CSPA key terms

A defined measure of insurance risk volume or time that must match the loss and claim-count population used in a rate.

The number of included claims per matching exposure unit, sensitive to how claims, dates, coverage, and exposure are defined.

Average loss per included claim under specified coverage and expense definitions, distinguished from loss per policy or exposure.

Loss cost per exposure, equal to compatible frequency times severity when the component definitions use the same population and basis.

Change from an earlier reported valuation toward a later or ultimate loss estimate, subject to the chosen data and development assumptions.

The entity represented by a row, such as a policy-period, claim, or payment transaction, essential for valid joins and aggregation.

Use of information in training or evaluation that would not legitimately be available under the intended prediction and validation design.

A categorical transformation based on outcomes that requires careful training-only and out-of-fold construction to limit direct response leakage.

The time at which a predictor would actually be accessible for scoring, which can differ from the event date recorded in a source.

A factor representing how observations contribute to a target population estimate when selection rates or other relevant sampling features differ.

The specified transformation connecting a conditional response mean to a linear predictor, distinct from a transformation of each observed response.

A known contribution with a specified coefficient in a linear predictor, often log exposure in an appropriately specified insurance count model.

The ratio of event probability to its complement, a scale on which a logistic coefficient produces a multiplicative conditional effect.

A model term allowing the effect of one predictor to depend on another, requiring conditional interpretation of coefficients and mean ratios.

A parameter or diagnostic concept describing variability relative to a specified mean-variance relationship, whose exact convention depends on the model family.

A tree output for observations reaching a terminal region, computed according to the response, objective, and any supplied observation weights.

An ensemble approach that fits models to bootstrap resamples and combines their predictions, often reducing instability of flexible base learners.

A sequential fitting approach that adds model contributions guided by an objective, with learning rate and model complexity affecting behavior.

A predictive importance measure based on performance change after disrupting a feature, sensitive to correlation and the evaluation design.

A summary of model predictions as selected features vary across a reference distribution, which may involve unrealistic correlated-feature combinations.

Agreement between predicted probabilities or means and observed outcomes at appropriate aggregate or conditional levels, distinct from ranking quality.

A binary discrimination measure interpretable through positive-negative score ranking with appropriate handling of ties, not a direct measure of probability accuracy.

The fraction of predicted or selected positive cases that are truly positive under the specified threshold, population, and outcome definition.

The fraction of actual positive cases selected as positive, dependent on the threshold and distinct from precision or the true-negative rate.

An evaluation design separating model-selection procedures from outer performance estimation, intended to account for tuning as part of the fitted process.

A stated measure of benefit and cost used to connect predictions with an action, requiring assumptions about capacity, consequences, and uncertainty.

An interval intended to describe variation of a future outcome under assumptions, rather than only uncertainty in an estimated conditional mean.

A field associated with another characteristic whose omission does not necessarily remove its information or resolve fairness and permitted-use concerns.

An analysis with sufficiently controlled data, code, transformations, splits, environment, and assumptions to explain and reconstruct its stated results.

A planned process for reviewing inputs, outputs, eventual outcomes, operational behavior, and response triggers with clear responsibility after deployment.