v0.1.3

Durbyn v0.1.3

Diff since v0.1.2

Bug-fix release.

• Fixed InexactError when fitting ARIMA models on integer-valued data with regressors
• fitted() and residuals() now work on all fitted models
• Fixed-order ARIMA formulas (e.g. p(1) + d(1) + q(1)) no longer add hidden seasonal terms
• Fixed fractional confidence levels (level = [0.8, 0.95]) being misread in Diffusion, Theta, and Kolmogorov–Wiener forecasts
• ses / holt / holt_winters accept lambda = :auto; BATS/TBATS accept integer Box-Cox bounds and series with missing values
• Croston forecasts no longer return empty fitted values; added forecast(fit; h = ...) for Croston
• auto_arima approximation default now activates correctly for high-frequency series
• New exports: PDQ, NamedMatrix, Formula, arima_rjh, croston_classic / croston_sba / croston_sbj, Theta (OTM, STM, DSTM, DOTM) and Diffusion (Bass, Gompertz, GSGompertz, Weibull) model-type constants
• Repaired ~40 outdated code examples across the README and documentation
• Added Code of Conduct, Contributing guide, and issue/PR templates

Merged pull requests:

• Bump codecov/codecov-action from 4 to 6 (#31) (@dependabot[bot])
• Bump julia-actions/setup-julia from 2 to 3 (#32) (@dependabot[bot])
• Bump codecov/codecov-action from 6 to 7 (#33) (@dependabot[bot])
• Bump actions/checkout from 6 to 7 (#34) (@dependabot[bot])

Closed issues:

• Durbyn.jl 0.1.2 (#29)
• ARIMA Inexact Errors with regressors (#35)
• Method error in Time Series Preprocessing Pipeline example (mis-named kwarg) (#36)

v0.1.2

Durbyn v0.1.2

Diff since v0.1.1

New Features

• Grammar interface for Kolmogorov-Wiener filter — The KW optimal filter forecasting model can now be specified declaratively using Durbyn's @formula DSL:
  spec = KwFilterSpec(@formula(gdp = kw_filter()))
  spec = KwFilterSpec(@formula(gdp = kw_filter(filter=:hp, lambda=1600, output=:trend)))
  spec = KwFilterSpec(@formula(gdp = kw_filter(filter=:bandpass, low=6, high=32)))

  • Panel data support for KW filter — Fit KW models to grouped data with groupby:
    grouped = fit(spec, data, m=4, groupby=:country)
  • PanelData wrapper support — KW filter works with the PanelData convenience type.

  New Types

  • KwFilterTerm — Grammar term for KW filter specification
  • KwFilterSpec — Model specification (declarative, data-independent)
  • FittedKwFilter — Fitted model with forecast() support

  New Exports

  • kw_filter() — Grammar constructor for @formula usage
  • KwFilterSpec — Model spec type

  Documentation

  • Added Grammar Interface section to kolmogorov_wiener.md with single-series and panel data examples
  • Added KwFilterSpec and kw_filter to API reference

Merged pull requests:

• CompatHelper: add new compat entry for Statistics at version 1, (keep… (#28) (@taf-society)

Closed issues:

• Durbyn v0.1.1 (#26)

v0.1.1

Durbyn v0.1.1

Diff since v0.1.0

Clean-room rewrites:

• ARIMA: Kalman filter from Durbin & Koopman (2012) Algorithm 4.3, concentrated ML from Harvey (1989) eq 3.3.16, CSS from Box-Jenkins (2015) §7.1.3, parameter transforms from Jones (1980) and Monahan (1984), covariance initialization from Gardner (1980) and Hamilton (1994).

• Optimize: BFGS from Nocedal & Wright (2006) Algorithm 6.1, L-BFGS-B from Byrd, Lu, Nocedal & Zhu (1995), Nelder-Mead from Nelder & Mead (1965) and Lagarias et al. (1998), ITP from Oliveira & Takahashi (2021).

• Stats / Unit Root Tests: ADF critical values from MacKinnon (2010) response surface regressions, KPSS from Kwiatkowski et al. (1992), Phillips-Perron from Phillips & Perron (1988) with MacKinnon (2010) finite-sample critical values, Schwert (1989) bandwidth.

• Diffusion: Bass (1969), Gompertz initialization from Jukic, Kralik & Scitovski (2004), Weibull from Sharif & Islam (1980), meta-analytic presets from Sultan, Farley & Lehmann (1990), OLS init from Srinivasan & Mason (1986), Gamma/Shifted Gompertz from Bemmaor (1994).

• Decompose: Classical decomposition by moving averages from Brockwell & Davis (2016) §1.5.2.

• STL: Seasonal-Trend decomposition using LOESS from Cleveland, Cleveland, McRae & Terpenning (1990).

New model:

• Kolmogorov-Wiener: Optimal finite-sample filter from Schleicher (2004), implementing Propositions 1–4. Supports HP, bandpass, and Butterworth filters with ARIMA-based autocovariance structure. Wiener-optimal multi-step forecasting via Toeplitz system.

Merged pull requests:

• Bump julia-actions/cache from 2 to 3 (#25) (@dependabot[bot])

Closed issues:

• JuliaRegistrator register (#21)
• Support LTS? (#23)

v0.1.0

Durbyn v0.1.0

Closed issues:

• about === nothing and !== nothing (#5)
• box-cox back transformation after model fitting: (#6)
• Show method for ets (#7)
• Forecast Plot (#8)
• Error using Durbyn (#9)
• ETS(MMM, damped = false) fails (#10)
• arima and auto_arima code refactoring (#11)
• ets seasonal index calculation in forecast() (#12)
• ets module (#13)
• arima module (nothing isnothing) code improvement (#14)
• ets forecast short time series error (#15)
• holt method should be in different to m (#16)
• Remove Plots dependency (#17)
• Package fails to compile (#18)
• LinearAlgebra.SingularException(1) error when there are 16 training elements but not 15 or 17+ [ARAR model] (#20)