+3z,\qquad \frac{\partial C}{\partial z}=3y+4z.

At the specified portfolio, $2x+y=2(1)+2=4$. Therefore,

\frac{\partial C}{\partial x}=4(4)=16,\quad \frac{\partial C}{\partial y}=2(4)+3(1)=11,\quad \frac{\partial C}{\partial z}=3(2)+4(1)=10.

The marginal contributions are $16,\ 11,$ and $10$ million dollars per exposure unit. The first-order estimate for the simultaneous changes is

dC\approx16(-0.02)+11(0.03)+10(0.01) =-0.32+0.33+0.10=0.11

million dollars. Check the baseline: $C(1,2,1)=5+4^2+3(2)(1)+2(1)^2=29$. At the adjusted point $(0.98,2.03,1.01)$, $2x+y=3.99$, so

$$ C=5+(3.99)^2+3(2.03)(1.01)+2(1.01)^2 =5+15.9201+6.1509+2.0402=29.1112. $$

The exact increase is $29.1112-29=0.1112$ million dollars, close to the first-order estimate of $0.11$ million. The check also shows why each exposure change must be paired with its own derivative: omitting the negative $x$-change or assigning a derivative to the wrong coordinate would give a different estimate.

Common Exam Traps and Model Use

Confusing a partial derivative with a finite difference. A derivative is a local rate. For a finite change, use the total differential only as an approximation unless the model is linear over that interval.
Ignoring indirect pathways. If exposures enter through intermediate variables, differentiate through each pathway using the chain rule.
Changing more than one coordinate in a partial derivative. A partial derivative varies one exposure while holding the other coordinates fixed. For a simultaneous adjustment, combine the partial derivatives with the respective changes using the total differential.
Dropping signs or units. A negative exposure change contributes a negative first-order effect when its marginal contribution is positive. Report the derivative in capital units per exposure unit.
Interpreting a zero derivative too broadly. A zero derivative only removes the first-order effect at that point; nonlinear effects may still appear for finite changes.

In model review and capital planning, derivatives provide a disciplined way to describe local sensitivity and identify which exposure changes have the strongest modeled effect at the current portfolio. Their interpretation depends on the selected capital function, its differentiability, the exposure definitions, and the portfolio point. Marginal sensitivities should therefore be reported with that context rather than as universal constants for a business component.

This treatment is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus. Its central calculation is direct: differentiate the aggregate capital function with respect to a component, apply the chain rule when exposures are transformed, and interpret the result as a local rate of capital change.

Check Your Understanding

For each problem, compute the requested derivative at the stated portfolio and keep the units in mind.

A hypothetical model has $C(x,y)=6+3x^2+2xy+y^2$ million dollars. At $(x,y)=(1,2)$, find both marginal contributions. (Answer: $\partial C/\partial x=10$ and $\partial C/\partial y=6$ million dollars per exposure unit.)
A model has $C=20+a^2+3b^2$ million dollars, where $a=p+2q$ and $b=2p-q$. At $(p,q)=(1,1)$, find the marginal contributions with respect to $p$ and $q$. (Answer: $\partial C/\partial p=18$ and $\partial C/\partial q=6$ million dollars per exposure unit.)
A model has $C(u,v)=2u^2+uv+v^2$ million dollars. At $(u,v)=(2,1)$, use the total differential to estimate the change for $du=0.01,\ dv=-0.02$. (Answer: $\partial C/\partial u=9,\ \partial C/\partial v=4$, and the estimated change is $0.01$ million dollars.)
A model has $C(m,n)=10+(m+n)^2+2n^2$ million dollars. At $(m,n)=(1,2)$, find each marginal contribution and estimate the change if $n$ increases by $0.05$ while $m$ is fixed. (Answer: $\partial C/\partial m=6,\ \partial C/\partial n=14$ million dollars per exposure unit; estimated capital increase $=0.70$ million dollars.)
A model has $C(x,y)=x^2+4y^2$ million dollars. At $(x,y)=(2,1)$, calculate the marginal contribution of $x$ and the exact capital change if $x$ increases by $0.1$ with $y$ fixed. Compare with the first-order estimate. (Answer: $\partial C/\partial x=4$ million dollars per exposure unit; exact increase $=0.41$ million dollars and first-order estimate $=0.40$ million dollars.)

What You'll Learn