insurancerating 0.8.1

Overview of changes since 0.8.0

• The excess-loss workflow has been made more portfolio-oriented and easier to audit. Threshold assessments now preserve input column names, support portfolio-level claim counts and can be presented with as_gt(). redistribute_excess_loss() now combines capping and redistribution in one step and returns adjusted claim amounts for severity modelling.
• Large-loss cost is redistributed over claim-bearing rows using portfolio, risk-factor or partial credibility redistribution. The total observed claim cost is preserved. Users can include the allocation in one severity response or retain it as a separate excess-loading component.
• add_portfolio_experience() is now the primary helper for attaching observed portfolio experience to rating_table() objects. The former add_observed_experience() interface remains available as a deprecated compatibility wrapper.
• The documentation for rating_table(), add_prediction() and the excess-loss workflow has been expanded with more applied portfolio examples. The pkgdown configuration and website build workflow have also been updated.

Main API updates

• add_smoothing() now stops early with the affected column and missing-value count when model_variable contains NA. add_prediction() preserves NA predictions and reports their number in a warning.
• add_restriction() and deprecated restrict_coef() can include tariff levels that were not observed when the GLM was fitted. Their supplied relativities are treated as explicit tariff assumptions. Set allow_new_levels = FALSE for strict matching against observed levels.
• add_restriction() can add a fixed, expert-specified tariff factor that is present in the refinement data but absent from the fitted GLM. This requires the explicit opt-in allow_new_risk_factors = TRUE; the default remains FALSE. Deprecated restrict_coef() keeps its historical permissive behaviour and therefore uses TRUE by default.
• Models returned by refit() now print a concise refinement summary followed by the regular GLM output. The stored call is shown as a readable glm(...) call instead of the internal function definition.
• add_relativities() now uses coefficients from an earlier add_restriction() step automatically. The requested model variable and the effective restricted variable remain separately recorded, restrictions are not applied twice, and refinement-step ordering remains explicit. The final rating_table() reports only the derived split variable rather than also showing the intermediate restricted variable.
• Refinement documentation now distinguishes the editable rating_refinement specification from the fitted GLM returned by refit(). Functions that add or edit refinement steps reject ordinary and refitted GLMs with an actionable error. Iterative smoothing therefore retains the refinement specification and calls refit() after each adjustment.
• The add_relativities() example now demonstrates that several original model levels can be refined in one relativities() specification by using multiple split_level() calls.
• Deprecated rating_factors() again preserves the dynamic estimated-column name produced by rating_table(), and deprecated fit_gam() correctly forwards a user-supplied amount column.

Rating tables

• add_portfolio_experience() is now the primary API for enriching a rating_table() object with observed portfolio experience.
• add_observed_experience() is deprecated and remains available as a compatibility wrapper.
• add_portfolio_experience() can calculate portfolio experience automatically from portfolio data for all risk factors in a rating_table(), and accepts multiple existing factor_analysis() objects.
• autoplot.rating_table() accepts metric to choose the attached portfolio experience metric at plot time.

Excess-loss workflow

• assess_excess_threshold() now returns class "threshold_assessment", preserves the original group and exposure column names, supports an optional claim-count column and uses as_gt() for report-ready threshold comparisons.
• redistribute_excess_loss() replaces the former multi-step decomposition, allocation and loading workflow. It returns capped, excess, redistributed, adjusted and adjusted-average claim amounts using names derived from the supplied claim-amount column.
• Redistribution is claim-count weighted. Rows without claims remain zero, and the adjusted loss reconciles to the original observed loss.
• redistribute_excess_loss() now accepts redistribution_weight for risk-sensitive shares and receives_redistribution for selecting which claim-bearing rows receive redistributed excess loss.
• redistribute_excess_loss() can now identify large losses that should remain unadjusted and should not contribute their excess to the redistribution pool.
• The redistribution choice is now expressed through redistribution_method = "portfolio", "risk_factor" or "partial".
• Results now inherit from "excess_redistribution". Their summary() method audits contributed and received excess loss, net loss shifts and changes in average claim amount by risk factor or another selected portfolio column.
• redistribute_excess_loss() now exposes the risk-factor loading, credibility, portfolio loading, blended loading, preservation factor and final redistribution loading. Set calculation_details = FALSE to omit these row-level audit columns from modelling data; summary() retains the complete calculation audit.
• redistribute_excess_loss() now supports output = "excess_loading" in addition to the backwards-compatible "redistributed_claim" default. The new mode keeps capped severity separate and returns allocated excess loss plus an excess loading per unit of redistribution_weight.

insurancerating 0.8.0

Main API updates

Portfolio analysis

• factor_analysis() is now the primary function for univariate/factor-level portfolio analysis. It returns objects with primary class "factor_analysis" while retaining "univariate" for compatibility.
• plot_severity_distribution() was added for exploratory severity diagnostics by category. It shows individual claim observations with mean and median markers, optional direct labels, and optional firebrick highlighting for claims above a user-supplied threshold.
• univariate() is deprecated and remains available as a compatibility wrapper. The old NSE interface is still supported through the deprecated wrapper.
• factor_analysis() now validates metric columns and grouping variables early, with clearer error messages for missing columns.
• Metrics with a zero or missing denominator now return NA_real_ instead of Inf or NaN.
• autoplot.factor_analysis() is the primary plot method. The deprecated show_plots argument has been replaced by metrics.
• The factor-analysis plot keeps the established package styling and now uses a consistent grid, axis-line, tick and secondary-axis style.

Outlier histograms

• outlier_histogram() has clearer argument names: lower, upper, density, bar_fill, bar_color, tail_fill, tail_color, and density_color.
• Deprecated argument names remain supported: left, right, line, fill, color, and fill_outliers.
• The default colours now align with the package palette: light grey bars, orange tail bins, and blue density curve.
• Input validation has been expanded for missing columns, non-numeric variables, invalid cutoffs, invalid colours, invalid bin counts, constant variables and all-missing variables.
• histbin() is deprecated and remains available as a compatibility wrapper.

Risk factor GAMs and tariff segmentation

• risk_factor_gam() is the primary spelling for fitting GAMs to continuous risk factors. riskfactor_gam() and fit_gam() remain available for compatibility, with fit_gam() deprecated.
• risk_factor_gam() returns objects with primary class "risk_factor_gam"; compatibility classes are retained for older code.
• model = "pure_premium" replaces the older model = "burning" wording. The old value remains supported with a lifecycle warning.
• Input validation and documentation for frequency, severity and pure-premium GAMs have been improved.
• derive_tariff_segments() replaces construct_tariff_classes() as the primary API for deriving tariff segments from a fitted risk-factor GAM.
• derive_tariff_segments() returns objects with primary class "tariff_segments".
• add_tariff_segments() can add derived tariff segments back to a portfolio.
• construct_tariff_classes() remains available as a deprecated compatibility wrapper.
• Split extraction now handles decimal split points correctly.
• Tree-fitting and split-extraction failures now fail clearly instead of silently returning one broad tariff interval.
• The tariff-segment plot now recognises confidence interval columns produced by the package’s own GAM output.

Rating tables

• rating_table() is the primary API for interpreting fitted GLM coefficients in tariff-table form. rating_factors() and rating_factors2() are deprecated wrappers.
• rating_table() now returns objects with primary class "rating_table" while retaining the legacy "riskfactor" class for compatibility.
• exposure_output replaces the older exposure_name argument.
• significance replaces the older signif_stars argument.
• Deprecated rating-table wrappers and plotting code have been separated more clearly in the source structure.
• add_observed_experience() was added to attach factor_analysis() output to a rating_table() object before plotting. This replaces the earlier direct univariate_* arguments in autoplot.rating_table().
• autoplot.rating_table() now plots attached observed experience from add_observed_experience() and uses cleaner, package-consistent plot styling, including a subtle secondary exposure axis.

Prediction helpers

• add_prediction() now has clearer naming arguments: predictions, prefix, confidence, and interval_names.
• var and conf_int are deprecated in favour of predictions and confidence.
• Confidence interval columns now use _lower and _upper suffixes by default.
• The function now validates alpha, confidence settings, duplicate output names, name collisions with existing columns, missing models and non-GLM inputs.

Model data and rating grids

• extract_model_data() replaces model_data() as the primary API for extracting model data from fitted models.
• model_data() is deprecated and remains available as a wrapper.
• rating_grid() now uses base R internally and returns a regular data.frame.
• Plain GLM metadata extraction has been improved so rating_grid(glm) groups by model terms as expected.
• as_gt() now formats rating_table objects as grouped tariff tables, with optional significance stars and locale-aware formatting for fitted effects and exposure.
• Refinement metadata is now joined by the related original/new factor columns instead of being cross-joined onto every rating-grid row.

Model refinement

• The refinement API has been clarified around prepare_refinement() |> add_*() |> refit().
• prepare_refinement() now explains row-count differences caused by omitted model observations. Its error identifies missing or non-finite model inputs, including the source variables used inside transformed model terms.
• A large-loss workflow was added for adjusted severity modelling: assess_excess_threshold() and redistribute_excess_loss().
• assess_excess_threshold() compares candidate large-loss thresholds and shows the impact on excess loss, capped loss and pure premium.
• redistribute_excess_loss() caps claim amounts and redistributes the observed excess burden across claims before a severity GLM is fitted. It supports portfolio, risk-factor and partial redistribution.
• Automatic credibility in redistribute_excess_loss() uses the transparent formula Z = n / (n + credibility_threshold) with credibility_basis = "claims" or "excess_records".
• Redistribution always preserves the total observed claim cost and returns an adjusted average claim amount suitable for a claim-count-weighted severity GLM.
• add_smoothing() now uses model_variable and source_variable as the primary argument names.
• GAM and shape-constrained smoothing now validate the requested basis dimension against the number of unique grouped covariate values before model fitting. Polynomial smoothing similarly validates degree, with actionable errors when the requested curve is too complex for the available model levels.
• autoplot.rating_refinement() accepts x_max and y_max for smoothing steps, allowing users to limit the visible axis ranges without changing the smoothing fit or the underlying refinement data.
• edit_smoothing() now uses clearer in-object editing arguments for adjusting smoothing settings without supplying an external data frame.
• add_restriction() can now accept a partial restriction data frame. Missing levels are automatically filled with the already fitted GLM relativities, so users can adjust only selected levels.
• add_relativities() now uses model_variable and split_variable.
• relativities() replaces relativities_list() as the helper for building relativity specifications.
• restrict_coef(), smooth_coef() and refit_glm() remain deprecated compatibility wrappers and now link clearly to add_restriction(), add_smoothing() and refit().
• autoplot.rating_refinement() no longer carries an experimental badge and uses the package plot theme.
• The refinement documentation has been expanded with applied examples and a clearer explanation of smoothing, restrictions, relativities and refitting.

Reference levels

• set_reference_level() replaces biggest_reference() as the primary helper for choosing factor reference levels.
• The default method is method = "largest_weight".
• A manual level argument was added so a specific reference level can be selected explicitly.
• biggest_reference() remains available as a deprecated compatibility wrapper.

Time utilities

• split_periods_to_months(), merge_date_ranges() and active_rows_by_date() now avoid mutating caller-visible input data.
• active_rows_by_date() replaces rows_per_date() as the primary API for matching event dates, such as claim dates, to active portfolio rows.
• period_to_months(), rows_per_date() and reduce() remain available as deprecated compatibility wrappers.
• Date interval validation, column validation, aggregation validation, nomatch validation and mult validation have been improved.
• R CMD check notes from data.table helper columns in active_rows_by_date() have been resolved.

Model validation and performance

• bootstrap_performance() now has an explicit metric = "rmse" argument.
• sampling = c("bootstrap", "split") was added to distinguish bootstrap out-of-bag evaluation from split validation.
• Deprecated arguments n and frac remain supported as aliases for n_resamples and sample_fraction.
• Character and factor rating variables are handled more robustly across resamples so prediction does not fail when a level is absent from a training sample.
• autoplot.bootstrap_performance() now uses a package-consistent visual style: subtle grey histogram, transparent blue density, orange original-model reference line, subtle confidence interval lines and no gap between the bars and x-axis.
• bootstrap_rmse() remains available as a deprecated compatibility wrapper and returned objects retain class "bootstrap_rmse" for older code.
• check_overdispersion() now validates non-GLM input, checks for Poisson models and fails clearly when residual degrees of freedom are not positive.
• print.overdispersion() now bases its conclusion on the original p-value rather than a rounded display value.
• check_residuals() now validates inputs, uses all scaled residuals for the KS test, handles empty residual vectors clearly and documents the DHARMa-based residual workflow for actuarial users.
• autoplot.check_residuals() now has a controllable max_points argument and uses ASCII messages and the package plot theme.

Truncated severity distributions

• fit_truncated_severity() replaces fit_truncated_dist() as the primary API for fitting distributions to truncated claim severities.
• Returned objects use primary class "truncated_severity" while compatibility with "truncated_dist" is retained.
• fit_truncated_dist() remains available as a deprecated compatibility wrapper.
• Observations outside the truncation interval now fail clearly instead of only warning and continuing.
• Validation has been expanded for truncation bounds, optimisation starts, grid sizes, reporting options and random generator arguments.
• Public random generators rlnormt() and rgammat() now validate sample size, distribution parameters, finite intervals and positive truncation mass.
• Plot argument names were modernised to ecdf_geom, x_label, y_label, show_title, digits and truncation_digits, with old names supported for compatibility.

Fisher-Jenks classification

• fisher_classify() and fisher() are deprecated because Fisher-Jenks classification is a general-purpose grouping method and is not directly tied to the insurance-rating workflow.
• classInt moved from Imports to Suggests.

Documentation, website and tests

• The README and vignettes have been revised to present the package as a set of actuarial pricing building blocks rather than a prescribed pricing method.
• The former “Pricing principles” vignette was replaced by “Pricing workflow building blocks”.
• The refinement vignette was rewritten with a more practical tone and current API examples.
• pkgdown reference sections were reorganised; deprecated functions are grouped under “Deprecated” and internal S3 methods are no longer listed as primary reference topics.
• New and expanded tests cover tariff segmentation, rating tables, observed experience plotting, refinement workflows, model-data extraction, model performance, overdispersion, residual checks, outlier histograms, truncated distributions, time utilities and factor analysis.

insurancerating 0.7.5

• rating_factors() now always returns correct output when column with exposure in data is not named exposure
• intercept_only in update_glm() is added to apply the manual changes and refit the intercept, ensuring that the changes have no impact on the other variables.
• smoothing in smooth_coef() is added to choose smoothing specification
• The README has been revised

insurancerating 0.7.4

• bootstrap_rmse() now uses after_stat(density) instead of the deprecated dot-dot notation
• custom_theme in autoplot.univariate() is added to customize the theme

insurancerating 0.7.3

• autoplot.univariate() now generates a plot even when there are missing values in the rows
• rating_factors() now always returns the correct coefficients when used on a ‘refitsmooth’ or ‘refitrestricted’ class of GLM.

insurancerating 0.7.2

• update_glm() now always returns the correct interval in case the function is used in combination with smooth_coef()

insurancerating 0.7.1

• rotate_angle in autoplot.univariate() is added to rotate x-labels
• univariate() now accepts external vectors for x; vec_ext() must be used

insurancerating 0.7.0

• smooth_coef() now gives correct results for intervals with scientific notation
• reduce() now returns no errors anymore for columns with dates in POSIXt format

insurancerating 0.6.9

• refit_glm() is renamed to update_glm()
• construct_model_points() and model_data() are added to create model points

insurancerating 0.6.8

• show_total in autoplot.univariate() is added to add line for total of groups in case by is used in univariate(); total_color can be used to change the color of the line, and total_name is added to change the name of the legend for the line
• rating_factors() now accepts GLMs with an intercept only
• fit_truncated_dist() is added to fit the original distribution (gamma, lognormal) from truncated severity data
• join_to_nearest() now returns NA in case NA is used as input

insurancerating 0.6.7

• smooth_coef() now returns an error message when intervals are not obtained by cut()
• get_data() is added to return the data used in refit_glm()

insurancerating 0.6.6

• summary.reduce() now gives correct aggregation for periods “months” and “quarters”
• rows_per_date() is added to determine active portfolio for a certain date

insurancerating 0.6.5

• smooth_coef() and restrict_coef() are added for model refinement
• histbin() now uses darkblue as default fill color

insurancerating 0.6.4

• In summary.reduce(), name can be used to change the name of the new column in the output.
• Dataset MTPL now contains extra columns for power, bm, and zip.
• Some functions in insight are renamed, therefore insight::format_table() is replaced with insight::export_table().

insurancerating 0.6.3

• fit_gam() for pure premium is now using average premium for each x calculated as sum(pure_premium * exposure) / sum(exposure) instead of sum(pure_premium) / sum(exposure) (#2).
• histbin() is added to create histograms with outliers
• reduce now returns a data.frame as output

insurancerating 0.6.2

• check_normality() is now depreciated; use check_residuals() instead to detect overall deviations from the expected distribution
• rating_factors() now shows significance stars for p-values
• period_to_months() arithmetic operations with dates are rewritten; much faster
• univariate() now has argument by to determine summary statistics for different subgroups

insurancerating 0.6.1

• univariate_all() and autoplot.univ_all() are now depreciated; use univariate() and autoplot.univariate() instead
• check_overdispersion(), check_normality(), model_performance(), bootstrap_rmse(), and add_prediction() are added to test model quality and return performance metrics
• reduce() is added to reduce an insurance portfolio by merging redundant date ranges

insurancerating 0.6.0

• label_width in autoplot() is added to wrap long labels in multiple lines
• sort_manual in autoplot() is added to sort risk factors into an own ordering
• autoplot() now works without manually loading package ggplot2 and patchwork first
• rating_factors() now returns an object of class riskfactor
• autoplot.riskfactor() is added to create the corresponding plots to the output given by rating_factors()

insurancerating 0.5.2

• autoplot.univ_all() now gives correct labels on the x-axis when ncol > 1.

insurancerating 0.5.1

• A package website is added using pkgdown.
• construct_tariff_classes() and fit_gam() now only returns tariff classes and fitted gam respectively; other items are stored as attributes.
• univariate_frequency(), univariate_average_severity(), univariate_risk_premium(), univariate_loss_ratio(), univariate_average_premium(), univariate_exposure(), and univariate_all() are added to perform an univariate analysis on an insurance portfolio.
• autoplot() creates the corresponding plots to the summary statistics calculated by univariate_*.

insurancerating 0.5.0

• construct_tariff_classes() is now split in fit_gam() and construct_tariff_classes().
• A vignette is added on how to use the package.

insurancerating 0.4.3

• period_to_months() is added to split rows with a time period longer than one month to multiple rows with a time period of exactly one month each.

insurancerating 0.4.2

• In construct_tariff_classes(), model now also accepts ‘severity’ as specification.