Does fitting a score distribution demonstrate calibration?

No; calibration requires outcome-based evaluation.

AI Glossary / Statistics and probability / Advanced

Unit Weibull distribution

Definition

The unit Weibull distribution is a continuous distribution on the open interval from zero to one, obtained by transforming a positive Weibull variable with X = exp(-Y). It is a candidate model for bounded continuous measurements, including some proportions and normalized indices. It is not the ordinary Weibull distribution with a scale parameter set to one. The transformation changes the support and probability law, and exact zeros or ones are not assigned positive mass by the basic model.

What it means

A Weibull variable normally ranges over positive values without a finite upper bound. Applying a decreasing exponential maps those values into the unit interval: small original values become values near one, and large original values become values near zero. The resulting family offers an alternative to other bounded distributions. Whether it is useful depends on the measurement process and empirical fit. A bounded score, a proportion based on a count and a calibrated probability are different objects even when all are written between zero and one.

How it works

Using positive parameters alpha and beta, define the unit Weibull CDF as F(x) = exp[-alpha(-log x)^beta] for zero less than x less than one. Its density is alpha times beta times (-log x)^(beta minus 1) times F(x), divided by x. The quantile at probability p is exp[-((-log p)/alpha)^(1/beta)]. This parameterization corresponds to a positive variable Y with survival exp(-alpha y^beta). Maximum likelihood or other suitable methods can estimate the parameters. Finite-sample bias and model uncertainty should be examined rather than assuming favorable asymptotic properties guarantee accurate small-sample estimates.

Three practical examples

Example 1: A quality analyst compares the family with a beta distribution for a continuous bounded index. Example 2: A simulation generates unit Weibull values by applying its quantile function to uniform random numbers. Example 3: A researcher models conditional quantiles of a bounded response using a regression extension. These examples are potential applications; the basic two-parameter distribution does not automatically include predictors, endpoint masses or dependence between observations.

Think of it this way

Think of compressing an unbounded positive scale into a one-unit ruler. The exponential transformation changes both spacing and direction, so equal distances on the original scale no longer correspond to equal distances on the ruler. The new distribution describes how the transformed values are spread, rather than merely changing the original measurement's units.

Worked example

Take alpha = 2 and beta = 2. At x = exp(-1), approximately 0.367879, minus log x equals one, so F(x) = exp(-2), approximately 0.135335. The median is exp[-sqrt((-log 0.5)/2)], approximately 0.5550. Substituting this median into the CDF returns 0.5. For comparison, if beta were one with alpha still two, the CDF would simplify to exp(2 log x) = x squared, making F(0.5) equal to 0.25. This special case shows that parameters affect the distribution rather than directly representing its mean and standard deviation. All calculations use the stated rate-style parameterization; software using a Weibull scale convention may express the equivalent positive-variable parameters differently.

Illustrative case study

In this fictional case study, Aspen Scoring analyzes a continuous operational index bounded between zero and one. Analysts fit several candidate distributions and compare their predicted lower-tail probabilities with held-out measurements. They discover a small set of exact ones caused by the scoring system's cap, so they do not silently shift those observations inward. Instead, they investigate whether a separate endpoint component or another observation model is appropriate. The unit Weibull family remains a candidate for the interior values, subject to validation.

Common misconceptions and mistakes

Common mistakes include confusing unit Weibull with Weibull scale one, treating a continuous density as a probability at an exact score, and ignoring endpoints. Another mistake is assuming two parameters are the mean and variance or transferring formulas between parameterizations without checking. A good marginal fit does not establish that an AI score is calibrated. Small-sample estimation can be biased, and a model selected on the same data needs appropriate assessment before its predictive claims are trusted.

Business use

Businesses may use the family to summarize continuous bounded measurements or simulate scenarios when its shape fits the intended quantity. Compare relevant quantiles and tail behavior with other models, and keep the transformation and parameter convention explicit. If a score is a proportion formed from a numerator and denominator, consider whether a count-based model is more appropriate. If observations are repeated within customers or over time, account for that dependence rather than relying on an independent marginal distribution alone.

Check your understanding

What is F(exp(-1)) when alpha and beta are both two?

Answer: exp(-2), approximately 0.135335.

What is the median?

Answer: Approximately 0.5550.

What CDF results when beta equals one?

Answer: x raised to alpha.

Why should exact endpoint observations be investigated?

Answer: They are outside the positive-mass behavior of the basic continuous interior model and may reflect a distinct measurement mechanism.

Related terms

Quantile Regression
Related terms include Weibull distribution
beta distribution
bounded response
maximum likelihood and transformed distribution

Learn next

Learn next how change-of-variable formulas produce the density and how the quantile function supports simulation. Then compare bounded-response regression models and methods for endpoint observations. Examine finite-sample estimation and predictive diagnostics before interpreting a fitted family as a reliable description of future scores.

Sources

Sources: Mazucheli, Menezes and Ghitany, The unit-Weibull distribution and associated inference, author-uploaded article, https://www.researchgate.net/publication/324844515_The_unit-Weibull_distribution_and_associated_inference; Menezes and colleagues, Bias-Corrected Maximum Likelihood Estimators of the Parameters of the Unit-Weibull Distribution, https://www.ajs.or.at/index.php/ajs/article/view/1023. These primary papers define the model and examine estimation. The probability and quantile calculations are original and independently checked; the scoring case is fictional.

Previous: Unit vector Next: Unit-weighted regression