NEWS

---

fastcpd 1.3.1

• Resolve omitted model orders consistently in the R, Python, and standalone C++ generic and family APIs: AR(1), ARMA(1, 1), ARIMA(1, 1, 0), GARCH(1, 1), VAR(1), median quantile, and 100-feature KCP. Explicit invalid orders still fail validation.
• Fix plot() for change-point-only results by constructing residual and covariate layers only when detailed output is available.

fastcpd 1.3.0

• Coordinate the R, Python, and standalone C++ source interfaces at version 1.3.0. Required CI checks all three source versions, and Python release CI rejects a mismatched tag, wheel filename/metadata, or sdist filename/metadata before publication.
• Validate standalone C++ data, scalar controls, bounds, line-search values, regression shapes, and time-series orders before native dispatch. These checks run once per fit and add no work to PELT or SEN candidate loops.
• Add canonical standalone C++ detect_* wrappers for portable built-in families. Raw AR and VAR preprocessing, pure-AR ARMA/ARIMA routing, rank transformation, original-coordinate change points/residuals, and public family/order metadata now match the R and Python contracts.
• Add standalone C++ detect_kernel() / detect_kcp() with the same scalar seed stream, bandwidth fallback, random-feature generation order, default penalty, and detailed numerical output as R and Python. Random-feature preprocessing remains outside the mean-family PELT loop.
• Expose the shared Rice mean, median, block-regression, and ARMA-selection variance estimators in the standalone C++ API, including the same detailed AR-order table and AIC/BIC selections returned by R and Python.
• Expand standalone C++ regression coverage from one mean-change fixture to all 22 portable detector cases, comparing raw/trimmed change points, costs, residual matrices, and parameters against the R-generated contract.
• Add standalone C++ bootstrap and profile intervals for change points and Wald intervals for fitted parameters. All 13 shared confidence fixtures, including the advancing seeded KCP bootstrap stream, now run through the public C++ interface without adding work to detector candidate loops.
• Expand the shared R/Python numerical contract to every portable built-in detector family and the common bootstrap, profile, and Wald confidence paths. The fixtures now compare raw and trimmed change points, segment costs, residuals, parameters, and interval diagnostics. R results expose the native raw_cp_set alongside cp_set.
• Define the remaining detailed-output boundaries: multivariate residuals now use an observation-by-response matrix in both languages; VAR wrappers always accept raw unlagged series and reject the ambiguous Python-only pre-lagged p_response form; multivariate-LM Wald intervals fail clearly in both languages; and separated-binomial Wald intervals return missing standard errors instead of unstable language-dependent finite values. Rank results now report the public rank family and retain original observations in both languages, while fitted numerical details and confidence calculations use the same centered ranks.
• Keep callable multiple_epochs schedules as native R and standalone C++ extensions. Python rejects callbacks so the shared native detector remains GIL-free and no callback dispatch is added to performance-sensitive update paths.
• Make scalar Python seeds reproduce R's default Mersenne-Twister, inversion-normal, and rejection-sampling stream for kernel random features. NumPy generator objects retain their native streams. Constant-valued KCP input now uses a finite bandwidth fallback in both languages instead of producing missing R features and dense spurious change points.
• Continue the same R-compatible scalar-seed stream through segment bootstrap resampling and seeded KCP refits, producing matching confidence interval endpoints and detection rates across R and Python.
• Skip R's unused automatic regression-variance estimate for numeric Gaussian penalties, matching Python/C++ and avoiding estimator-only failures on degenerate designs. Character-criterion pure-AR ARMA and ARIMA calls now use the same variance scaling and detailed results as detect_ar.
• Document the three-language interface contract, including native callback extensions in R/C++, Python's GIL-free callback boundary, lightweight C++ result ownership, compatibility-name differences, and the installed CMake dependency requirements. Python callback errors now describe the actual language boundary instead of incorrectly calling callbacks R-only.

fastcpd 1.2.2

• Unify ARIMA detection in R and Python behind a shared native likelihood: each candidate segment is differenced independently, change-point indices stay in the original-series coordinate system, and d = 0 is identical to ARMA. The unified likelihood is zero-mean, so include.mean = TRUE is no longer supported.
• Route pure AR ARMA(p, 0) calls through the same lagged Gaussian model as the dedicated AR wrappers in both languages.
• Extend Python confidence intervals to the R-supported binomial, Poisson, quantile, and ARIMA profile families and binomial/Poisson Wald intervals. ARIMA profile intervals in both languages now reuse the shared native segment-local likelihood. Custom callback families are intentionally R-only; Python continues to expose the built-in GIL-free native families.
• Remove the R fastcpd_ts() / fastcpd.ts() compatibility umbrella in favor of detect() or family-specific wrappers. Python does not expose a fastcpd_ts counterpart and continues to require the direct APIs.

fastcpd 1.2.1

• Remove the redundant detect_time_series() / detect_ts() umbrella from the R and Python APIs. Time-series family wrappers now call detect() directly; the long-standing R fastcpd_ts() / fastcpd.ts() entry points remain available for compatibility.
• Remove the corresponding Python time_series and ts aliases; use detect(..., family=...) or a family-specific wrapper instead.

fastcpd 1.2.0

• Publish language-specific source projections at fastcpd-r, fastcpd-py, and fastcpd-cpp, with the combined source maintained in the private Depot monorepo.
• Make the unified detect*() and estimate_variance*() APIs the primary R implementations. The existing fastcpd*() and variance_*() functions remain exported compatibility aliases and are not deprecated yet.
• Add confint.fastcpd() for confidence intervals on change-point locations and fitted segment parameters, with bootstrap, profile, and Wald methods where supported.
• Add plot.fastcpd_confint() to visualize confint() results: change-point intervals as shaded bands over the data and segment parameter intervals as point ranges.
• Add Python confidence-interval helpers via fastcpd.confint() and CpdResult.confint().
• Fix WebAssembly builds by treating the used Abseil facilities as header-only instead of linking the host platform's static Abseil archive.

fastcpd 1.1.2

• Add a standalone C++ library interface (<fastcpd/fastcpd.h>) backed by the existing C++ core, with CMake package metadata and a plain Make example for non-R consumers.
• Add GitHub Actions coverage for building and running the standalone C++ library example.

fastcpd 1.1.1

• Improve performance for matrix-backed wrappers by bypassing avoidable data-frame and model-frame construction when the input is already a numeric matrix or vector.
• Reduce SEN hot-path allocations for built-in GLM-style families by using direct row access helpers, cached segment lookup, and cached Poisson log-factorials.
• Reduce copying in compiled custom-cost adapters by passing Armadillo data and parameter objects by const reference internally.
• Cache MBIC/MDL log segment-length penalties for moderate-size PELT runs.
• Refresh README benchmark output and remove the obsolete cheatsheet links.

fastcpd 1.1.0

• Add fastcpd_kcp() / fastcpd.kcp(): kernel change-point detection via random Fourier feature approximation of the RBF kernel, routed through the PELT mean path.
• Add fastcpd_rank() / fastcpd.rank(): distribution-free rank-based change-point detection; replaces each column with its centred global rank and detects location shifts via PELT mean.
• Replace RProgress dependency with a built-in tqdm-format progress bar (src/fastcpd_progress.h); remove progress from LinkingTo. Rename the user-facing parameter r.progress → show.progress; default changed from TRUE to FALSE (progress bar is off by default).
• Unify ARMA and MA C++ implementation: remove the #ifndef NO_RCPP split in src/families/arma.h and src/families/ma.h. Both families now use a single native C++ path (exact OLS for AR, diffuse Kalman ML + Nelder-Mead for ARMA/MA) in both R and Python builds, eliminating the runtime call to stats::arima().
• Python: add garch, lm, binomial, poisson, arma, ar, and arima to fastcpd.segmentation; all 13 non-custom families now supported.
• Fix NEWS.md history: fastcpd 0.99.9 (Add small shiny app) corrected to fastcpd 0.9.9.

fastcpd 1.0.8

• Add quantile regression family (fastcpd_quantile): robust change-point detection via pinball loss with IRLS fitting; resistant to outliers unlike least-squares methods.

fastcpd 1.0.7

• Change default trim from 0.05 to 0: change points near segment boundaries are no longer suppressed by default.
• Reduce GARCH example sizes: n reduced from 1501/200/300 to 401/120/150 for faster documentation builds and test runs.
• Replace boxplot with violin plot in README benchmark figure.

fastcpd 1.0.6

• Performance improvements: absl::InlinedVector for pruned set (O(1) memory vs O(n) pre-allocation), cross-platform prefetch via absl::PrefetchToLocalCache, and baseline-subtracted benchmark plot with horizontal layout.

fastcpd 1.0.5

• Add favicon.

fastcpd 1.0.4

• Update macOS CI runner from deprecated macos-13 to macos-latest.

fastcpd 1.0.3

• Remove glmnet and Matrix from Imports; lasso fitting now uses a pure C++ coordinate descent implementation with 5-fold cross-validation for lambda selection, eliminating two heavy runtime dependencies.
• Replace Matrix::nearPD() with a pure-R eigendecomposition-based nearest positive-definite projection (nearest_pd_).

fastcpd 1.0.2

• Use a fairer comparison in README.

fastcpd 1.0.1

• Add exponential family to Python package (fastcpd.segmentation.exponential).
• Fix Bazel build for Python CI: add missing fastcpd_optim.h and families/exponential.h to hdrs in src/BUILD.bazel.
• Improve README benchmark methodology: use subprocess isolation via callr::r() inside microbenchmark with baseline subtraction for fair large-scale (n = 10^8) algorithm comparison.
• Skip slow tests during coverage runs to prevent CI timeout.

fastcpd 1.0.0 (2026-06-16)

• Add exponential family (fastcpd_exponential / fastcpd.exponential) for piecewise-constant rate change detection in exponentially distributed data.
• Add support for compiled C++ cost functions via Rcpp::XPtr in family = "custom", eliminating R-call overhead in the hot PELT/SeGD loop. All three parameters (cost, cost_gradient, cost_hessian) accept tagged external pointers alongside R closures; arity is conveyed via attr(xptr, "fastcpd_cost_arity").
• Remove bundled third-party source files (ref_fastglm_*, ref_tseries*); GLM and GARCH fitting now uses internal implementations (FitGlm, FitGarch) with no external source copies.
• Move fastcpd_impl.h from inst/include/ to src/; remove inst/include/ directory entirely.

fastcpd 0.99.9

• Add JSS publication reference: Li and Zhang (2026) doi:10.18637/jss.v116.i06.
• Update CITATION to include the Journal of Statistical Software article.

fastcpd 0.20.0

• Use templates generation.

fastcpd 0.19.0

• Relicense from GPL-3.0 to Apache 2.0.

fastcpd 0.18.0

• Release wheel for Python on Linux and macOS.

fastcpd 0.17.0

• Introduce Python version.

fastcpd 0.16.3

• Move fastcpd to the top in the README.

fastcpd 0.16.2 (2025-04-25)

• Remove natural language support.
• Update mean variance cost function.
• Use larger default trim value.

fastcpd 0.16.1 (2025-03-20)

• Skip tests on CRAN.

fastcpd 0.16.0 (2025-03-13)

• Performance improvement.
• Less package dependencies.

fastcpd 0.15.0

• Remove constants.
• Add parameter variance_estimation allowing users to specify the variance or covariance matrix if known.

fastcpd 0.14.7

• Simplify methods initialization.

fastcpd 0.14.6 (2024-11-05)

• Update package comparison and remove packages breaking seed setting.

fastcpd 0.14.5

• Add pre-compiled vignettes.

fastcpd 0.14.4

• Simplify code.
• Expose fastcpd_impl API for use in other packages.

fastcpd 0.14.3 (2024-04-26)

• Skip certain tests on CRAN.

fastcpd 0.14.2

• Fix CRAN noSuggests error.

fastcpd 0.14.1 (2024-04-18)

• Fix possible numeric issues on CRAN breaking updates of other packages.

fastcpd 0.14.0 (2024-04-16)

• Correct mBIC.
• Add citation.

fastcpd 0.13.2

• Remove package check in examples.
• Use proper pruning coefficients for MBIC and MDL.
• Make residuals matrices.
• Use preprocess for faster mean change detection.
• Update examples demonstrating faster mean change.
• Replace comparison packages in vignettes with strange dependencies with results and eval = FALSE.

fastcpd 0.13.1 (2024-04-01)

• Default no pruning for lasso.
• Comment out gfpop due to https://github.com/doccstat/fastcpd/issues/10.

fastcpd 0.13.0

• Update all documentations.

fastcpd 0.12.4

• Customizable and pretty plots.

fastcpd 0.12.3

• Remove pruning parameter and replace with convexity_coef = -Inf.

fastcpd 0.12.2 (2024-03-19)

• Update vignettes.

fastcpd 0.12.1 (2024-03-12)

• Remove useless C++ codes.
• Add more debug points in C++.
• Add more examples for data well_log.
• Add detection comparison for well_log data.
• Add a variance estimator for median change.
• Deprecate winsorize_minval and winsorize_maxval.

fastcpd 0.12.0

• Add Rice estimation for ARMA model variance estimation.
• Add time comparison using Well-log data in vignettes.

fastcpd 0.11.3

• Add Rice estimator for mean change variance estimation.

fastcpd 0.11.2

• Export variance estimator function for linear models.

fastcpd 0.11.1

• Add package comparison with CptNonPar, gfpop, InspectChangepoint, jointseg, Rbeast and VARDetect.

fastcpd 0.11.0

• Note: From now on, MBIC is used as the default penalty selection for beta parameter.

• Add penalty selection criteria using

  1. BIC: (p + 1) * log(nrow(data)) / 2
  2. Modified BIC: (p + 2) * log(nrow(data)) / 2 with adjusted cost function.
  3. MDL: (p + 2) * log(nrow(data)) / 2 with adjusted cost function.

  In the mean time, a numeric value can be passed to beta as well to explicitly specify the penalty for BIC.

fastcpd 0.10.3 (2024-01-24)

• Add package check in examples and tests.

fastcpd 0.10.2

• Remove bcp according to

  Package ‘bcp’ was removed from the CRAN repository.
  
  Formerly available versions can be obtained from the archive.
  
  Archived on 2024-01-12 as email to the maintainer is undeliverable.
  
  A summary of the most recent check results can be obtained from the check results archive.
  
  Please use the canonical form https://CRAN.R-project.org/package=bcp to link to this page.
  

fastcpd 0.10.1 (2024-01-09)

• Increase test coverage.
• Use interactive() to check if the current R session is interactive.

fastcpd 0.10.0 (2023-12-21)

• Add package comparison with other packages.

fastcpd 0.9.9 (2023-12-11)

• Add small shiny app.

fastcpd 0.9.8

• Preliminary support for ARMA(p, q) model with parameter order = c(p, q) and family "arma".
• Add fastcpd.arma / fastcpd_arma for ARMA(p, q) model.
• Add adaptive increasing beta values.

fastcpd 0.9.7

• Add lower and upper parameters to denote the lower and upper bounds of the parameters.
• Add line search.
• Add hardcoded ARMA(3, 2).

fastcpd 0.9.6

• Add bitcoin and well_log data.
• Add residual calculation for mean family.
• Add plots for bitcoin data.
• Fix residual calculation for time series data when the seegments are too small.
• Handle variance estimation errors.

fastcpd 0.9.5

• Add wrapper functions of AR(p) family: fastcpd.ar / fastcpd_ar, ARIMA(p, d, q) family: fastcpd.arima / fastcpd_arima, GARCH(p, q) family: fastcpd.garch / fastcpd_garch, linear regression family: fastcpd.lm / fastcpd_lm, logistic regression family: fastcpd.binomial / fastcpd_binomial, poisson regression family: fastcpd.poisson / fastcpd_poisson, penalized linear regression family: fastcpd.lasso / fastcpd_lasso, MA(q) model: fastcpd.ma / fastcpd_ma, mean change: fastcpd.mean / fastcpd_mean, variance change: fastcpd.variance / fastcpd_variance, mean or variance change: fastcpd.meanvariance / fastcpd_meanvariance / fastcpd.mv / fastcpd_mv.
• Replace "gaussian" family with "lm".
• Add progress bar.
• Fix design matrix from formula bug.

fastcpd 0.9.4

• Fix sanity check.
• Add small AR(1) data in gallery.
• Fix VAR(p) model bug.
• Add VAR(2) example in Gallery.
• Remove commented code.

fastcpd 0.9.3

• Deprecate "vanilla" family by vanilla_percentage parameter.
• Add check utility functions.
• Add MA(4) example.
• Fix the bug when beta is updated but the old beta is still in use.
• Remove tests estimating the variance in the "gaussian" family dynamically.
• Merge beta updating into get_segment_statistics.

fastcpd 0.9.2

• Add gallery to vignettes.
• Remove cheatsheets pdf from the package.
• Use forecast package for ARIMA model.
• Use fGarch package for GARCH model.

fastcpd 0.9.1

• Wrap && around || by parentheses.
• Add ma(4) example using custom cost function.
• Add full support for AR(p), MA(q) and ARIMA(p, d, q) models.

fastcpd 0.9.0 (2023-10-19)

• Submit for CRAN update.

fastcpd 0.8.4

• Add cheatsheets.
• Refactor code and utilize the cost_function_wrapper.
• Optimize warm start.

fastcpd 0.8.3

• Add fastcpd.ts / fastcpd_ts for time series data.
• Fix pre segmengation bug for lasso.
• Fix bug related to vanilla_percentage parameter for lasso.
• Add tests with invalid family for fastcpd.ts.
• Remove the cp_only = TRUE default when the family is "custom".
• Improved plotting for "ar" and "var" families.
• Add test coverage for cp_only = TRUE and fastcpd_ts.
• Increase test coverage.
• Provide user selection when ggplot2 is not installed.

fastcpd 0.8.2

• Add cheatsheets WIP.
• Add smaller examples test for penalized linear regression.

fastcpd 0.8.1

• Add new "ar" family for autoregressive models.
• Add new "var" family for vector autoregressive models.

fastcpd 0.8.0

• Deal with the following:

  Due to the excessive calls to `glmnet` between R and C++,
  it is better to use the R implementation of `fastcpd` for lasso.
  

• Separate the use of internal C++ cost functions and user-defined R cost functions.

• Add Codecov Icicle plot in README.

• Remove cost_optim and cost_update from RcppExports.R.

• Estimate the variance in the "gaussian" family dynamically.

fastcpd 0.7.6

• Move default cost functions definition inside the fastcpd definition.
• Define constant unordered set to store the family sets.
• Avoid using length(formals(cost)) to check the number of arguments of cost function.
• Introduce an internal family "vanilla".

fastcpd 0.7.5

• Add variance estimation example in linear regression.
• Update reference page.
• Add validation for family.
• Add user selection when ggplot2 is not installed.
• Add AR(1) using forecast example in the tests.

fastcpd 0.7.4

• Update website UI.
• Update fastcpd documentation.

fastcpd 0.7.3

• Allow multiple response variables in the formula.
• Add fastcpd logo to README.

fastcpd 0.7.2 (2023-09-23)

• Add suggested package checking in tests.

• Try to solve the amazing clang-ASAN error on CRAN:

  Error in dyn.load(file, DLLpath = DLLpath, ...) :
    unable to load shared object '/data/gannet/ripley/R/test-clang/mvtnorm/libs/mvtnorm.so':
    /data/gannet/ripley/R/test-clang/mvtnorm/libs/mvtnorm.so: undefined symbol: _ZNK7Fortran7runtime10Terminator5CrashEPKcz
  Calls: <Anonymous> ... asNamespace -> loadNamespace -> library.dynam -> dyn.load
  

fastcpd 0.7.1

• Add package citation.

fastcpd 0.7.0 (2023-09-21)

• Remove C++ unit tests using catch and commented out the code since the new version of development version of Rcpp is not yet available on CRAN. Related pull request: https://github.com/RcppCore/Rcpp/pull/1274.
• Add more documentation for fastcpd method.

fastcpd 0.6.5

• Add more experiments.

fastcpd 0.6.4

• Check warning messages in tests.
• Further encapsulation of all FastcpdParameters members.

fastcpd 0.6.3

• Add CRAN release badge.

fastcpd 0.6.2 (2023-09-14)

• Address CRAN comments.
• Add more experiments.

fastcpd 0.6.1

• Address CRAN comments.

fastcpd 0.6.0

• Submit for CRAN release.

fastcpd 0.5.7

• Fix loss function for custom mean or variance change.

fastcpd 0.5.6

• Add stargazers in README.

fastcpd 0.5.5

• Add example and test for multivariate mean shift.
• Add example and test for multivariate variance change.
• Add example and test for multivariate mean and variance change.
• Add test for linear regression with multi-dimensional responses.

fastcpd 0.5.4

• Fix bug when no change point is detected.

fastcpd 0.5.3

• Add more experiments but commented out for the sake of test time without affecting the test coverage.
• Add examples in README.
• Add CRAN manual using R CMD Rd2pdf . --output=man/figures/manual.pdf --force --no-preview from stackoverflow.
• Add example for multiple epochs using custom cost functions.
• Add table of contents to README.

fastcpd 0.5.2

• Add one-dimensional linear regression example with plot.

fastcpd 0.5.1

• Prepare for CRAN release.

fastcpd 0.5.0

• Rewrite the whole package in C++ except LASSO due to the excessive calls between R and C++ of glmnet.

fastcpd 0.4.0

• Add the transition from vanilla PELT to SeN by using vanilla_percentage parameter.

fastcpd 0.3.3

• Merge the implementation of vanilla PELT and SeN.
• Encapsulate the implementation of binding new coefficients into the previous coefficients.
• Rewrite fastcpd parameters updating in C++.

fastcpd 0.3.2

• Integrate initialization and update for theta_hat, theta_sum and hessian.
• Combine theta estimation into a single function.
• Add a parameter vanilla_percentage to denote the method switching between vanilla PETL and SeN.
• Add more documentation to cp_only parameter.
• Add preparation for merging vanilla PELT with SeN.

fastcpd 0.3.1

• Add examples as tests for fastcpd.
• Rearrange C++ functions.
• Add more precondition check.

fastcpd 0.3.0

• Bump test coverage for class methods of fastcpd.

fastcpd 0.2.9

• Fix Poisson regression bug related to lfactorial.

fastcpd 0.2.8

• Make penalized linear regression estimated coefficients output sparse.

fastcpd 0.2.7

• Fix mean change example bug.
• Update documentation to redirect README to pkgdown generated webpage.
• Add contact methods and ways to file a ticket.

fastcpd 0.2.6

• Add C++ sanity check for Logistic regression data, i.e. binomial family.
• Add examples as tests for fastcpd.
• Rename C++ source files to follow Unix convention.
• Update documentation link in README.

fastcpd 0.2.5

• Hide internal functions from the documentation.
• Export fastcpd class.

fastcpd 0.2.4

• Add column name for thetas slot in fastcpd class.
• Fix plot where residuals and responses appear in the same plot.
• Default cp_only to FALSE.
• Remove residuals from summary method.

fastcpd 0.2.3

• Add missing examples for linear regression and LASSO.

fastcpd 0.2.2

• Add more examples to illustrate the use of the fastcpd function.
• Indicating internal functions so that the users should not use them.

fastcpd 0.2.1

• Add more examples to the README.

fastcpd 0.2.0

• Added a NEWS.md file to track changes to the package.