ASA exams

SRM

Statistics for Risk Modeling

5 discussions

Study strategy

SRM: PCA depends on scale and is not a supervised predictor ranking

Principal components summarise variation in the predictors. They do not automatically select the directions best associated with an outcome. Interpret loadings with the preprocessing convention and the units in mind.

Worked example or practice scenario

A data matrix contains annual income in dollars and a satisfaction score from 1 to 5. Without scaling, income may dominate variance solely because of units. A correlation-based PCA answers a different question from covariance-based PCA on the raw measurements.

Try this next

Run a small example under both scale conventions and explain the difference. A component explaining 80% of predictor variance need not explain 80% of an outcome’s variation. Keep that distinction visible in your written answer.

Reading sources

Official syllabus and resources

ActNet editorial guide · October 1, 2026 · Original illustrative examples.

0 repliesOct 1

Explanation

SRM: Regularisation trades fit for stability

A penalty changes the coefficient estimation problem. Ridge shrinks coefficients using a squared penalty; lasso can set some coefficients to zero. Compare models on a consistent scale, because predictor units affect penalised estimation.

Worked example or practice scenario

A predictor measured in dollars and another measured in thousands of dollars encode the same information but have different coefficient magnitudes. Penalising raw coefficients without accounting for scale can treat the two versions differently. Standardisation can make the intended penalty comparison clearer.

Try this next

Standardise using training-set means and scales, then tune the penalty with an appropriate validation scheme. Record whether the intercept is penalised. Distinguish selecting a useful predictive model from claiming that every selected variable has a causal effect.

Reading sources

Official syllabus and resources

ActNet editorial guide · October 1, 2026 · Original illustrative examples.

0 repliesOct 1

Explanation

SRM: Cross-validation chooses a model without reusing the final test set

Training error favours flexible models. Validation estimates how a modelling procedure performs on data it did not fit. A final test set should remain outside repeated tuning decisions; otherwise it becomes another training signal.

Worked example or practice scenario

Suppose degree-2 and degree-8 polynomial models have training MSE 12 and 3, but validation MSE 14 and 29. The smaller training error does not support choosing degree 8 for prediction. Repeat a suitable resampling procedure to assess variability in that comparison.

Try this next

Fit scaling and any feature selection using only the training portion within each fold. For time-dependent observations, consider time-aware validation instead of assuming random folds imitate deployment. Explain what decision the validation result is supporting.

Reading sources

Official syllabus and resources

ActNet editorial guide · October 1, 2026 · Original illustrative examples.

0 repliesOct 1

Explanation

SRM: Statistical significance does not establish business importance

A p-value addresses a specified statistical hypothesis under the model. It does not measure the size of the effect, prove causation or guarantee useful predictions. Report effect size and uncertainty alongside the test result.

Worked example or practice scenario

A very large dataset can make a fitted 0.2% difference statistically detectable. Whether that difference justifies a business change depends on consequences, uncertainty, stability and the cost of implementation. A large p-value also does not prove that an effect is exactly zero.

Try this next

Write two separate conclusions for a model output: one about the evidence against the null hypothesis and one about the estimated effect. Identify confounding or dependence that could invalidate the model’s assumptions.

Reading sources

Official syllabus and resources

ActNet editorial guide · October 1, 2026 · Original illustrative examples.

0 repliesOct 1

Explanation

SRM: A log-link coefficient is a multiplicative effect

In a GLM with log link, a coefficient acts on the log of the conditional mean. Exponentiating gives a multiplicative effect with other predictors fixed. That is not an additive percentage-point change in a probability.

Worked example or practice scenario

If a binary predictor has coefficient .20, its fitted mean ratio is e^.20≈1.2214. With a Poisson count model and exposure offset, the expected count for exposure e is e×exp(linear predictor). Doubling exposure doubles expected count while leaving the fitted rate unchanged.

Try this next

Distinguish the exposure offset from an estimated exposure coefficient. Then explain why a logit-link coefficient instead changes odds, and why multiplying a probability directly by e^.20 can produce an invalid answer.

Reading sources

Official syllabus and resources

ActNet editorial guide · October 1, 2026 · Original illustrative examples.

0 repliesOct 1