Deriving Memorylessness For Exponential Losses

Why Memorylessness Matters for Losses

For a loss model, a threshold may represent an amount already exceeded, such as a deductible or a level used to identify large claims. A natural question is then: given that a loss has exceeded that level, how likely is it to exceed a still higher level? For an exponential loss, the answer depends only on the distance between the two levels, not on the first threshold itself.

This property is called memorylessness. It is a consequence of the exponential survival function, rather than an additional assumption about how a loss is generated. The result is useful for simplifying conditional tail calculations, but it is specific to the exponential model; it should not be assumed for other severity distributions.

Definition: A nonnegative random variable $X$ is memoryless at thresholds $s,t\geq 0$ when $P(X>s)>0$ and $$ P(X>s+t\mid X>s)=P(X>t). $$ The condition $P(X>s)>0$ is required so that the conditional probability is defined. For an exponential random variable with positive rate, this condition holds for every finite $s\geq 0$.

Deriving the Exponential Memoryless Property

Suppose $X$ is an exponential loss with rate $\lambda>0$, measured in inverse units of loss. Its survival function is $S_X(x)=P(X>x)=e^{-\lambda x}$ for $x\geq 0$. In the mean parameterization, the mean is $1/\lambda$. The conditional probability can be written as a ratio of survival probabilities because the event $\{X>s+t\}$ is contained in $\{X>s\}$.

\begin{aligned} P(X>s+t\mid X>s) &=\frac{P(X>s+t,\ X>s)}{P(X>s)}\\ &=\frac{P(X>s+t)}{P(X>s)}\\ &=\frac{e^{-\lambda(s+t)}}{e^{-\lambda s}}\\ &=e^{-\lambda t}\\ &=P(X>t). \end{aligned}

The cancellation of $e^{-\lambda s}$ is the key step: the threshold already exceeded, $s$, drops out. Thus the additional amount $X-s$, conditional on $X>s$, has the same distribution as $X$. More precisely, for every $t\geq0$,

P(X-s>t\mid X>s)=P(X>t).

This also gives a practical conditional interval formula. For $0\leq a<b$, the probability that the excess above $s$ falls between $a$ and $b$, conditional on having exceeded $s$, is $P(a<X-s\leq b\mid X>s)=e^{-\lambda a}-e^{-\lambda b}$. Because the exponential distribution is continuous, the choice of open or closed endpoints does not change these probabilities.

P(a<X-s\leq b\mid X>s) =P(X-s>a\mid X>s)-P(X-s>b\mid X>s) =e^{-\lambda a}-e^{-\lambda b}.

The formula concerns the excess over the threshold, not the original loss measured from zero. To find the conditional probability of exceeding an absolute level $c>s$, use $t=c-s$: $P(X>c\mid X>s)=e^{-\lambda(c-s)}$. This distinction is central to the examples below.

Worked Example: A Higher Liability-Loss Threshold

A hypothetical liability loss $X$ is exponential with mean $\$8{,}000$, so $\lambda=1/8{,}000$ per dollar. Given that the loss exceeds $\$6{,}000$, find the probability that it exceeds $\$13{,}000$. The severity is measured in dollars at the time of loss; no interest or mortality assumption applies.

Method. Apply the exponential survival ratio, recognizing that the additional amount is $\$13{,}000-\$6{,}000=\$7{,}000$.

First calculate the survival probabilities at both absolute thresholds:

S_X(6{,}000)=e^{-6{,}000/8{,}000}=e^{-0.75}=0.4723666, \qquad S_X(13{,}000)=e^{-13{,}000/8{,}000}=e^{-1.625}=0.1969117.

Therefore,

P(X>13{,}000\mid X>6{,}000) =\frac{0.1969117}{0.4723666} =e^{-7{,}000/8{,}000} =e^{-0.875} =0.4168620.

To four decimal places, the answer is 0.4169. As a check, the conditional probability of being between $\$6{,}000$ and $\$13{,}000$, given an exceedance of $\$6{,}000$, is $1-0.4168620=0.5831380$. The two conditional events partition losses above $\$6{,}000$, and their probabilities sum to 1.

Common mistake avoided: Using $e^{-13{,}000/8{,}000}$ alone gives the unconditional probability of exceeding $\$13{,}000$, not the probability conditional on already exceeding $\$6{,}000$.

Conditional Probabilities for a Loss Band

A question may ask for a loss to fall below a second threshold, or within a range, given that it has already exceeded a first threshold. One approach is to subtract conditional tail probabilities. Another is to work directly with the excess $X-s$. In either approach, first translate absolute thresholds into amounts above $s$.

Worked Example: Probability of a Loss Falling Below a Second Threshold

A hypothetical property loss $X$ has an exponential distribution with mean $\$12{,}000$, so $\lambda=1/12{,}000$ per dollar. Given that the loss exceeds $\$4{,}000$, calculate the probability that it is no greater than $\$13{,}000$, and verify the result using survival probabilities. The loss is considered at the time it occurs; no interest or mortality assumption applies.

Method. Conditional on $X>\$4{,}000$, the excess has the same exponential distribution as $X$. The allowed excess is at most $\$13{,}000-\$4{,}000=\$9{,}000$.

P(X\leq13{,}000\mid X>4{,}000) =P(X-4{,}000\leq9{,}000\mid X>4{,}000) =1-e^{-9{,}000/12{,}000}.

Since $e^{-0.75}=0.4723666$, the required probability is $1-0.4723666=0.5276334$, or 0.5276 to four decimal places.

For the verification, calculate the survival probabilities: $S_X(4{,}000)=e^{-1/3}=0.7165313$ and $S_X(13{,}000)=e^{-13/12}=0.3384654$. The probability of being above $\$4{,}000$ but no greater than $\$13{,}000$ is their difference, $0.7165313-0.3384654=0.3780659$. Dividing by the conditioning probability gives $0.3780659/0.7165313=0.5276334$, agreeing with the excess calculation.

Common mistake avoided: The excess bound is $\$9{,}000$, not $\$13{,}000$. Applying the survival function at $\$13{,}000$ without accounting for the $\$4{,}000$ already exceeded answers a different question.

Thresholds, Units, and Complement Checks

When the rate is supplied directly, its units must match the loss units used in the exponent. For example, if losses are measured in thousands of dollars, a rate of $0.08$ is interpreted as $0.08$ per thousand dollars. The quantity $\lambda t$ must be dimensionless. An easy check on a calculation is to confirm that a conditional tail probability decreases as the additional threshold increases, and that a conditional probability and its complement sum to 1.

Worked Example: Conditional Probabilities With Losses in Thousands

A hypothetical equipment-loss severity $X$ is exponential with rate $\lambda=0.08$ per thousand dollars. Thus $X$ is measured in thousands of dollars and has mean $1/0.08=12.5$ thousand dollars. Given $X>7$, calculate (a) the probability that $X>20$, and (b) the probability that $10<X\leq20$. No interest or mortality assumption applies.

Method. The first event requires an additional loss of $20-7=13$ thousand dollars. For the second event, the excess $X-7$ must be greater than 3 and no greater than 13 thousand dollars.

(a) By memorylessness,

P(X>20\mid X>7) =e^{-0.08(20-7)} =e^{-1.04} =0.3534547.

To four decimal places, this is 0.3535.

(b) For the interval, subtract the conditional tails at excesses 3 and 13:

\begin{aligned} P(10<X\leq20\mid X>7) &=P(3<X-7\leq13\mid X>7)\\ &=e^{-0.08(3)}-e^{-0.08(13)}\\ &=e^{-0.24}-e^{-1.04}\\ &=0.7866279-0.3534547\\ &=0.4331732. \end{aligned}

To four decimal places, this is 0.4332. As a partition check, the conditional probability of $7<X\leq10$ is $1-e^{-0.24}=0.2133721$; adding this to the interval probability $0.4331732$ and the probability above 20, $0.3534547$, gives 1.0000000 before rounding.

Common mistake avoided: The rate is per thousand dollars, so the thresholds must also be entered in thousands. Using 13,000 in $e^{-0.08(13{,}000)}$ would mix units and produce a meaningless exponent.

What Memorylessness Does—and Does Not—Say

Memorylessness means that after conditioning on an exceedance, the distribution of the additional loss is unchanged for an exponential model. It does not mean that the original loss $X$ is independent of the event $X>s$; that event directly constrains $X$. Nor does it imply that a severity model is exponential merely because one conditional probability happens to match an exponential calculation.

The result is useful when a problem gives an exponential severity and asks about multiple thresholds: conditional tail probabilities reduce to a survival probability evaluated at the difference of the thresholds. It also helps expose common misreadings. In particular, identify whether a question asks for an absolute loss level or an additional amount above the level already exceeded, and keep all amounts in units consistent with the rate.

For an absolute threshold $c>s$, use $P(X>c\mid X>s)=e^{-\lambda(c-s)}$.
For an excess interval $a<X-s\leq b$, use $e^{-\lambda a}-e^{-\lambda b}$.
For a conditional probability of being no greater than an excess bound $b$, use $1-e^{-\lambda b}$.

Key takeaway: For an exponential loss, conditioning on $X>s$ leaves an excess $X-s$ with the original exponential distribution. The conditional probability depends on the additional amount beyond $s$, not on the threshold already exceeded.

Check Your Understanding

For each question, identify the amount by which the loss threshold changes after conditioning. Probabilities are rounded to four decimal places.

A hypothetical exponential loss has mean $\$5{,}000$. Given $X>\$3{,}000$, what is the probability that $X>\$8{,}000$? (Answer: 0.3679)
A hypothetical exponential loss has mean $\$20{,}000$. Given $X>\$12{,}000$, what is the probability that $X>\$27{,}000$? (Answer: 0.4724)
A hypothetical exponential loss has mean $\$6{,}000$. Given $X>\$4{,}000$, what is the probability that $X\leq\$10{,}000$? (Answer: 0.6321)
A hypothetical exponential loss has rate $0.0001$ per dollar. Given $X>\$5{,}000$, what is the probability that $X>\$18{,}000$? (Answer: 0.2725)
A hypothetical exponential loss has mean $\$10{,}000$. Given $X>\$7{,}000$, what is the probability that the excess $X-\$7{,}000$ is greater than $\$2{,}000$ and no greater than $\$6{,}000$? (Answer: 0.2699)

What You'll Learn