Continuity Corrections for Lattice Aggregate Losses

Why Continuity Corrections Matter

Aggregate losses from discrete claim counts and discrete severities often take values only on a grid. A normal distribution, by contrast, assigns probability to intervals on a continuous scale. If a normal approximation is used directly at the discrete values, the probability assigned to a region can be noticeably misaligned with the event being approximated.

A continuity correction shifts the boundary of the normal probability by half the spacing between adjacent lattice points. The adjustment depends on the event: an inclusive interval needs a correction at both ends, while a cumulative probability needs only one. This tutorial develops those distinctions for aggregate losses and is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Definition: A random variable $S$ is supported on a lattice with span $h>0$ if its possible values lie on points separated by $h$, such as $0,h,2h,\ldots$. For a lattice aggregate loss, $h$ is the spacing between adjacent attainable values in the loss grid. A continuity correction replaces a discrete event by a continuous normal interval whose boundaries lie halfway between the relevant lattice points.

Suppose $S$ is approximated by a normal random variable with the same mean $\mu$ and standard deviation $\sigma$. For a lattice with points separated by $h$, and lattice endpoints $a$ and $b$, the principal correction rules are:

\begin{aligned} P(S\leq k)&\approx \Phi\!\left(\frac{k+h/2-\mu}{\sigma}\right),\\ P(S<k)&\approx \Phi\!\left(\frac{k-h/2-\mu}{\sigma}\right),\\ P(S\geq k)&\approx 1-\Phi\!\left(\frac{k-h/2-\mu}{\sigma}\right),\\ P(S>k)&\approx 1-\Phi\!\left(\frac{k+h/2-\mu}{\sigma}\right),\\ P(a\leq S\leq b)&\approx \Phi\!\left(\frac{b+h/2-\mu}{\sigma}\right) -\Phi\!\left(\frac{a-h/2-\mu}{\sigma}\right). \end{aligned}

Here, $\Phi$ is the standard normal cumulative distribution function. These forms assume the stated cutoffs are lattice points. For a loss lattice consisting of multiples of a dollar amount, first express the event using those attainable amounts. Do not apply the rules as though every dollar were an attainable loss if the aggregate can only change in larger steps.

Why the Half-Step Adjustment Works

The correction follows from associating a continuous interval with each discrete lattice value. The interval for a lattice point extends halfway to the neighboring points on either side. Thus, the lattice value $k$ represents the continuous interval from $k-h/2$ to $k+h/2$. Summing these intervals gives the boundary adjustments in the formulas.

Locate the lattice points.
If $S$ can take values $0,h,2h,\ldots$, identify the spacing $h$ and rewrite the event using those values.

Represent each included value by its cell.
For a lattice value $s$, use the continuous cell from $s-h/2$ to $s+h/2$.

Combine the cells and standardize.
For an inclusive interval, the outer cells determine the corrected endpoints. Convert those endpoints to $z$-scores using $(x-\mu)/\sigma$.

\bigcup_{s=a,a+h,\ldots,b} \left[s-\frac h2,\ s+\frac h2\right] = \left[a-\frac h2,\ b+\frac h2\right].

The normal probability over this combined interval is the difference between its upper and lower normal cumulative probabilities. This explains why an inclusive interval is extended outward by half a lattice step at both ends. For $S\leq k$, the last included lattice value is $k$, so its cell ends at $k+h/2$. For $S<k$, the last included point is $k-h$, so its cell ends at $k-h/2$. The strict and inclusive forms therefore have different boundaries.

Worked Applications

Worked Example: An Inclusive Interval With Fixed Payments

A hypothetical benefit program pays $80 for each successful outcome. The number of successes is $N\sim\operatorname{Binomial}(12,0.25)$, and the aggregate payment is $S=80N$. Use a normal approximation with a continuity correction to estimate $P(160\leq S\leq400)$. State the aggregate mean, variance, and standard deviation.

Method and model. The aggregate is a fixed payment per success, so its possible values are multiples of $80 and the lattice span is $h=80$. For $N\sim\operatorname{Binomial}(n,p)$, $E[N]=np$ and $\operatorname{Var}(N)=np(1-p)$. Therefore:

E[S]=80(12)(0.25)=240,\qquad \operatorname{Var}(S)=80^2(12)(0.25)(0.75)=14{,}400, \qquad \sigma=120.

The event $160\leq S\leq400$ includes the attainable values $160,240,320,400$, corresponding to $N=2,3,4,5$. Extend the lower and upper endpoints by half of $80:

P(160\leq S\leq400) \approx P(120\leq Z_{\!S}\leq440),

where $Z_{\!S}$ denotes the continuous normal approximation to $S$, with mean 240 and standard deviation 120. Standardizing gives:

z_{\text{lower}}=\frac{120-240}{120}=-1,\qquad z_{\text{upper}}=\frac{440-240}{120}=\frac53.

Thus the approximation is $\Phi(5/3)-\Phi(-1)$. Using $\Phi(5/3)\approx0.9522$ and $\Phi(-1)\approx0.1587$, the probability is $0.9522-0.1587=0.7935$ if those intermediate values are rounded to four decimals. Using the unrounded cumulative probabilities, $0.9522096-0.1586553=0.7935543\ldots$, which rounds to 0.7936 to four decimal places.

Check and common mistake. The corrected endpoints are $120 and $440, each exactly $40 from the nearest included lattice value. Using $160 and $400 directly as continuous bounds omits the two half-cells and gives a different approximation. The final answer uses unrounded cumulative probabilities until the final rounding step.

Worked Example: A Compound Poisson Loss and an Interval

A hypothetical portfolio has $N\sim\operatorname{Poisson}(9)$ claims. Each claim costs exactly $20, independently of the count, so $S=20N$. Approximate the probability that total claims cost at least $100 and at most $240, inclusive.

Method and model. For a compound Poisson aggregate with independent claim amounts $X_j$ and count $N\sim\operatorname{Poisson}(\lambda)$, $E[S]=\lambda E[X]$ and $\operatorname{Var}(S)=\lambda E[X^2]$. Here $X=20$ with probability 1 and $\lambda=9$, so:

E[S]=9(20)=180,\qquad \operatorname{Var}(S)=9(20^2)=3{,}600,\qquad \sigma=60.

The lattice span is $h=20$. The inclusive interval has lattice endpoints $100 and $240, so its corrected continuous bounds are $100-10=90$ and $240+10=250$. The standardized bounds are:

z_{\text{lower}}=\frac{90-180}{60}=-1.5,\qquad z_{\text{upper}}=\frac{250-180}{60}=\frac76.

Therefore:

P(100\leq S\leq240) \approx \Phi(7/6)-\Phi(-1.5) \approx 0.8783-0.0668 =0.8115.

The estimate is 0.8115 to four decimal places. The event includes the attainable totals $100, $120, and so on through $240; it does not include every real-valued total between those amounts. The correction accounts for the half-cell on each side of that set.

Common mistake avoided. A $10 correction would be wrong if one treated the loss as having a $10 lattice span. The aggregate changes only in $20 increments, so the proper half-step is $10.

Worked Example: Distinguishing Strict and Inclusive Cumulative Events

A hypothetical program pays $50 per successful trial. The number of successes is $N\sim\operatorname{Binomial}(20,0.10)$, and $S=50N$. Approximate both $P(S>100)$ and $P(S\geq100)$, using the normal approximation to the aggregate.

Aggregate moments. The lattice span is $h=50$, and:

E[S]=50(20)(0.10)=100,\qquad \operatorname{Var}(S)=50^2(20)(0.10)(0.90)=4{,}500,\qquad \sigma=\sqrt{4{,}500}\approx67.0820.

For $S>100$, the value $100 is excluded. The event begins at the next attainable value, $150, so the continuous boundary is halfway between $100 and $150: $100+25=125$. Hence:

P(S>100)\approx 1-\Phi\!\left(\frac{125-100}{\sqrt{4{,}500}}\right) =1-\Phi(0.3727) \approx 1-0.6453 =\mathbf{0.3547}.

For $S\geq100$, the value $100 is included. The continuous boundary is halfway between $50 and $100, or $75. Therefore:

P(S\geq100)\approx 1-\Phi\!\left(\frac{75-100}{\sqrt{4{,}500}}\right) =1-\Phi(-0.3727) \approx 1-0.3547 =\mathbf{0.6453}.

The inclusive probability is larger because it includes the lattice value $100. The two approximations add to 1 here because the events $S>100$ and $S\geq100$ use complementary half-cell boundaries around that cutoff. Both results are rounded to four decimal places.

Common mistake avoided. Applying the same boundary, $125, to both events treats the strict and inclusive inequalities as identical. The boundary for $S\geq100$ is $75, not $125.

Choosing the Correct Boundary

For an inclusive interval, identify the first and last included lattice values before adjusting anything. For cumulative events, attend to whether the cutoff itself is included. The table summarizes the boundaries when $k$ is a lattice value.

Discrete event	Continuous normal region
$S\leq k$	Up to $k+h/2$
$S<k$	Up to $k-h/2$
$S\geq k$	From $k-h/2$ upward
$S>k$	From $k+h/2$ upward
$a\leq S\leq b$	From $a-h/2$ to $b+h/2$

A correction is not a change to the aggregate mean or variance; it changes only the normal region used to approximate a discrete event. It is most directly justified for lattice-supported variables. If the aggregate is not lattice-supported, or if a stated cutoff is not on the lattice, first express the event in terms of the actual attainable values rather than applying a half-step mechanically.

The correction improves the alignment of a continuous approximation with a discrete event, but it does not make a normal model exact. The approximation may still be poor when the aggregate distribution is far from normal. Use the aggregate moments and the normal approximation as intended, and keep the continuity correction separate from any question about the quality of the approximation itself.

Key takeaway: Find the aggregate loss lattice and its span before standardizing. Extend an inclusive interval outward by half a lattice step at each end; for a cumulative event, choose the half-step according to whether the cutoff is included.

Check Your Understanding

For each problem, use a normal approximation with the appropriate continuity correction. Round probabilities to four decimal places.

A hypothetical portfolio has $N\sim\operatorname{Poisson}(25)$ claims, each costing exactly $10. Approximate $P(200\leq S\leq300)$. (Answer: 0.7287)
A program pays $20 per success, with $N\sim\operatorname{Binomial}(16,0.5)$ and $S=20N$. Approximate $P(S\leq140)$. (Answer: 0.4013)
A hypothetical insurer has $N\sim\operatorname{Poisson}(4)$ claims, each costing exactly $15. Approximate $P(S\geq60)$. (Answer: 0.5987)
A program pays $20 for each success, where $N\sim\operatorname{Binomial}(25,0.5)$. Approximate $P(150\leq S\leq350)$. (Answer: 0.9722)
A hypothetical aggregate has $N\sim\operatorname{Poisson}(5)$ claims. Independently, each claim costs $10 or $20 with probability $0.5$ each. Approximate $P(S\leq70)$, using the aggregate mean and variance. (Answer: 0.5000)

Discrete event	Continuous normal region
\(S\leq k\)	Up to \(k+h/2\)
\(S<k\)	Up to \(k-h/2\)
\(S\geq k\)	From \(k-h/2\) upward
\(S>k\)	From \(k+h/2\) upward
\(a\leq S\leq b\)	From \(a-h/2\) to \(b+h/2\)

What You'll Learn