IFoA SA10 key terms

A modelling approach where an algorithm learns a mapping from input features to a known outcome using labelled training examples, then predicts that outcome for new, unlabelled cases.

Techniques that find structure in data without a labelled outcome, such as grouping similar records or reducing many variables to a few underlying dimensions.

Predicting which discrete category an observation belongs to, for example whether a policyholder will lapse or a claim is fraudulent, and reporting the predicted class or its probability.

Predicting a continuous numeric outcome from one or more input variables, such as modelling claim frequency or severity as a function of rating factors.

The subset of data used to fit a model's parameters. It must be representative of the population the model will be applied to, otherwise predictions will be unreliable.

Data held back from fitting and used to tune model choices such as hyperparameters or variable selection, giving a less optimistic view of performance than training error.

Data kept entirely separate from fitting and tuning, used once to estimate how a final chosen model will perform on genuinely unseen cases.

When a model captures noise and idiosyncrasies of the training data rather than general patterns, so it fits historical data well but predicts poorly on new data.

The tension between a model too simple to capture real structure (high bias) and one so flexible it fits noise (high variance), with predictive error minimised somewhere between the two.

Adding a penalty on model complexity, such as a constraint on the size of coefficients, to reduce variance and improve out-of-sample prediction, especially with many correlated inputs.

A resampling procedure that repeatedly splits data into fitting and held-out folds to estimate out-of-sample performance and support model or hyperparameter selection.

A regression framework combining a distributional assumption for the response with a link function relating its mean to a linear combination of predictors, widely used for claim counts and amounts.

A generalised linear model for binary outcomes that models the log-odds of the event as a linear function of predictors, producing probabilities between zero and one.

A model that recursively splits data on predictor thresholds to form homogeneous groups, producing an interpretable rule set but often unstable and prone to overfitting if grown too deep.

An ensemble of many decision trees fitted on bootstrapped samples with randomised variable selection at each split, then averaged to reduce variance and improve accuracy.

An ensemble method that builds models sequentially, each new weak learner fitted to the residual errors of the current combination, typically giving strong predictive accuracy at the cost of interpretability.

Combining predictions from several models, by averaging, voting or stacking, to produce a result more stable and usually more accurate than any single component model.

A flexible model of layered interconnected units that learns complex non-linear relationships from data, requiring large samples and careful tuning and offering limited transparency.

Grouping observations so that members of a group are more similar to each other than to those outside it, used for segmentation without any predefined outcome variable.

Techniques that represent many correlated variables with fewer derived components, reducing complexity and noise while retaining most of the underlying information.

Choosing which input variables to include in a model, on grounds of predictive value, stability, cost of collection or regulatory acceptability, to improve performance and interpretability.

A measure of how much each input variable contributes to a model's predictions, useful for interpretation and challenge but not by itself evidence of a causal relationship.

The degree to which a model's reasoning can be understood and communicated to stakeholders, regulators and customers, whether intrinsically or through post-hoc interpretation tools.

A table cross-tabulating predicted against actual classes, from which accuracy, sensitivity, specificity and related performance measures are calculated for a classifier.

A plot of true positive rate against false positive rate across all classification thresholds, summarised by the area under the curve as a threshold-independent measure of discrimination.

A situation where one outcome category is far rarer than another, so overall accuracy misleads and resampling, weighting or threshold choice must be considered.

The fitness of data for its intended use, covering completeness, accuracy, consistency, timeliness and appropriate granularity, all of which directly affect model reliability.

Detecting and correcting errors, duplicates, inconsistencies and implausible values in a dataset before analysis, with all changes documented and reproducible.

Absent values in a dataset, which may arise completely at random, depend on observed variables, or depend on the unobserved value itself, each requiring different handling.

When information unavailable at the point of prediction contaminates model development, such as using post-outcome variables, producing unrealistically good performance that fails in production.

Datasets whose volume, variety, velocity or veracity exceed what traditional processing tools handle comfortably, requiring distributed storage and computation and different quality controls.

Information without a predefined tabular schema, such as free text, images, audio or sensor logs, which must be transformed before conventional statistical modelling is possible.

Methods for extracting and using information from written or spoken language, applied in insurance to documents such as claim narratives, medical reports and policy wordings.

Data transmitted from devices in vehicles or wearables, such as distance, speed, braking and location, used to refine risk pricing and offer usage-based insurance products.

Using historical data and statistical models to forecast future outcomes such as claims, lapse or fraud, supporting pricing, reserving, underwriting and marketing decisions.

A family of algorithms that improve predictive performance by learning patterns from data rather than being explicitly programmed with rules, often at the expense of transparency.

The risk of adverse consequences from decisions based on incorrect or misused model outputs, including errors, inappropriate use and failure to monitor performance over time.

Independent assessment of whether a model is conceptually sound, appropriately implemented and performing as expected, including checks on data, assumptions, outputs and limitations.

Comparing a model's predictions with actual observed outcomes over a period, to confirm the model remains adequate and to detect deterioration requiring recalibration or rebuild.

The principle that models should not produce systematically disadvantageous outcomes for protected groups, requiring scrutiny of proxies, proxies for protected characteristics and disparate impact.

A personal attribute such as age, sex, disability, race or religion that anti-discrimination law shields from unfair differential treatment in pricing or underwriting decisions.

An input that is not itself a protected characteristic but correlates strongly with one, so using it may indirectly reproduce unlawful or unfair discrimination.

Legal and regulatory obligations governing the collection, storage, use and sharing of personal data, including lawful basis, minimisation, retention limits and individual rights.

Processing personal data so that individuals can no longer be identified, including by combining datasets, which removes it from many data protection obligations if genuinely irreversible.

The policies, roles and controls governing data ownership, access, quality, lineage and use, providing accountability for analytics and supporting regulatory compliance.

Reasoned professional opinion applied where data or models are insufficient, which must be documented, challenged and balanced against model output rather than replaced by it.