IFoA SA10 flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Bias-variance trade-off

Total expected error splits into squared bias, variance, and irreducible noise. More flexible models reduce bias but raise variance, so model selection targets out-of-sample error rather than training fit.

Data leakage

Information from outside the training sample, or from the target, contaminates model fitting. Typical causes are scaling or imputing before splitting and using future information. Leakage inflates validation results.

Stratified k-fold cross-validation

Data are split into k folds that preserve the outcome class proportions. Each fold serves once as validation. It gives more stable estimates than plain k-fold when events are rare.

Multiple imputation

Missing values are replaced by several plausible draws from a conditional model, each dataset analysed separately, and results combined using Rubin's rules. Valid when data are missing at random.

MCAR, MAR and MNAR

Missing completely at random means missingness is unrelated to any variable. Missing at random means it depends on observed variables. Missing not at random means it depends on unobserved values, requiring explicit modelling.

Class imbalance metrics

Accuracy is misleading when one class dominates. Use precision, recall, F1, and precision-recall curves. Threshold tuning and class weights change decisions but not the underlying event rate.

Reproducible modelling environment

A model is reproducible when data version, code, random seeds, package versions, and compute environment are recorded. This enables independent re-runs, peer review, and regulatory audit.

Feature engineering

Transforming raw variables into model inputs, for example ratios, aggregations, binning, or date-derived tenure. It encodes actuarial judgement and often affects performance more than algorithm choice.

GLM pricing structure

A GLM links the mean to a linear predictor through a link function, with variance proportional to a power of the mean. Poisson, gamma, and Tweedie families cover frequency, severity, and aggregate loss.

Tweedie for aggregate claims

A compound Poisson-gamma distribution with a point mass at zero and continuous positive values. It models pure premium directly in one GLM, avoiding separate frequency and severity stages.

Claim-level reserving models

Instead of modelling aggregate triangles, individual claim models track case estimates, payments, and settlement. They capture operational behaviour and claim characteristics that aggregate development factors cannot represent.

Fraud detection precision-recall

Fraud is rare, so ROC curves look optimistic. Precision-recall curves and lift charts show performance at realistic event rates. Costs of false positives and false negatives should be weighed explicitly.

Lee-Carter mortality model

Log central death rates equal an age-specific profile plus a time index multiplied by an age sensitivity, plus error. The time index is usually forecast with a random walk with drift.

Customer lifetime value

Expected present value of future margins from a customer. It combines expected retention duration, usually from a survival or churn model, with expected margin per period, each validated separately.

Metamodelling capital simulations

A fast surrogate model, such as a neural network or gradient boosting model, approximates a slow simulation engine. It must be validated on out-of-sample scenarios and checked in the tail.

Model drift monitoring

Input distributions and predictive performance are tracked over time using stability indices and recalibration tests. Drift signals that the deployed model no longer reflects the current population.

Ridge and lasso regression

Ridge adds an L2 penalty, shrinking coefficients toward zero. Lasso adds an L1 penalty, which also sets some coefficients to zero for variable selection. Elastic net combines both penalties.

Gradient boosting mechanics

Trees are added sequentially, each fitted to the negative gradient of the loss from the current ensemble. A learning rate shrinks each tree's contribution, and depth and subsampling control complexity.

Random forest

Many deep decision trees are grown on bootstrap samples with random feature subsets at each split. Predictions average across trees, reducing variance relative to a single unstable tree.

ROC-AUC and Gini

ROC-AUC is the probability a random positive case scores above a random negative case. Gini equals twice AUC minus one and is common in credit and insurance discrimination reporting.

Log loss

Average negative log-likelihood of the observed outcomes under predicted probabilities. It rewards both sharpness and calibration, and heavily penalises confident wrong predictions.

Probability calibration

Predicted probabilities match observed frequencies across score bands. Check with reliability diagrams or held-out recalibration fits. Discrimination can be good while calibration is poor.

SHAP values

Shapley-value attributions from cooperative game theory assign each feature a contribution to a single prediction. Contributions sum to the prediction minus a baseline, but correlated features split credit.

Principal component analysis

A linear projection onto orthogonal components ordered by explained variance. It reduces dimensionality and collinearity but destroys direct interpretability, and it is sensitive to feature scaling.

k-means clustering

Partitions observations into k groups by minimising within-cluster sum of squares. k must be chosen in advance, results depend on initialisation, and the method favours spherical, similarly sized clusters.

Regularisation strength selection

The penalty parameter is chosen by cross-validation on training data, never on the test set. Features should be standardised first so penalties apply comparably across variables.

Protected characteristics in rating

Many jurisdictions restrict or prohibit using attributes such as age, sex, disability, or ethnicity in pricing. Proxy variables can recreate the effect indirectly, so fairness testing is needed.

Model documentation standards

Documentation records purpose, data, assumptions, methodology, validation, limitations, and ownership. It supports independent review and demonstrates compliance with actuarial professional standards.

Right to explanation

Some data protection regimes give individuals the right to meaningful information about automated decisions that significantly affect them, and sometimes a right to human intervention.

Purpose limitation principle

Personal data collected for one stated purpose should not be reused for an incompatible purpose without a new lawful basis. Reusing claims data for a new model may need assessment.

Model risk tiering

Models are graded by materiality and complexity, and validation depth, monitoring frequency, and approval level scale with the tier. Higher-tier models need independent review.

Communicating model uncertainty

Present results as ranges with sensitivity analysis and stated assumptions rather than single point estimates. Overstating precision to decision-makers undermines professional integrity.