Skip to contents

Changelog

Source: NEWS.md

frpp 0.0.0.9000

  • Use dresidual, presidual, qresidual and rresidual as the primary residual_law arguments. Retain density, cdf, quantile and rng as compatibility aliases and preserve positional calls and saved laws. Reject conflicting pairs and keep per-operation requirements: generator-only simulation, density-only univariate likelihood/fitting, and additional CDF/quantile functions only when needed.

  • Add optional independent numeric fitting with fit_backend = "independent". FRPP owns observation preparation and calls maxLik or stats directly, using the public emhawkes parameter contract. Preserve the bridge default and distinguish fitting execution from the likelihood backend; auto selects support before optimization and records fallback reasons without retrying.

  • Register and restore numeric kernels and raw residual parameters through the public emhawkes parameter-plan contract. Preserve hfit optimization, parameter coordinates, residual freezing and saved results; callback/older-version reconstruction remains available.

  • Reuse a public prepared emhawkes kernel plan in the independent simulator, avoiding repeated validation of fixed kernel parameters and baseline. Keep changing-state/residual checks, exact paths and RNG order, and bridge default. Older event-wise API installations retain the existing checked calculation.

  • Add an independent numeric FRPP simulation event loop using public event-wise emhawkes kernel APIs. Keep bridge as the default; optional state/auto backends preserve the all-dimension residual refresh rule and keep callback/older-version support in the bridge. Backend support is selected before any random draws.

  • Add Windows, Linux and macOS package checks against CRAN and development emhawkes. Run the quick exact-result/RNG and paired-fit regression suite with both optimizer providers, recording dependency identity and test coverage. Released versions without the kernel API exercise the bridge and verify the explicit state capability error; development jobs require all state checks.

  • Document the public state API in GitHub development emhawkes 1.0.1.9000. Limit the default quick suite to its common-kernel contract; retain the separately pending simulate/nobs checks behind an explicit –extended option.

  • Add a quick joint regression runner with a fixed small fitting fixture, isolated test sessions and source/installed-code provenance. Keep saved-fit replay and expensive full re-estimation as separate validation steps.

  • Convert fspec to an independent S4 composition class with typed slots, semantic validation, component accessors and S3/S4 method dispatch. Upgrade earlier S3 specifications and fitted models while retaining list-style component access, callback bindings, calculation and RNG behavior.

  • Verify the 10,000-row migrated fit against the original emhawkes interface in three paired repetitions; document exact estimates and optimizer curvature, restored likelihoods, unavailable Wald covariance and elapsed fitting times.

  • Preserve the full original emhawkes FRPP examples in a compatibility vignette, alongside the new-interface examples. Document the planned migration while retaining simple residual-law uses and Hawkes residual diagnostics in emhawkes.

  • Add fsim() as the primary simulation function. Existing simulate() methods call it, preserving path and RNG behavior and old serialized classes.

  • Assemble numeric state-backend likelihoods using the experimental public emhawkes::exp_kernel_replay() interface, without Hawkes likelihood evaluation or full component histories when inference is not requested. This requires the emhawkes development interface; the bridge remains available with 1.0.0.

  • Use fspec() as the primary composition constructor. Keep frpp_spec() and old serialized classes compatible; add as_fspec().

  • Add ffit() with maxLik and stats optimizers, optional joint residual estimation, shared/fixed parameters, constraints and a domain validator.

  • Add fitted-model restoration, initialization replay, coefficient, likelihood, covariance, summary and diagnostic methods. Residual parameters stay fixed by default; invalid Wald covariance is withheld rather than repaired.