Changes in version 1.8.2 (2026-07-06)                  

Packaging

  - Added the .Rbuildignore file so that top-level development artefacts
    (the RStudio project file, the technical notes document, local tool
    settings and cran-comments.md) are excluded from the source tarball.
    The 1.8.1 release notes stated this exclusion was already in place,
    but the file itself was missing; R CMD check --as-cran consequently
    reported NOTEs about hidden and non-standard top-level files. Both
    NOTEs are now resolved.
  - Added cran-comments.md (not shipped in the tarball) documenting the
    test environments and the resubmission context.

DESCRIPTION metadata

  - Added the maintainer's ORCID iD to Authors@R.
  - Simplified the roles of Stefanos Georganos to "aut" (the previous
    c("aut", "ctb") was redundant).
  - Added Language: en-GB, matching the British spelling used throughout
    the documentation.

Documentation

  - The \donttest{} examples of grf() and predict.grf() now set nthreads
    = 1 explicitly. This was already CRAN-compliant (ranger defaults to
    two threads) and is purely a defensive clarification.

                        Changes in version 1.8.1                        

CRAN Resubmission

  - Prepared the package for CRAN resubmission after archival. The known
    CRAN archival issue was an Rd cross-reference NOTE for unanchored
    external links to ranger; all such references now use
    \link[ranger]{ranger}.
  - Version and date updated to 1.8.1 / 2026-06-09.
  - Top-level development artefacts are excluded from R CMD build
    through .Rbuildignore.
  - Removed "active development" wording from help pages to better match
    CRAN's publication-quality expectation.

Dependency Reduction

  - The package now depends only on R in Depends: and imports ranger,
    stats, graphics and utils.
  - Removed the runtime dependencies on caret and randomForest. The two
    affected functions, rf.mtry.optim() and grf.bw(), are now
    implemented directly with ranger and base R helpers.
  - Replaced broad namespace exports/imports with explicit export(),
    S3method() and importFrom() declarations.

grf() Robustness

  - grf() no longer repeatedly refits local forests when ranger returns
    a NaN out-of-bag prediction for the focal observation. The default
    fallback is now a leave-one-out local refit: the focal row is
    removed from the local training subset and predicted out-of-sample.
    This preserves finite local diagnostics without noisy console
    output.
  - Added oob.fallback = c("loo", "inbag") to grf(). The default "loo"
    uses the new leave-one-out fallback; "inbag" reproduces the older
    in-sample fallback behaviour.
  - The focal observation is now forced to be the first row of each
    local subset before reading ranger's first training prediction. This
    avoids incorrect focal OOB extraction when two or more observations
    share the same coordinates.
  - Degenerate zero-distance local kernels now receive equal weights
    instead of producing non-finite bi-square weights.
  - The local-fit loop no longer rebuilds a full copy of dframe on every
    iteration. Local neighbourhoods are indexed by row position,
    reducing memory traffic for medium-sized data sets.
  - Added validation for coords, bw, ntree, kernel and formula inputs,
    plus clearer fixed-bandwidth diagnostics when too few neighbours are
    available.
  - grf() now returns an object of class "grf" so predict() dispatches
    to predict.grf().
  - LocalModelSummary now reports AIC and AICc for both OOB and
    predicted residuals.

Bandwidth and mtry Selection

  - grf.bw() no longer uses caret::postResample(). Its Mixed and
    Low.Local diagnostics are computed by an internal base R helper
    equivalent to squared correlation.
  - grf.bw() returns Best.BW = NA_real_ with an informative warning when
    every evaluated bandwidth produces a non-finite local R-squared,
    instead of returning numeric(0).
  - grf.bw() gains verbose = TRUE, allowing bandwidth-search progress
    messages to be suppressed in tests, vignettes and batch scripts.
  - grf.bw() now caps an over-large mtry at the number of predictors and
    uses a one-predictor-safe default.
  - rf.mtry.optim() gains the arguments num.trees, cv.repeats,
    num.threads and verbose, and supports three evaluation strategies
    via cv.method: "oob" (default), "cv", "repeatedcv".
  - rf.mtry.optim() no longer returns a caret train object. The new
    return value is a list with best.mtry, results, cv.method, num.trees
    and call. Code that read out$bestTune$mtry should now read
    out$best.mtry.

Documentation, Examples and Tests

  - Added a tests/testthat/ test suite (testthat 3rd edition) covering
    the public API of every exported function.
  - Added a vignette, vignettes/SpatialML.Rmd, with a complete workflow:
    mtry tuning, bandwidth search, GRF fit, local-importance plotting
    and prediction at new locations.
  - The vignette now uses the knitr::knitr engine with markdown output,
    avoiding the RStudio/Quarto/Pandoc probe path during local package
    checks.
  - The \donttest{} examples in grf.Rd, grf.bw.Rd, predict.grf.Rd and
    rf.mtry.optim.Rd now use small, single-threaded examples suitable
    for CRAN checks. Heavier Income demonstrations are kept in
    \dontrun{} blocks.
  - The maintainer's website (https://stamatisgeoai.eu/) is linked from
    the package help page and vignette in addition to DESCRIPTION.

                        Changes in version 1.7.0                        

CRAN compliance

  - Fixed the missing package anchors in Rd cross-references that
    triggered the CRAN archival on 2025-11-28: every \link{ranger} is
    now \link[ranger]{ranger} (in grf.Rd, grf.bw.Rd, predict.grf.Rd and
    SpatialML-package.Rd).
  - Moved the heavy run-time dependencies (ranger, caret, randomForest)
    from Depends to Imports in line with current CRAN best practice.
    Depends now contains only the minimum R version.
  - Replaced exportPattern("^[[:alpha:]]+") in NAMESPACE with explicit
    export() statements; removed the unused import(randomForest) in
    favour of a narrow importFrom().
  - Added explicit stats:: and graphics:: namespacing to all internal
    calls (predict, formula, terms, setNames, dist, sd, rnorm, rpois,
    runif, plot).
  - Added a package-level help page (?SpatialML).
  - Switched URLs and DOIs in the documentation to the canonical
    \doi{...} macro.

Bug fixes

  - grf(): pre-allocation of the local-forest list previously read
    as.list(rep(NA, length(ntrees))), producing a one-element list
    because ntrees is a scalar; fixed to vector("list", Obs).
  - grf(): replaced subset(DataSetSorted, DNeighbour <= bw) with bracket
    indexing to avoid the standard CRAN check NOTE about non-standard
    evaluation.
  - grf(): the function now returns an object of class "grf" so that
    predict(obj, ...) dispatches correctly to predict.grf() without the
    user having to call the S3 method directly.
  - random.test.data(): when dep.var.dis = "normal" the dependent
    variable is now drawn with rnorm() (was incorrectly drawn with
    runif()).
  - random.test.data(): fixed the operator-precedence bug for (i
    in 1:vars.no - 1) (which evaluated to 0:(vars.no - 1)); the
    predictor matrix is now filled vectorised and column-named X1, X2,
    ....
  - predict.grf(): replaced object[[1]] by object$Global.Model;
    validates that object$Forests exists and that x.var.name /
    y.var.name match columns of new.data; ... is now propagated to both
    global and local predict() calls.
  - rf.mtry.optim(): caret::trainControl() arguments are now passed by
    name (cv.method was previously matched positionally and could land
    in the wrong slot).

New features and improvements

  - grf(), grf.bw(), random.test.data(): added match.arg()-based
    validation for the kernel and dep.var.dis arguments and explicit
    validation of coords, bw, ntree, dataset and step.
  - grf(): improved diagnostic stop messages when the fixed bandwidth
    yields too few neighbours.
  - grf(): returned LocalModelSummary now also reports AIC and AICc for
    both OOB and predicted residuals.
  - rf.mtry.optim(): new plot.it = TRUE argument lets the user disable
    plotting when running non-interactively or in vignettes.
  - Removed every hard-coded set.seed() from inside the package
    functions (CRAN policy). Reproducibility is now under user control
    via a set.seed() call before each function.

Documentation

  - Fixed typos: "Radmom" -> "Random", "lcoal" -> "local", "Neightbours"
    -> "Neighbours", "tradition" -> "traditional".
  - grf.Rd: the \value{} block was renamed LocalModelSummary to match
    the actual returned name (was incorrectly lModelSummary).
  - grf.bw.Rd: the worked example used weighted = TRUE (a non-existent
    argument); replaced with geo.weighted = TRUE.
  - predict.grf.Rd: examples now call predict() (which dispatches via
    the new "grf" class) instead of the explicit predict.grf().
  - Expanded the Description field of every Rd file with the bi-square
    weighting equation and a step-by-step description of the algorithm.