Normal-Reference Bandwidth Selection

Why Choose a Normal-Reference Bandwidth?

A kernel density estimate replaces each observed loss with a small, smooth contribution and adds those contributions to estimate the loss density. The bandwidth $h$ controls the width of each contribution: a small $h$ can leave a noisy estimate, while a large $h$ can smooth away important features. Earlier material on bandwidth asymptotics explains this bias-variance tradeoff. A normal-reference rule turns it into a practical starting value using the sample size and a scale estimate.

The rule is a reference calculation, not a claim that loss data are normally distributed. Its value is that it gives a reproducible baseline. Its limitation is that its calibration depends on a smooth normal-shaped density, whereas losses can be skewed, have long tails, or be concentrated near zero. The bandwidth methods here are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: For $n$ independent, exact observations $X_1,\ldots,X_n$ from a common continuous distribution, a kernel density estimate with kernel $K$ and bandwidth $h>0$ is $$ \widehat f_h(x)=\frac{1}{nh}\sum_{j=1}^{n}K\!\left(\frac{x-X_j}{h}\right). $$ The bandwidth $h$ is measured in the same units as the observations. A normal-reference bandwidth is a rule-of-thumb value calibrated as though the underlying density were normal.

The Two Common Normal-Reference Rules

Let $s$ denote the sample standard deviation, and let $\operatorname{IQR}=Q_3-Q_1$, where $Q_1$ and $Q_3$ are the first and third sample quartiles. The direct normal-reference rule uses $s$. The robust version substitutes an IQR-based scale when that is smaller than $s$.

Rule of thumb: The normal-reference bandwidth for a Gaussian kernel is $$ h_{\mathrm{NR}}=1.06\,s\,n^{-1/5}. $$ A common robust variant is $$ h_{\mathrm{rob}}=0.9\min\!\left(s,\frac{\operatorname{IQR}}{1.34}\right)n^{-1/5}. $$ The factor $n^{-1/5}$ means $1/\sqrt[5]{n}$. The factor $1.34$ approximately converts the normal-distribution IQR to a standard-deviation scale. These rules assume the observations are independent, exact measurements from one distribution; they do not by themselves adjust for truncation, censoring, or other sampling complications.

The robust rule is less sensitive to an unusually large observation affecting the scale estimate. It does not make a skewed or heavy-tailed distribution normal, nor does it ensure that one bandwidth describes every region of the density well. Both rules give a single global bandwidth.

Why the Normal-Reference Constant Appears

A short derivation shows how the $1.06$ constant is obtained. Under the usual smoothness conditions for a second-order kernel, the leading terms in the asymptotic mean integrated squared error (AMISE) are $$ \operatorname{AMISE}(h)=\frac{R(K)}{nh}+\frac{h^4\mu_2(K)^2R(f'')}{4}, $$ where $R(g)=\int g(x)^2\,dx$, $\mu_2(K)=\int u^2K(u)\,du$, and $f''$ is the second derivative of the density. This approximation applies to a smooth density and a symmetric kernel with finite second moment; it is not an exact finite-sample error formula.

Differentiate the two terms with respect to $h$ and set the derivative equal to zero:

-\frac{R(K)}{nh^2}+h^3\mu_2(K)^2R(f'')=0 \quad\Longrightarrow\quad h_{\mathrm{AMISE}}= \left(\frac{R(K)}{\mu_2(K)^2R(f'')}\right)^{1/5}n^{-1/5}.

For the standard normal kernel $K$, $R(K)=1/(2\sqrt{\pi})$ and $\mu_2(K)=1$. Now use a normal reference density with standard deviation $\sigma$. Its second derivative is $f''(x)=((x/\sigma)^2-1)f(x)/\sigma^2$, and direct integration gives $R(f'')=3/(8\sqrt{\pi}\sigma^5)$. For example, the integral uses $\int (z^2-1)^2e^{-z^2}\,dz=3\sqrt{\pi}/4$. Substituting,

h_{\mathrm{AMISE}} =\left(\frac{1/(2\sqrt{\pi})}{3/(8\sqrt{\pi}\sigma^5)}\right)^{1/5}n^{-1/5} =\left(\frac{4}{3}\right)^{1/5}\sigma n^{-1/5} \approx 1.06\,\sigma n^{-1/5}.

Replacing the reference $\sigma$ by the sample standard deviation $s$ gives the direct rule. The robust version instead uses a scale estimate less affected by extreme observations. It is a practical modification of the reference calculation, not the result of assuming that the IQR and standard deviation are interchangeable for every distribution.

Calculating the Bandwidth

Worked Example: Direct Normal-Reference Rule

A hypothetical portfolio contains 32 independent, exact claim amounts. The sample standard deviation is $s=\$4.50$ thousand. Calculate the direct normal-reference bandwidth for a Gaussian kernel. State the answer in thousands of dollars.

Method. Use $h_{\mathrm{NR}}=1.06s n^{-1/5}$. Since $32=2^5$, $n^{-1/5}=1/2=0.5$.

Substitute and calculate. $$ h_{\mathrm{NR}}=1.06(4.50)(0.5)=4.77(0.5)=2.385. $$ Thus, rounded only at the final step to two decimal places, $h_{\mathrm{NR}}=\$2.39$ thousand.

The bandwidth has the same units as the claim amounts, so its unrounded value is $\$2.385$ thousand. The calculation is internally consistent: the standard deviation is first multiplied by a dimensionless factor, leaving the result in thousands of dollars. A common mistake avoided here is using $n^{-1/2}$ instead of the fifth-root factor $n^{-1/5}$.

Worked Example: Applying the Robust Scale

For a hypothetical set of 243 exact repair losses, suppose $s=\$12$ thousand, $Q_1=\$42$ thousand, and $Q_3=\$55.40$ thousand. Calculate the robust normal-reference bandwidth.

Method. First calculate the IQR-based scale and choose the smaller of that value and $s$. Here $n=243=3^5$, so $n^{-1/5}=1/3$.

Calculate the spread. $$ \operatorname{IQR}=55.40-42=13.40\text{ thousand}, \qquad \frac{\operatorname{IQR}}{1.34}=\frac{13.40}{1.34}=10.00\text{ thousand}. $$ The robust scale is $\min(12,10)=10$ thousand.

Calculate the bandwidth. $$ h_{\mathrm{rob}}=0.9(10)\left(\frac{1}{3}\right)=\frac{9}{3}=3.00\text{ thousand}. $$ The result is $\$3.00$ thousand. For comparison, the direct rule would give $1.06(12)/3=4.24$ thousand. The smaller robust bandwidth follows from the smaller scale estimate, not from a different value of $n$. A common mistake avoided is to use the larger of $s$ and $\operatorname{IQR}/1.34$ in the robust rule.

What a Robust Scale Can and Cannot Do

Using the minimum of two scale estimates can reduce the influence of a large observation on the bandwidth. That is useful when the standard deviation is much larger than the central spread. But the operation only changes the scale input; it does not alter the kernel, add a second bandwidth, or model the shape of the tails.

Worked Example: A Highly Uneven Hypothetical Loss Sample

A hypothetical sample of 1,024 exact losses is strongly right-skewed. Its summary statistics are $s=\$40$ thousand, $Q_1=\$4.95$ thousand, and $Q_3=\$25.05$ thousand. An unusually large loss contributes to the large standard deviation. Calculate both bandwidth rules and assess what the comparison does and does not establish.

Method and scale calculation. Since $1{,}024=4^5$, $n^{-1/5}=1/4=0.25$. The IQR and IQR-based scale are $$ \operatorname{IQR}=25.05-4.95=20.10\text{ thousand}, \qquad \frac{\operatorname{IQR}}{1.34}=15.00\text{ thousand}. $$ The robust scale is $\min(40,15)=15$ thousand.

Calculate both values. $$ h_{\mathrm{NR}}=1.06(40)(0.25)=10.60\text{ thousand}, $$ $$ h_{\mathrm{rob}}=0.9(15)(0.25)=3.375\text{ thousand}. $$ Rounded to two decimal places, the direct bandwidth is $\$10.60$ thousand and the robust bandwidth is $\$3.38$ thousand.

The difference indicates that the standard deviation is much larger than the central IQR-based scale. It does not prove that $\$3.38$ thousand is the best bandwidth: the strongly right-skewed shape still departs from the normal reference, and one global bandwidth may smooth the central region and the far tail differently. The calculation is checked by noting that $3.375<10.60$, as expected because the robust scale $15$ is smaller than $40$. A common mistake avoided is to treat the robust rule as a correction that removes tail misspecification; it only changes the scale used in the rule.

When the Rule Is Unsuitable

Normal-reference selection is most defensible as an initial calculation when the data are a reasonably smooth, roughly symmetric sample without a severe tail problem. For loss data, there are reasons to treat the result cautiously:

Strong skewness. A single normal reference does not reflect a long right tail or a sharply concentrated body. The IQR-based scale can be more stable, but it does not reproduce the asymmetric shape.
Heavy tails or influential extreme losses. The sample standard deviation can be driven upward by a few observations, producing a very large direct bandwidth. The robust alternative reduces that sensitivity, but may then reflect the center more than the tail.
Different levels of detail across the distribution. A single $h$ applies the same smoothing width everywhere. It can be too wide for a tightly grouped region and too narrow for a sparse region, even when its overall value appears reasonable.
Losses near zero. A symmetric kernel centered on a small positive loss can place estimated density below zero. That is a boundary issue, not something fixed by selecting a different normal-reference scale; boundary methods require separate consideration.

The normal-reference value is therefore a benchmark to inspect, not a guarantee of good fit. If the estimated density changes substantially under reasonable bandwidth choices, that sensitivity is itself important to report. The rule also does not convert incomplete or selected observations into an independent, exact sample; its stated assumptions must be checked before applying it.

Common Exam Traps

Using the wrong power of $n$.
The bandwidth scales with $n^{-1/5}$, not $n^{-1/2}$. For $n=243$, the factor is $1/3$, since $3^5=243$.

Mixing the two rules.
The direct rule uses $1.06s$. The robust rule uses $0.9\min(s,\operatorname{IQR}/1.34)$. Do not use the direct-rule constant with the robust scale unless a problem explicitly defines that alternative.

Ignoring units or rounding early.
Both scale estimates and $h$ have the units of the observations; $n^{-1/5}$, $1.06$, $0.9$, and $1.34$ are dimensionless. Keep intermediate values unrounded, then round the final bandwidth as requested.

Calling the rule optimal for every loss distribution.
The $1.06$ factor comes from a normal reference under an asymptotic approximation. It is not a distribution-free optimum, and the robust variant does not remove skewness or heavy tails.

Key takeaway: Calculate the fifth-root factor first, select the stated scale rule, and preserve the data units. Interpret a normal-reference bandwidth as a baseline whose suitability depends on the shape and sampling conditions of the loss data.

Check Your Understanding

For each problem, use the specified normal-reference rule and keep the stated measurement units. These hypothetical problems test both the arithmetic and the interpretation.

A sample of 32 exact losses has $s=8$ thousand dollars. Find the direct normal-reference bandwidth in thousands. (Answer: $4.24$ thousand dollars)
A sample of 243 exact losses has $s=9$ thousand dollars and $\operatorname{IQR}=10.72$ thousand dollars. Find the robust bandwidth. (Answer: $2.40$ thousand dollars)
A sample of 1,024 exact claims has $s=20$ thousand dollars and $\operatorname{IQR}=16.08$ thousand dollars. Calculate the robust bandwidth and the direct bandwidth. (Answer: Robust $2.70$ thousand dollars; direct $5.30$ thousand dollars)
A sample of 3,125 exact losses has $s=25$ thousand dollars, $Q_1=18.95$ thousand dollars, and $Q_3=39.05$ thousand dollars. Find the robust bandwidth and identify which scale controls the minimum. (Answer: $2.70$ thousand dollars; $\operatorname{IQR}/1.34=15$ thousand dollars controls)
A strongly right-skewed sample has $n=243$, $s=18$ thousand dollars, and $\operatorname{IQR}=13.40$ thousand dollars. Calculate its robust normal-reference bandwidth and state one limitation that remains. (Answer: $3.00$ thousand dollars; the robust scale does not make the skewed density normal or guarantee that one bandwidth suits both center and tail)

What You'll Learn