Expected Loss as a Distributional Average

Expected Loss Is a Probability-Weighted Average

A loss distribution describes a range of possible outcomes and how probability is assigned to them. Expected loss compresses that distribution into one number: its probability-weighted average. This is useful for comparing average costs across risks, but it does not predict the loss in any particular period. A low expected loss can coexist with a small probability of a very large loss.

Definition: For a loss random variable $X$, the expected loss, written $E[X]$, is the probability-weighted average of the possible values of $X$. For a discrete distribution with possible losses $x_j$ and probabilities $p_j=P(X=x_j)$, $$ E[X]=\sum_j x_jp_j. $$ For a continuous distribution with density $f_X(x)$, $$ E[X]=\int_{-\infty}^{\infty}x f_X(x)\,dx, $$ provided the expectation exists. For a nonnegative loss, the integral can be restricted to the range of possible losses. The units of $E[X]$ are the same as the units of $X$.

The probability weights matter. A loss with a small probability contributes less to the average than an equally large loss with a greater probability. Conversely, even a rare large loss can make a material contribution. For a loss taking finitely many values, the probabilities must be nonnegative and sum to 1. For a density, the density must be nonnegative and its integral over the possible range must equal 1.

Why the Discrete Formula Works

For a discrete loss, the expected value is not an unweighted average of the listed outcomes. Each outcome is represented in proportion to how often it occurs in the probability model. One way to see this is to write the loss as a sum of its possible values multiplied by indicators for whether each value occurs.

X=\sum_j x_j I(X=x_j) \quad\Longrightarrow\quad E[X]=\sum_j x_j E[I(X=x_j)] =\sum_j x_j P(X=x_j) =\sum_j x_jp_j.

Here $I(X=x_j)$ equals 1 when $X=x_j$, and 0 otherwise, so its expectation is the probability of that outcome. The derivation uses linearity of expectation. It explains why rare outcomes are included but weighted by their probabilities. For a continuous distribution, the corresponding average is formed by integrating values against the density rather than summing probability masses.

Worked Example: A Discrete Loss Table

A hypothetical one-year equipment policy has the following modeled claim outcomes. Calculate the expected payment, identify the most likely outcome, and explain whether the expected payment is guaranteed.

Payment $x_j$	Probability $p_j$	Contribution $x_jp_j$
$0	0.92	$0
$5,000	0.06	$300
$40,000	0.02	$800

Step 1: Check the probabilities. Their sum is $0.92+0.06+0.02=1.00$, so the table specifies a probability distribution.

Step 2: Weight each payment by its probability. The contributions are $0\times0.92=\$0$, $\$5{,}000\times0.06=\$300$, and $\$40{,}000\times0.02=\$800$. Therefore,

E[X]=\$0+\$300+\$800=\$1{,}100.

Step 3: Interpret the result. The most likely outcome is a payment of $0, with probability 0.92. The expected payment is $1,100, but that amount is not one of the modeled outcomes. It is an average over the probability distribution, not a promised payment or a prediction that this policy will pay exactly $1,100.

Method and mistake avoided. This is a probability-weighted sum. Simply averaging the three listed payments would give $(\$0+\$5{,}000+\$40{,}000)/3=\$15{,}000$, which incorrectly treats the outcomes as equally likely.

Expected Loss from a Density

When $X$ is continuous, individual values generally have probability zero; probabilities are assigned to ranges of values. The density $f_X(x)$ supplies the weights in the expected-value integral. It is not itself a probability at a point. Before integrating $x f_X(x)$, check that the density covers the full range of possible losses and integrates to 1.

Worked Example: Integrating a Loss Density

A hypothetical repair cost $X$, measured in thousands of dollars, has density

f_X(x)=\frac{2x}{10{,}000},\quad 0\le x\le100,

and density 0 elsewhere. Verify that this is a density and calculate the expected repair cost in dollars.

Step 1: Check nonnegativity and normalization. The expression $2x/10{,}000$ is nonnegative on $0\le x\le100$. Its integral over that range is

\int_0^{100}\frac{2x}{10{,}000}\,dx =\left[\frac{x^2}{10{,}000}\right]_0^{100} =\frac{10{,}000}{10{,}000}=1.

Thus it is a valid density on the stated range.

Step 2: Integrate loss times density. Since $x$ is in thousands of dollars, the expected value initially comes out in thousands of dollars:

E[X]=\int_0^{100}x\frac{2x}{10{,}000}\,dx =\frac{2}{10{,}000}\left[\frac{x^3}{3}\right]_0^{100} =\frac{2(1{,}000{,}000)}{30{,}000} =\frac{200}{3} =66.666\ldots

The final result is $66.666\ldots$ thousand dollars, or $66,666.67 when rounded to the nearest cent. The density gives relatively more weight to larger values of $x$, so the average is greater than the midpoint of the range, which is 50 thousand dollars.

Method and mistake avoided. The method is to integrate $x f_X(x)$ over the support. Integrating $f_X(x)$ alone only checks total probability; it does not calculate an average loss.

Expected Aggregate Loss from Frequency and Severity

A frequency-severity description separates the number of claims from the size of each claim. Let $N$ be the claim count and let $Y_1,Y_2,\ldots$ be claim severities. The aggregate loss is $S=\sum_{i=1}^{N}Y_i$, with $S=0$ when $N=0$. Under the assumptions below, its expected value is the expected claim count multiplied by the mean severity.

Result: Suppose $E[N]$ is finite, the severities $Y_i$ have a common finite mean $E[Y]$, and the claim severities are independent of $N$ (with the same mean for each claim). Then $$ E[S]=E[N]E[Y]. $$ The formula concerns the expected aggregate loss over the stated period and exposure. It does not give the probability of any particular aggregate amount.

To justify the result, condition on the claim count. If $N=n$, the aggregate is a sum of $n$ severities, so its conditional mean is $nE[Y]$. Averaging this conditional mean over the possible values of $N$ gives $E[N]E[Y]$. Independence, or at least a constant conditional mean severity for each possible claim count, matters: if the expected severity changes with the claim count, the simple product need not hold.

Worked Example: Expected Loss from Claim Count and Severity

A hypothetical portfolio has the following claim-count distribution in one year. Each claim has severity independent of the count and of other claims, with the severity distribution shown. Calculate expected annual aggregate loss.

Claim count $N$	Probability
0	0.20
1	0.50
2	0.30

Severity per claim	Probability
$2,000	0.25
$4,000	0.50
$6,000	0.25

Step 1: Calculate the expected count.

$$ E[N]=0(0.20)+1(0.50)+2(0.30)=0+0.50+0.60=1.10. $$

Step 2: Calculate the mean severity.

E[Y]=\$2{,}000(0.25)+\$4{,}000(0.50)+\$6{,}000(0.25) =\$500+\$2{,}000+\$1{,}500 =\$4{,}000.

Step 3: Apply the result under the stated independence assumption.

E[S]=E[N]E[Y]=1.10(\$4{,}000)=\$4{,}400.

As a check, conditioning on the count gives expected aggregate amounts of $0, $4,000, and $8,000 for counts 0, 1, and 2. Their probability-weighted average is $0(0.20)+\$4{,}000(0.50)+\$8{,}000(0.30)=\$4{,}400$, which agrees.

Method and mistake avoided. The method uses the expected count times the mean cost per claim. Multiplying the most likely count, 1, by the mean severity would give $4,000, not the expected aggregate loss; it ignores the full count distribution.

Interpreting the Average and Avoiding Common Errors

Expected loss is a property of a probability model over a specified period, population, and loss definition. It is useful as an average-cost measure, but it is not a statement that the modeled loss will occur at that amount. A single observed loss can be far below or above the expected value. Nor does the mean alone describe how probability is distributed across outcomes.

For instance, a loss that is $0 with probability 0.99 and $100,000 with probability 0.01 has expected loss $0(0.99)+\$100{,}000(0.01)=\$1{,}000$. The most likely loss is $0, while the expected loss is $1,000. Neither number is a guarantee about a particular period. The expected value can also be an amount that cannot occur: if a loss is $0 or $10 with equal probability, its expected value is $5 even though the loss is never $5.

Keep the horizon and units consistent. An annual expected loss should not be compared directly with a monthly expected loss without converting one to the other's time basis under an appropriate model. In frequency-severity calculations, the severity mean and count must refer to compatible claim definitions and periods. A model for claim payments, for example, should not be silently combined with a severity mean defined for gross losses.

Identify the loss and its units.
State what $X$ measures and the time period or exposure to which it applies.

Check the distribution.
For a table, confirm probabilities sum to 1. For a density, confirm it is nonnegative and integrates to 1 over its support.

Weight all possible outcomes.
Use a probability-weighted sum for discrete losses, an integral for a density, or the expected-count and mean-severity product when its assumptions hold.

Interpret the result as an average.
Do not label it the most likely outcome, a guaranteed payment, or a prediction of one particular period's loss.

A frequent exam error is to average listed outcomes without their probabilities, or to use only the most likely outcome in place of the expected value. Another is to forget that a calculated mean inherits the units used in the model. Keep the probability weights, stated assumptions, and units visible in the calculation. These distinctions are aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Check Your Understanding

For each problem, calculate the requested expected loss and distinguish it from an individual outcome where relevant.

A hypothetical annual loss is $0 with probability 0.80, $3,000 with probability 0.15, and $20,000 with probability 0.05. Find its expected value and most likely outcome. (Answer: Expected loss = $1,450; most likely outcome = $0.)
A discrete loss is $2,000 with probability 0.40, $8,000 with probability 0.35, and $20,000 with probability 0.25. Calculate its expected loss. (Answer: $8,600.)
A hypothetical continuous loss $X$, in thousands of dollars, has density $f_X(x)=1/12$ for $0\le x\le12$, and zero elsewhere. Verify normalization and calculate $E[X]$ in dollars. (Answer: The density integrates to 1; $E[X]=6$ thousand dollars = $6,000.)
A portfolio has $N=0,1,3$ claims with probabilities 0.40, 0.40, and 0.20, respectively. Claim severity is independent of $N$, with severity $1,000 or $5,000, each with probability 0.50. Find expected annual aggregate loss. (Answer: $E[N]=1.00$, $E[Y]=\$3{,}000$, and $E[S]=\$3{,}000$.)
A hypothetical book has claim count $N=1,2,4$ with probabilities 0.20, 0.50, and 0.30. Independent claim severities are $5,000 or $10,000, each with probability 0.50. Find expected aggregate loss and explain why it need not be the loss in any one year. (Answer: $E[N]=2.40$, $E[Y]=\$7{,}500$, and $E[S]=\$18{,}000$; this is a distributional average, not a guaranteed annual amount.)

What You'll Learn