NEWS

---

gratis 1.9

Major Bug Fixes

• Corrected ARIMA mean/drift conventions, MA coefficient signs, and seasonal polynomial expansion. Preserved long seasonal lags and zero-sum seasonal coefficients; unconverged MA approximations now warn.
• Corrected ARIMA innovation scaling during startup. Automatic burn-in now uses an adaptive impulse-energy estimate for repeated poles and near AR-MA cancellation. Nonconvergence raises an error requesting explicit burn-in; explicit burn-in bypasses the search without bypassing stationarity checks.
• Random stationary AR models above order two now allow every possible count of real roots instead of excluding valid stationary regimes. Sampling has full support but is not uniform over coefficient space.
• Fixed ETS parameter resampling that could hang on rejected parameters. Added bounded admissibility checks, corrected seasonal-state ordering, and required positive multiplicative states and seasonal factors.
• Target-search weights now cover the full mixture simplex, including zero weights. Unmet tolerances return the best series with a warning; callback errors propagate and nonfinite candidates are rejected.
• Made sample standardization overflow-safe, preserving exactly constant series and small variations around large offsets. Mixture covariance matrices must be finite, symmetric and positive definite; weights are normalized and log densities remain stable in distribution tails.
• Fixed single-component target indexing, zero-burn-in MAR lengths, and short mixture-series indexing. Corrected variance scaling for external GARCH paths.
• Repaired GA selection and crossover bookkeeping, scalar local optimization, parallel and optimized population updates, oversized suggestions, elitism, nonuniform mutation, and permutation ranges. Parameters, scores and stored series remain paired; supplied clusters remain caller-owned.
• Batch generation seeds once per batch, restores the previous RNG state, and preserves calendar indices in target-generated data.

Performance

• Added sparse MAR lag evaluation, vectorized zero-order mixtures, and compiled single-component AR filtering while retaining the order of random draws.
• Vectorized conditional mixture densities, skipped zero-weight conditional means, and avoided repeated scalar-mixture validation and variance allocation.
• Automatic ARIMA burn-in is calculated once per generation batch.
• Target search selects paired final scores and series without retaining unused generation histories or reevaluating final feature callbacks.

Compatibility

• Removed the bundled Shiny applications, app_gratis() launcher, and app-only suggested dependencies.
• ARIMA/ETS noise scales whose squared variance lies outside the positive normal floating-point range are rejected, rather than silently overflowing or underflowing.
• Seeded random-model draws and default ARIMA paths can differ from earlier versions because sampling and initialization have changed.

Documentation And Tests

• Updated roxygen documentation and the README with parameter conventions, initialization limits, and the distinction between external GARCH volatility and residual-coupled MAR-GARCH dynamics. Updated simulate_target() examples to use dplyr::reframe() instead of summarise() for dplyr 1.2.0.
• Added testthat regressions for numerical safety, simulation boundaries, seeded recursion, GA bookkeeping, and real parallel workers. Cross-language checks compare controlled paths and conventions; Monte Carlo checks compare reference MAR means, variances and lag-one correlations with analytic values.

gratis 1.0.7 (2024-04-10)

• Fixed dependency issue on CRAN.

gratis 1.0.5 (2023-08-28)

• Fixed the 'docType' issue.

gratis 1.0.3 (2023-03-27)

• Fixed CRAN checking issues.

gratis 1.0.0 (2022-01-14)

• Improved documentation
• Added new functions for MAR simulation: mar_model(), simulate.mar()
• Added ets_model() and arima_model() to randomly specify ETS and ARIMA models
• Added generate.ets, generate.mar, and generate.Arima to return tsibbles
• Added generate_target() and simulate_target() functions.
• Deprecated generate_msts(), generate_ts(), and generate_ts_with_target().
• Removed some redundant functions

gratis 0.2.1 (2020-10-08)

Improvements

• Added new parameter output_format, giving users an option to transform time series output data into a tsibble format.
• Update vignette content

Bug fixes

• Corrected dependencies in R code, which is namespaces in Imports field not imported from: 'rlang' 'shinydashboard'

gratis 0.2.0 (2020-08-03)

• First release.

New features

• The R package gratis (previously known as tsgeneration) provides efficient algorithms for generating time series with diverse and controllable characteristics.