Diagnosing Normal Approximation Errors Under Skewness

Why Skewness Matters in Aggregate Tails

A normal approximation summarizes an aggregate loss distribution using its mean and variance. That can be useful, but those two moments do not describe asymmetry. When aggregate losses are right-skewed, the upper tail may be heavier than the corresponding normal tail, while the lower tail may be lighter. The difference is not determined by skewness alone, but skewness provides a useful first diagnostic.

The purpose of a skewness correction is not to replace an exact calculation or certify that an approximation is accurate. It is to assess whether the normal estimate is likely to be materially affected by asymmetry. The correction is especially worth checking when the event is several standard deviations from the mean, because even a modest absolute change can be a large relative change in a small tail probability. This diagnostic is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Definition: For an aggregate loss $S$ with mean $\mu_S$, variance $\sigma_S^2>0$, and third central moment $E[(S-\mu_S)^3]$, its standardized skewness is $$ \gamma_S=\frac{E[(S-\mu_S)^3]}{\sigma_S^3}. $$ Positive skewness indicates a longer or more influential upper tail relative to the lower tail; negative skewness indicates the reverse. Skewness describes asymmetry, not the exact shape or tail probabilities of the distribution.

The First-Order Skewness Correction

Standardize the aggregate loss by setting $Z=(S-\mu_S)/\sigma_S$. A normal approximation treats $Z$ as standard normal. A first-order Edgeworth approximation adds a term involving $\gamma_S$, the standardized third cumulant. Write $\phi(z)$ for the standard normal density and $\bar{\Phi}(z)=1-\Phi(z)$ for its upper-tail probability.

First-order diagnostic: For a standardized threshold $z=(s-\mu_S)/\sigma_S$, the first-order approximations are $$ P(S\le s)\ \approx\ \Phi(z)+\frac{\gamma_S}{6}(1-z^2)\phi(z) $$ and $$ P(S>s)\ \approx\ \bar{\Phi}(z)+\frac{\gamma_S}{6}(z^2-1)\phi(z). $$ Use these as diagnostics for an approximately normal aggregate distribution, not as exact formulas. They can be unreliable for extreme thresholds, large skewness, or distributions with important lattice effects.

The two formulas are consistent: the upper-tail correction is the negative of the correction to the cumulative probability at the same boundary. If $\gamma_S>0$, the first-order correction to the upper tail is positive for $z>1$, zero at $z=1$, and negative for $0\le z<1$. Thus, positive skewness does not mean that every upper-tail estimate must be increased. The threshold matters.

Why the Correction Has This Form

A short derivation explains both the factor $1/6$ and the threshold term. In a first-order Edgeworth expansion, the approximate standardized density is $$ f_Z(z)\ \approx\ \phi(z)\left[1+\frac{\gamma_S}{6}(z^3-3z)\right]. $$ Integrating the correction from $-\infty$ to $z$ gives the cumulative-probability correction because $$ \frac{d}{dz}\left[(1-z^2)\phi(z)\right]=(z^3-3z)\phi(z). $$ The expression $(1-z^2)\phi(z)$ tends to zero as $z$ tends to $-\infty$, so integration gives the stated CDF approximation. Subtracting that CDF from 1 gives the upper-tail formula.

This derivation justifies the first-order expression, not its accuracy for every aggregate model. Edgeworth expansions are asymptotic approximations, and their usefulness depends on the distribution being sufficiently close to normal and on the event not being too far into the tail. A discrete aggregate can also require a continuity correction or a lattice adjustment at the boundary. If an adjusted probability is negative or exceeds 1, that is a warning that the approximation is being used outside a sensible range.

Compute Skewness From Aggregate Moments

For a compound Poisson aggregate $S=\sum_{j=1}^{N}X_j$, suppose $N$ is Poisson with mean $\lambda$, independent of identically distributed claim amounts $X_j$. If the severity moments are finite, the aggregate cumulant of order $r$ is $\lambda E[X^r]$. In particular, the mean is $\lambda E[X]$, the variance is $\lambda E[X^2]$, and the third cumulant is $\lambda E[X^3]$. For any distribution, the third cumulant equals the third central moment, so these quantities give $\gamma_S$.

Worked Example: A Compound Poisson Upper-Tail Diagnostic

A hypothetical portfolio has Poisson claim frequency with mean $\lambda=100$. Each claim amount is exponentially distributed with mean 1 loss unit, independently of the other claims and of frequency. Estimate the probability that aggregate loss exceeds its mean by two standard deviations using the normal approximation and the first-order skewness diagnostic. Treat this as a smooth-threshold diagnostic; do not apply a separate lattice correction.

Method. Compute the aggregate mean, variance, and third cumulant from the compound Poisson moments, then compare the normal tail with its first-order correction. The exponential severity moments are $E[X]=1$, $E[X^2]=2$, and $E[X^3]=6$.

The aggregate mean is $100(1)=100$, and its variance is $100(2)=200$. Thus $\sigma_S=\sqrt{200}\approx14.1421$. The third cumulant is $100(6)=600$, giving $$ \gamma_S=\frac{600}{200^{3/2}}=\frac{600}{2828.4271}\approx0.2121. $$

The threshold is $s=100+2\sqrt{200}$, so $z=2$. The normal upper tail is $\bar{\Phi}(2)\approx0.0227501$, and $\phi(2)\approx0.0539910$. The skewness correction is $$ \frac{0.2121}{6}(2^2-1)(0.0539910)\approx0.005727. $$

Adding the unrounded normal tail and correction gives approximately $0.028477$, or 0.0285 to four decimal places. The normal estimate alone is about 0.0228, so the first-order diagnostic suggests that positive skewness could matter for this tail. It does not establish the exact tail probability.

Check. The correction is positive because $\gamma_S>0$ and $z=2>1$. A negative correction here would indicate a sign or tail-formula error.

Compare the Center and Both Tails

The correction can be small near the center while still changing a remote tail by a noticeable percentage. Conversely, a large standardized skewness does not, by itself, prove a large error at every threshold: the correction also contains $(z^2-1)\phi(z)$. Since $\phi(z)$ becomes very small far out in the tail, the first-order absolute correction eventually shrinks, even though an asymptotic expansion may also become less trustworthy there.

Worked Example: Positive Skewness and Opposite Tail Changes

A hypothetical aggregate-loss model has mean $\mu_S=500$ loss units, standard deviation $\sigma_S=40$ loss units, and standardized skewness $\gamma_S=0.30$. Use the first-order diagnostic to estimate the probabilities below 420 and above 580. Also state what the normal approximation alone would give.

Method. Standardize each threshold. The lower threshold has $z=(420-500)/40=-2$, and the upper threshold has $z=(580-500)/40=2$. By symmetry of the normal distribution, each unadjusted tail is $\bar{\Phi}(2)\approx0.0227501$, and $\phi(2)\approx0.0539910$.

For the upper tail, the correction is $$ \frac{0.30}{6}(2^2-1)\phi(2) =0.05(3)(0.0539910) \approx0.0080986. $$ The unrounded normal tail plus correction is approximately $0.0227501+0.0080986=0.0308487$, or 0.0308 to four decimal places.

For the lower tail, use the CDF formula at $z=-2$. Its correction is $$ \frac{0.30}{6}(1-(-2)^2)\phi(2) =0.05(-3)(0.0539910) \approx-0.0080986. $$ The adjusted lower-tail diagnostic is $0.0227501-0.0080986\approx0.0146515$, or 0.0147 to four decimal places. The normal approximation gives about 0.0228 for each tail; the first-order adjustment moves probability toward the upper tail, as expected for positive skewness.

Check. The upper- and lower-tail adjustments have equal magnitude and opposite signs at these symmetric thresholds. That follows directly from the two formulas and is a useful sign check.

Assessing Whether an Error May Be Material

A practical diagnosis uses both the absolute and relative size of the correction. The absolute change is the correction term itself. When the normal tail is small, divide its absolute value by the normal tail to see the relative scale. This ratio is only a comparison, not a percentage error against the unknown true probability.

Worked Example: Central Thresholds Versus Tail Thresholds

A hypothetical aggregate loss is standardized to $Z=(S-\mu_S)/\sigma_S$ and has skewness $\gamma_S=0.24$. Compare the first-order upper-tail diagnostic at $z=0.5$, $z=1$, and $z=1.5$. Use $\bar{\Phi}(0.5)\approx0.308537$, $\bar{\Phi}(1)\approx0.158655$, $\bar{\Phi}(1.5)\approx0.066807$, $\phi(0.5)\approx0.352065$, and $\phi(1.5)\approx0.129518$.

Method. At each threshold, calculate $(\gamma_S/6)(z^2-1)\phi(z)$. Here $\gamma_S/6=0.04$.

At $z=0.5$, the correction is $0.04(0.25-1)(0.352065)\approx-0.010562$. The adjusted upper-tail diagnostic is $0.308537-0.010562\approx0.297975$, or 0.2980 to four decimal places.

At $z=1$, the factor $z^2-1$ is zero, so the first-order diagnostic equals the normal tail, approximately 0.1587. This does not imply that the true probability is exactly normal; only this particular first-order correction vanishes.

At $z=1.5$, the correction is $0.04(1.25)(0.129518)\approx0.006476$. The adjusted estimate is $0.066807+0.006476\approx0.073283$, or 0.0733 to four decimal places. This is an increase of about $0.0065$ in absolute terms, or about $0.006476/0.066807\approx9.7\%$ of the normal estimate.

Check. The correction changes sign between $z=0.5$ and $z=1.5$, exactly as the factor $z^2-1$ requires. The relative comparison at $z=1.5$ indicates potential importance, not a verified error against the true probability.

Common Mistakes and Limits of the Diagnostic

A correction calculation is short, so small formula or interpretation errors can have a large effect. In particular, do not omit the factor $z^2-1$ in the upper-tail expression. For example, with $\gamma_S=0.30$, $z=2$, and $\phi(2)\approx0.0540$, the correction is $$ \frac{0.30}{6}(2^2-1)(0.0540)=0.05(3)(0.0540)=0.0081, $$ not $0.0027$. The factor $z^2-1=3$ is essential.

Other checks help keep the interpretation disciplined:

Use standardized skewness. The denominator is $\sigma_S^3$, not the variance or standard deviation alone.
Match the formula to the event. The CDF correction is $(1-z^2)\phi(z)$; the upper-tail correction is $(z^2-1)\phi(z)$.
Do not label the adjusted result exact. It is a first-order approximation, not a bound or a substitute for an exact recursion when that calculation is available.
Check lattice effects. If aggregate losses lie on a grid, the event boundary may call for a continuity correction. A smooth-threshold Edgeworth calculation does not automatically account for that change.
Be cautious in extreme tails. A small correction does not guarantee accuracy there; higher-order terms, tail shape, or a poor normal approximation may dominate.

The most useful conclusion is therefore qualitative and comparative: if the correction is sizeable relative to the normal tail, skewness is a reason to question the normal estimate and seek a more reliable calculation. If it is small, the diagnostic finds no large first-order skewness effect at that threshold, but it does not rule out other sources of error.

Key takeaway: Compute $\gamma_S$, standardize the threshold, and evaluate the signed correction $(\gamma_S/6)(z^2-1)\phi(z)$ for an upper tail. Judge its size relative to the normal estimate, then check whether discreteness, extreme-tail behavior, or other distributional features limit the comparison.

Check Your Understanding

For each problem, use the first-order skewness diagnostic where requested. Round probabilities to four decimal places.

A standardized aggregate has $\gamma_S=0.24$. At $z=2$, calculate the upper-tail correction using $\phi(2)\approx0.0540$. Is it upward or downward? (Answer: approximately $0.0065$, upward.)
A standardized aggregate has $\gamma_S=0.18$. At $z=1.5$, use $\phi(1.5)\approx0.1295$ to calculate the upper-tail correction. (Answer: approximately $0.0049$.)
A standardized aggregate has $\gamma_S=0.30$. At $z=0.5$, use $\phi(0.5)\approx0.3521$. Calculate the upper-tail correction and state its direction. (Answer: approximately $-0.0132$, downward.)
A compound Poisson aggregate has frequency mean $\lambda=50$, with independent exponential claim amounts of mean 1 loss unit. Find its aggregate skewness, then calculate the first-order upper-tail correction at two standard deviations above its mean, using $\phi(2)\approx0.0540$. (Answer: $\gamma_S=0.30$; correction approximately $0.0081$, upward.)
A standardized aggregate has $\gamma_S=0.12$. At $z=2$, the normal upper tail is approximately 0.0228 and $\phi(2)\approx0.0540$. Find the first-order adjusted tail and the correction as a percentage of the unrounded normal tail $0.0227501$. (Answer: adjusted tail approximately $0.0260$; correction approximately 14.2% of the normal tail.)

What You'll Learn