NEWS.md

In reviser: Analyzing Revisions in Real-Time Time Series Vintages

reviser 0.3.1

Internal

• Removed the dependency on calculus, which CRAN has scheduled for archival. The four symbolic operators it provided (%mx%, %diff%, %sum%, %prod%) were used only in kk_equations(), to build the nlsystemfit() formulas for the Kishor-Koenig model from character matrices of parameter names. They are replaced by small internal equivalents (kk_sym_mx(), kk_sym_diff_mat(), kk_sym_prod_mat(), kk_sym_add()) scoped to that exact use. Verified to produce identical formulas to the calculus-based implementation across all models and e values. No user-visible change.

reviser 0.3.0

Breaking changes

• The two fitted-model classes now inherit from a common parent class, revision_model, and the two vintages classes from a common parent class, tbl_vintage. A fitted model therefore carries the class attribute c("kk_model", "revision_model", "list") and a vintages object c("tbl_pubdate", "tbl_vintage", "tbl_df", "tbl", "data.frame"). Ordinary use is unaffected: objects built by the package's own functions gain the parent class automatically, and the generics dispatch as before. Two situations do change.
  • Code that replaced the class attribute wholesale, as in class(x) <- c("tbl_release", "tbl_df", "tbl", "data.frame"), must now include the parent class as well. validate_vintages() reports objects that do not. Code that prepended to the existing attribute, as in class(x) <- c("tbl_release", class(x)), keeps the parent class and is unaffected.
  • Objects serialized with an earlier version no longer dispatch to the print, summary, plot or extractor methods, because their stored class attribute predates the parent class. Re-create them, or add the parent class (tbl_vintage, and revision_model for fitted models) to the stored object.
• The pillar header printed for a long-format tbl_pubdate changed: it now reports the number of distinct dates rather than the number of rows (see below), and gained the Format row that tbl_release already showed.
• The bundled gdp data set is now itself a vintages object, with class attribute c("tbl_pubdate", "tbl_vintage", "tbl_df", "tbl", "data.frame"). print(), summary(), plot() and validate_vintages() therefore work on it as loaded, rather than only after a release-extraction step. The data are unchanged; what changes is that printing gdp now shows the vintages header instead of the plain tibble header.
• vintages_long() no longer warns when it is handed long data that carry no vintages class. Attaching the class is real work there, and it is the documented way to recover the class after an operation that dropped it, so the warning made the recommended idiom noisy. Long input that is already a vintages object still warns, because the call is then a no-op.

Bug fixes

• The pillar header shown when printing a long-format tbl_pubdate reported the number of rows as the number of time periods, which overstates it once a series has more than one vintage. It now counts distinct dates, matching tbl_release and the summary() method.
• summary() on a vintages object whose columns no longer match either documented layout -- for example after dplyr::select() dropped value or the vintage key, which leaves the class attribute in place -- failed inside as.Date() with "character string is not in a standard unambiguous format". It now reports what the object is missing and points to ?validate_vintages. print() on such an object falls back to the plain tibble header instead of failing, so it can still be inspected.
• Standard errors requested with se_method = "qml" in kk_nowcast() inverted the Hessian with a general solve() and, on failure, silently applied a ridge. This path now uses the same Cholesky-based invert_hessian() helper as the other standard-error methods, so a Hessian that is not positive definite is reported rather than masked. Estimates are unchanged; standard errors change only where the previous ridge was silently applied.
• Several validation messages contained hard-wrapped newlines and source indentation, which appeared verbatim in the console. They are now single lines.
• plot() on a model fitted with return_states = FALSE failed with the base error "argument is of length zero", because it read the dropped states component without checking for it. It now reports the cause in the same words as states(), fitted(), residuals() and predict(), all of which now share a single definition of that message.
• validate_vintages() misdiagnosed a vintages object that had lost a long-layout column: because the layout was inferred from the column names alone, an object missing value was reported as a wide object whose column names were "not labeled correctly". It now reports what the object is actually missing, matching the message summary() gives for the same object. Objects whose class attribute contradicts their columns are still reported as the class mismatch they are.
• predict() on a model fitted with h = 0 returned a zero-row tibble with no explanation. It now says which argument decides that.
• The Kalman filter's stationary initial-state covariance in jvn_nowcast() fell back to a ridge-regularized solve near a non-stationary boundary without reporting it, unlike the parameter-covariance fallback added earlier in this version. The converged estimate's fitted object now carries a p0_regularized flag, which summary() reports when TRUE. Trial parameter vectors evaluated during optimization are unaffected and still regularize silently, since a momentarily non-stationary trial point is expected there and reporting it would be noise, not diagnosis.

New features

• The standard extractor methods on fitted models -- coef(), vcov(), logLik(), nobs(), fitted(), residuals() and predict() -- together with states(), print(), summary() and plot(), are now defined once for revision_model and inherited by kk_model and jvn_model, instead of being registered separately for each. The same applies to print(), summary(), plot() and pillar::tbl_sum() on tbl_vintage. See ?revision_model and ?tbl_vintage for the class contracts and the extension points a new model family or vintages representation must supply.

Documentation

• Help page titles now use a consistent title-case style throughout the package.
• The Kishor-Koenig and Jacobs-Van Norden vignettes now reach fitted models through the extractor generics -- coef(), logLik(), AIC(), BIC(), nobs() and states() -- rather than indexing into fit$params and fit$states.
• ?reviser-vintages-classes and ?tbl_vintage no longer restate the same material: the former documents the data contract and validate_vintages(), the latter the class hierarchy and the methods the parent provides.
• inst/CITATION reports the current version and title.
• The ?kk_nowcast example reached into the fitted object with result$params. It now uses coef() and logLik(), matching the vignettes and the rest of the documentation.
• ?validate_vintages gains an "Operations that drop the class" section, which names tidyr::drop_na() as the case most likely to be met in a vintages workflow and gives the idiom for recovering the class.

Internal

• The shared behavior of the two model families, and of the two vintages representations, is now expressed through S3 inheritance rather than through per-class methods forwarding to common helper functions. The model families differ only in model_family(), spec_lines(), signal_state(), target_column() and default_plot_state(); the vintages representations only in vintage_labels(), vintage_value_cols() and vintage_detail(). These are also the methods a new family or representation has to supply. Estimates, plots and printed output for kk_model, jvn_model and tbl_release objects are unchanged.
• Test coverage of the multi-series (id-aware) code paths in revisions.R, and of the revision_summary print and diagnose branches, has been substantially extended. Every method that depends on the state estimates is now tested to report return_states = FALSE as the cause.
• Comments in jvn.R no longer contain non-ASCII typographic quotes.
• The stationary initial-state covariance in jvn_nowcast() is now obtained by diagonalizing the transition matrix and solving the resulting Lyapunov equation elementwise in the eigenbasis, rather than by forming and solving the dense m^2 x m^2 linear system vec(P) = (I - T %x% T)^{-1} vec(S). This is cheaper (O(m^3) versus O(m^6)) and exploits the same transition-matrix structure -- an AR companion block plus diagonal news/noise blocks -- that motivated the earlier Cholesky and tcrossprod() changes in this version. The dense solve remains as a fallback for the rare case of a non-diagonalizable transition matrix. Estimates are unchanged.

reviser 0.2.0

Bug fixes

• summary() on a long-format tbl_pubdate no longer fails with "character string is not in a standard unambiguous format". The method assumed a wide layout and treated the pub_date and value column names as publication dates, so it failed on every get_revisions() result. The reported number of time periods and vintages was also wrong for long input.
• print() and summary() on a kk_model now report which specification was estimated. model = "Howrey" and model = "Classical" previously produced identical headers, because the fitted object never recorded the model argument. jvn_model objects likewise report whether news, noise or both were estimated.

New features

• kk_model and jvn_model objects gain the standard extractor methods: coef(), vcov(), logLik(), nobs(), fitted(), residuals() and predict(). AIC() and BIC() therefore work, and reproduce the values shown by summary().
• New states() generic to access the estimated state paths of a fitted revision model, replacing direct use of fit$states.
• New validate_vintages() checks a tbl_pubdate or tbl_release object against the documented class contract. See ?"reviser-vintages-classes" for the contract itself.

Internal

• Covariance matrices are obtained through a Cholesky factorization, which exploits the symmetry of the Hessian and reports when it is not positive definite instead of silently applying a ridge. Delta-method transformations exploit the diagonal structure of the Jacobian, and the Kalman recursions use tcrossprod(). Estimates are unchanged.

reviser 0.1.1

• Updated repository, issue tracker, and documentation links to the rOpenSci organization and docs site.
• Updated package documentation and README badges to use rOpenSci URLs.
• Added rOpenSci R-universe installation instructions to the README.
• Removed the package-specific code of conduct file in favor of the rOpenSci project-wide code of conduct.
• Disabled automatic pkgdown deployment to GitHub Pages and replaced the legacy website with a redirect page.

reviser 0.1.0

• Initial CRAN release.
• Added Jacobs-Van Norden nowcasting support via jvn_nowcast().
• Improved estimation methods and solver behavior in kk_nowcast().
• Expanded examples, tests, and documentation.

reviser documentation built on Sept. 14, 2026, 9:09 a.m.