Why Transform Losses to Uniform Values?
A goodness-of-fit test compares observed data with a candidate distribution. For a continuous candidate model, there is a useful change of scale: apply the model’s cumulative distribution function to each observation. If the candidate model is correct, the resulting values have a common reference distribution, the uniform distribution on \([0,1]\). This gives a direct way to think about how observations sit within the candidate model’s probability scale.
The transform is useful only with its assumptions understood. In particular, the candidate distribution must be fully specified for the standard reference behavior described here, and the original observations must be independent if we want the transformed observations to be independent. A set of transformed values that looks irregular is not, by itself, proof that the model is wrong; random samples vary.
“Fully specified” means every parameter needed to calculate \(F_0(x)\) is fixed. For example, an exponential candidate with mean 2 is fully specified; an exponential candidate with an unknown mean that has first been estimated from the same observations is not. This distinction matters when assessing the reference behavior of a goodness-of-fit statistic.
Why the Transform Is Uniform
Suppose first that \(F_0\) is continuous and strictly increasing. For any \(u\) between 0 and 1, let \(x_u\) be the value satisfying \(F_0(x_u)=u\). Since \(F_0\) is increasing, the event \(F_0(X)\leq u\) is the same as the event \(X\leq x_u\). Therefore:
The CDF of a uniform variable on \([0,1]\) equals \(u\) for \(0\leq u\leq1\), so this establishes the result. More generally, the probability integral transform also holds for a continuous CDF that is not strictly increasing; flat portions of the CDF carry no probability mass. The strictly increasing case makes the argument especially direct.
The transform also preserves independence when the original observations are independent. Each \(U_i\) is a function of its corresponding \(X_i\) alone. Applying separate functions to independent random variables leaves them independent. Formally, for any \(u_1,\ldots,u_n\) in \([0,1]\), independence gives:
That joint distribution is the distribution of independent uniform variables. Thus, if \(X_1,\ldots,X_n\) are independent observations from the fully specified continuous model \(F_0\), their transformed values are independent and each is uniform on \([0,1]\). If the original losses are dependent, the transforms can each be uniform without being independent.
Worked Transformations
In each example, the candidate distribution and its parameters are stated explicitly. The observations are hypothetical and treated as independent draws; the calculations illustrate the transform, not a complete hypothesis test.
Worked Example: Transforming Exponential Losses
A hypothetical portfolio is assessed against an exponential loss model with mean 2 dollars. Three observed losses are \(1\), \(2\), and \(4\) dollars. Find their probability integral transforms and order the results.
Method. For an exponential distribution with mean \(\theta\), the rate is \(1/\theta\), and the CDF for \(x\geq0\) is \(F_0(x)=1-e^{-x/\theta}\). Here \(\theta=2\), so the fully specified candidate CDF is \(F_0(x)=1-e^{-x/2}\).
For \(x=1\), \(U_1=1-e^{-1/2}=1-0.6065=0.3935\), to four decimal places.
For \(x=2\), \(U_2=1-e^{-1}=1-0.3679=0.6321\), to four decimal places.
For \(x=4\), \(U_3=1-e^{-2}=1-0.1353=0.8647\), to four decimal places.
The transformed values are \(0.3935,\ 0.6321,\ 0.8647\), in the same order as the losses because this CDF is increasing. The check is that each value lies in \([0,1]\) and increases as the loss increases. The method avoids a common mistake: using the survival probability \(e^{-x/2}\) as the PIT value. That would produce \(0.6065,\ 0.3679,\ 0.1353\), which are survival probabilities, not \(F_0(x)\).
Worked Example: Weibull Model and an Interval Probability
A hypothetical severity model has a Weibull distribution with shape parameter 2 and scale parameter 10 dollars. Its CDF is \(F_0(x)=1-\exp[-(x/10)^2]\) for \(x\geq0\). Calculate the transforms of losses \(5\) and \(15\) dollars, then find the model probability that a loss is greater than 5 and at most 15 dollars.
Method. Substitute each loss into the stated CDF. For \(x=5\), \((5/10)^2=0.25\), so \(F_0(5)=1-e^{-0.25}=1-0.7788=0.2212\).
For \(x=15\), \((15/10)^2=2.25\), so \(F_0(15)=1-e^{-2.25}=1-0.1054=0.8946\). Thus the two transformed values are \(0.2212\) and \(0.8946\), to four decimal places.
For a continuous distribution, the probability of an interval is the difference of its endpoint CDF values: \(P(5<X\leq15)=F_0(15)-F_0(5)=0.8946-0.2212=0.6734\), to four decimal places. This agrees with the difference between the transformed endpoints. The result is checked by the ordering \(0.2212<0.8946\) and by the fact that the interval probability is between 0 and 1. A common mistake is to subtract the endpoint survival probabilities in the wrong order; the CDF difference \(F_0(15)-F_0(5)\) is the probability in the interval.
Worked Example: Normal Severity Model
A hypothetical loss amount is assessed using the fully specified model \(X\sim N(100,20^2)\), where the mean is 100 dollars and the standard deviation is 20 dollars. Transform losses of \(80\) and \(130\) dollars, and calculate the model probability that a loss lies between them.
Method. Standardize each loss using \(Z=(X-100)/20\), then apply the standard normal CDF \(\Phi\). At \(x=80\), \(z=(80-100)/20=-1\), so \(F_0(80)=\Phi(-1)=0.1587\).
At \(x=130\), \(z=(130-100)/20=1.5\), so \(F_0(130)=\Phi(1.5)=0.9332\). The transforms are therefore \(0.1587\) and \(0.9332\), to four decimal places.
The probability between the losses is \(P(80<X\leq130)=F_0(130)-F_0(80)=0.9332-0.1587=0.7745\), to four decimal places. The check is that the lower loss has the lower transformed value, and the interval probability equals their difference. This calculation uses a standard deviation of 20, not a variance of 20; the stated variance is \(20^2=400\).
Using the Uniform Reference Correctly
Under the stated null model, each \(U_i\) is uniform, so it is reasonable to inspect whether transformed observations are spread over \([0,1]\) in a way consistent with that reference. For example, repeated values near 0 or near 1 can be a sign to examine the model’s lower or upper tail. A small sample can nevertheless show uneven spacing by chance. A formal conclusion requires the appropriate goodness-of-fit statistic and its calibration, not a visual impression alone.
For an independent sample from a continuous, fully specified model, transforming does not discard the model’s probability scale: small \(U_i\) values correspond to losses in the lower part of the candidate distribution, and values near 1 correspond to losses in its upper part. This connection is why probability transforms are useful in goodness-of-fit work. They put different continuous candidate models on a common scale.
Several conditions can change how the result should be used:
- Parameters estimated from the sample. If the parameters of \(F_0\) are estimated using the same observations, the transformed values are generally not independent uniform variables under the simple fully specified-model result. The fitting step affects the reference distribution used for a test.
- Truncation or censoring. If only losses above a threshold are observed, or some losses are censored, the observed data do not have the unconditional distribution \(F_0\). Transforming them as though they were an ordinary complete sample can give the wrong reference behavior.
- Rounded or discrete observations. The uniform result is for a continuous model and unrounded observations. Rounding can create ties, while a discrete model’s ordinary CDF transform is not generally uniform.
- Dependence. Marginal transforms may each be uniform under the correct marginal model, but dependence between losses remains dependence between their transforms.
Check Your Understanding
For each problem, use the stated fully specified continuous model. Give probabilities and transformed values to four decimal places.
- A hypothetical loss follows an exponential distribution with mean 4 dollars. Find the transform of a loss of 4 dollars. (Answer: \(1-e^{-1}=0.6321\))
- A hypothetical severity follows a Weibull distribution with shape 1 and scale 8 dollars, with CDF \(F_0(x)=1-\exp[-(x/8)]\) for \(x\geq0\). Find the transform of a loss of 8 dollars. (Answer: \(1-e^{-1}=0.6321\))
- A hypothetical loss follows \(N(50,10^2)\). Find the transforms of losses 40 and 65 dollars. (Answer: \(0.1587,\ 0.9332\))
- A hypothetical loss is uniform on the interval from 10 to 30 dollars. Find the transform of a loss of 22 dollars using its CDF on that interval. (Answer: \((22-10)/(30-10)=0.6000\))
- A hypothetical loss follows an exponential distribution with mean 5 dollars. Find \(P(2<X\leq7)\) by transforming both endpoints. (Answer: \(F_0(7)-F_0(2)=e^{-0.4}-e^{-1.4}=0.4237\))