tseries 0.8.0

Time-series analysis for Dart: ARIMA, SARIMA and SARIMAX on the ctsa C library via FFI, automatic order selection (autoArima) and pure-Dart forecasting primitives.

History before the first public release was kept inside the Sweet Limit monorepo under the name tsforge and is not reproduced here.

0.8.0 — 2026-10-06 — first public release

A framework-agnostic time-series SDK for Dart (no Flutter dependency).

• Fixed-order models (native core). fitSarima — ARIMA/SARIMA fit and forecast (MLE, CSS, Box-Jenkins) with ctsa's fit status, σ², log-likelihood/AIC (null when not genuinely computed). fitSarimax — regression with (seasonal) ARIMA errors: named regressors, intercept, drift, regressor screening (ExogPolicy), scaling of regressors before the fit, coefficient SEs and covariance.
• Forecast ladders. arimaForecast / sarimaxForecast — point forecast, standard errors and confidence band over an ordered list of candidate orders, reporting every attempt.
• Automatic order selection. autoArima — pure Dart over the native fit, after Hyndman & Khandakar (2008): d by successive KPSS tests (kpssLag), D by the seasonal strength F_S of a classical decomposition, stepwise (ArimaSearch.stepwise) or exhaustive (ArimaSearch.exhaustive) search, AICc/AIC/BIC, AR/MA root margin (rootMargin), fit budget (maxModels). Returns the model, forecast and band, a DifferencingReport and the full search journal; AutoArimaNoModelException carries the reason and the journal when no model is selected. Written clean-room from the published methodology (doc/auto_arima_clean_room.md).
• Pure-Dart primitives. kpssLevelTest (KPSS level stationarity), classicalDecomposition / seasonalStrength, allRootsOutside / minRootModulus (root location of lag polynomials, no root finding), InformationCriterion, holtForecast (damped-trend exponential smoothing), linearForecast (OLS trend + prediction SE), simpleMovingAverage.
• Health check. checkNativeCore() returns NativeCoreHealth or throws TseriesNativeUnavailableException when the native core cannot be loaded.
• Typed errors. TseriesNumericException, TseriesInternalException, ModelTooLargeError (models whose exact-likelihood state exceeds 64 MiB), SeriesTooShortError, ExogenousColumnError.
• Native core. The ctsa C library, vendored under third_party/ctsa/ at a pinned commit, reached through a thin C shim that exports four functions (tseries_smoke, tseries_model_state_bytes, tseries_sarima_fit, tseries_sarimax_fit) and keeps every ctsa struct on the C side. Bindings are generated with ffigen; no Pointer reaches the public API. Compiled from source by a Dart build hook (hook/build.dart, native_toolchain_c) without floating-point contraction (-ffp-contract=off; /fp:precise on MSVC), with hidden symbol visibility and dead-code elimination.
• Defects in ctsa fixed or refused. Memory-safety defects, uninitialised reads, likelihood errors and leaks found in the vendored code are patched or refused before the native call; each one is documented in third_party/ctsa/PROVENANCE.md ("Local modifications").
• Licences. The package is BSD-3-Clause (LICENSE); the notices of the vendored code are in THIRD_PARTY_NOTICES; the per-file licence audit is in PROVENANCE. No LGPL code remains. One question is open: the Applied Statistics algorithms AS 154 / AS 182 (with AS 75) on every fit are distributed by the Royal Statistical Society "provided that no fee is charged" — see PROVENANCE, "Open licensing questions".
• Verification. Fits are checked against R and statsmodels reference values (airline model, SARIMAX cases, AR(1) likelihood, white-noise forecasts); clean-room replacements against scipy/numpy. Per-platform results, the FMA gate and how to re-run them are in doc/platform_verification.md; known limitations of autoArima are listed in the README, "Limitations of this version".
• Platforms. Verified: Windows x64 (MSVC), Android arm64-v8a and x86_64 (API 36 emulator), Linux x64 (Docker). Built but not run: Android armeabi-v7a. Not verified: iOS, macOS. Web is not supported (dart:ffi).