CSPA cheat sheet

Frequency = matching claim count / exposure; severity = matching loss / claim count.

Pure premium = compatible frequency times severity = loss / exposure.

Check gross/net, deductible/limit, expense, time, and population definitions before combining ratios.

Accident year groups occurrences; payment-year and other date views answer different questions.

Under an explicitly uniform pro-rata assumption, earned premium follows the elapsed covered fraction.

A supplied ultimate factor multiplies the matching reported loss measure.

A specified unpaid amount equals ultimate minus paid; define IBNR terminology explicitly.

Aggregate results can change because the coverage or exposure mix changes.

Define the row unit and join cardinality before summing policy or claim quantities.

Reconcile keys, counts, exposure, and losses before and after joins.

Missing, observed zero, unmatched, and late-entered values are not automatically equivalent.

Learn imputation, scaling, category handling, and feature selection inside training folds.

Target encoding needs appropriate out-of-fold construction and held-out separation.

Retain only information that would genuinely be available at the prediction timestamp.

A verified large claim may be a genuine tail observation rather than a removable error.

Use weights and uncertainty when sampling or group exposure sizes are unequal.

A GLM link maps conditional mean to the linear predictor; the inverse link returns the mean.

Log-link coefficient beta implies a mean ratio exp(beta) for the specified change.

An interaction changes the relevant conditional effect; include it before exponentiating.

Log-exposure offset has a specified coefficient and scales expected count with exposure.

Logit eta gives probability 1 / (1 + exp(-eta)); odds equal p / (1-p).

Poisson variance equals mean; gamma variance is proportional to squared mean under its dispersion convention.

Standard positive gamma support does not itself represent a no-claim point mass.

Penalization, collinearity, separation, and likelihood comparisons require their own assumptions.

A tree predicts within its learned regions; a squared-error leaf uses the appropriate weighted mean.

Split improvement compares parent loss with the weighted or summed child loss on a consistent basis.

Bagging resamples observations with replacement and averages fitted models.

Correlated errors reduce the variance benefit of adding more models.

Random-forest predictor subsampling helps diversify trees.

Boosting updates a fitted predictor sequentially; learning rate and stage count interact.

Early stopping uses an appropriate validation role rather than repeatedly consulting the final test.

Permutation importance, partial dependence, and local explanations are descriptive tools, not automatic causal proof.

Choose temporal or group separation to reflect deployment rather than using a convenient row split.

Tune within training/validation procedures and reserve genuinely held-out evaluation for final use.

MAE averages absolute errors; MSE averages squared errors; RMSE returns to response units.

Precision = TP/(TP+FP); recall = TP/(TP+FN); specificity = TN/(TN+FP).

ROC AUC describes ranking, whereas calibration checks observed outcomes against predicted probabilities.

F1 combines precision and recall but does not encode every decision cost.

Under the specified simple cost model, threshold = false-positive cost / sum of false-positive and false-negative costs.

Report subgroup uncertainty, immature outcomes, drift, operational constraints, and the actual decision context.

Expected claim cost is one input to pricing, not automatically the final charged premium.

A payment-horizon target is not automatically an ultimate reserve target.

A high fraud score supports appropriate review, not an accusation or proof of intent.

Review priority can depend on recoverable amount and investigation cost as well as probability.

A confidence interval for a mean is not an individual outcome prediction interval.

Removing a protected field does not remove all proxies or establish fairness.

Hashing linkable identifiers does not by itself make data anonymous.

Preserve data versions, split logic, feature code, environment, limitations, ownership, and response procedures.