insurancerating (version 0.8.2)

audit_refinement: Audit the effect of a fitted model refinement

Description

Compare an unrestricted GLM with the model returned by refit() on the same observed portfolio. The audit records the refinement specification and quantifies how the fitted response or fitted rate changes for the portfolio and for each final tariff-factor level.

Usage

audit_refinement(
  object,
  exposure = NULL,
  risk_factors = NULL,
  scale = c("auto", "response", "per_exposure"),
  metric = NULL
)

Value

An object of class refinement_audit. The object contains package and model metadata, the ordered refinement steps, portfolio-level results, results by risk factor and level, and the model points used in the calculation. Use summary.refinement_audit() for a concise audit report, as.data.frame() for the level results and as_gt() for a formatted table.

Arguments

object

A fitted model returned by refit(). Ordinary GLMs do not contain the stored baseline and refinement metadata required for the comparison.

exposure

Optional character string naming the exposure column. With scale = "auto", the function attempts to infer a single exposure column from the original model offset. Supply this argument explicitly when that interpretation is ambiguous.

risk_factors

Optional character vector identifying final tariff factors for the level comparison. If NULL, the final factors reported by rating_table() and available in rating_grid() are used.

scale

Character string. "per_exposure" compares fitted values per unit of exposure. "response" compares predictions on the response scale. "auto" selects "per_exposure" when one exposure variable can be identified from the model offset and otherwise selects "response".

metric

Optional character string used to describe the audited measure, for example "risk_premium", "frequency" or "average_severity". If NULL, the audit uses "fitted_rate" or "fitted_response".

Author

Martin Haringa

Details

Direct coefficient comparisons are generally not a sufficient refinement audit. A coefficient can change because the intercept or another model term changes, while the combined fitted value for a policy remains similar. audit_refinement() therefore compares predictions from the original and refined models on common observed model-point combinations.

The model points are obtained with rating_grid(). Portfolio and level results are weighted by the number of records or, when supplied, by exposure. With scale = "per_exposure", predictions that include an exposure offset are divided by total exposure after aggregation. This gives an exposure-weighted fitted rate rather than an unweighted average over unique model points.

The resulting measure should be named according to the model being audited. For a frequency model it is normally a fitted frequency; for a severity model it is a fitted average severity; and for a direct risk-premium model it can be labelled "risk_premium". A complete risk-premium comparison requires either a direct risk-premium model or an explicit combination of frequency and severity predictions.

See Also

prepare_refinement(), refit(), rating_grid(), summary.rating_refinement()

Examples

portfolio <- data.frame(
  claims = c(1, 2, 1, 3, 2, 4),
  exposure = rep(1, 6),
  risk_class = factor(c("A", "B", "A", "B", "A", "B"))
)

base_model <- glm(
  claims ~ risk_class + offset(log(exposure)),
  family = poisson(),
  data = portfolio
)

refinement <- prepare_refinement(base_model, data = portfolio) |>
  add_restriction(data.frame(
    risk_class = "B",
    risk_class_restricted = 1.15
  ))

summary(refinement)

refined_model <- refit(refinement)
audit <- audit_refinement(
  refined_model,
  exposure = "exposure",
  metric = "frequency"
)

summary(audit)
as.data.frame(audit)

if (requireNamespace("gt", quietly = TRUE)) {
  as_gt(audit)
}