Changelog

This changelog includes end-user changes only. For internal development notes, see the repository changelog.

[5.1.0]

Changed

  • Breaking change: OnlineLowess with missing = "drop" now returns :dropped for non-finite points instead of nothing, so drops are distinguishable from a still-filling window.
  • Online fits sort the sliding window by x before each update so non-monotonic streams stay correct.
  • process_chunk rejects chunks that are not non-decreasing in x when merged with the previous overlap; feed ascending chunks or pre-sort the series.
  • append! on LowessResult now rejects results that include cv_scores, diagnostics, or predict_model; only streaming-style per-point fields can be concatenated.

Fixed

  • Streaming no longer panics when derivatives were requested but missing; it throws instead.
  • Extend/Zero boundary pads use unit spacing when leading/trailing x ties would otherwise stack pads.
  • Fixed retained-model predict_handle ownership so error paths no longer leak the handle and successful conversion no longer risks double-free.
  • Free native result buffers with try/finally, and clear finalizer-owned handles after free.
  • Fixed FFI length fields truncating on 64-bit Windows by using pointer-sized sizes instead of Culong.

[5.0.0]

Added

  • Enabled the optional wgpu DirectX 12 backend for GPU-enabled Windows builds. DXC is loaded dynamically with an FXC fallback, avoiding eager imports of dxcompiler.dll and dxil.dll.
  • Added GLES for Android GPU builds with a target-scoped wgpu feature; Windows continues to omit the GLES-only loader imports.
  • Added grouped outputs and cv keywords for Lowess, grouped outputs for Streaming and Online constructors, and retained-model prediction.
  • Added a grouped intervals keyword with residual-bootstrap intervals for Lowess, StreamingLowess, full-update OnlineLowess, and predict.
  • Added an "Alternative Software" guide comparing FastLOWESS.jl with the general-purpose LOESS implementation in Loess.jl.

Changed

  • Clarified that Batch residual_sd is 1.4826 * MAD, while Streaming reports the cumulative sample standard deviation of emitted residuals.
  • Breaking change: replaced individual return_* output keywords with outputs=[...] for Batch, Streaming, Online, and prediction.
  • Breaking change: replaced flat interval keywords and predict's interval levels with intervals=(confidence=..., prediction=..., bootstrap=...), and replaced flat CV keywords and the nested cv seed with cv=(method=..., k=..., fractions=...) plus one outer seed shared by CV and bootstrap. predict has its own seed; seed=0 is a valid seed.
  • Represent unavailable diagnostic metrics as nothing instead of NaN sentinels.

Fixed

  • Fixed CPU Gaussian standard errors for fitted and queried values to use the full unbounded kernel support.
  • Made even-sample medians and mean-absolute residual scaling overflow-resistant for large finite values, and centered Batch/Streaming R-squared accumulation to preserve one-ULP response variation at large offsets.
  • Made Batch RMSE, MAE, and R-squared reductions scale-safe for large finite values, preserved bisquare downweighting when tuned scales exceed the numeric range, and normalized custom weights before summation in local and all-tied fits.
  • Kept AIC finite when raw residual-square sums overflow, and made local/all-tied WLS invariant to common scaling of large finite case weights.
  • Fixed global OLS fits treating predictor values with a large offset as degenerate; translated inputs now retain their fitted slope.
  • Fixed fraction-1 global fits ignoring custom weights, including when Batch sorts observations by x.
  • Fixed Batch missing = "drop" accepting custom weights with a length different from the original input; weights are validated before rows are dropped.
  • Fixed standard errors for global weighted fits to account for observation weights and weighted prediction leverage.
  • Fixed local Batch and retained-model prediction standard errors ignoring custom_weights; local SE moments now include the per-observation case weights.
  • Fixed global fits with all-zero custom weights to honor the configured zero-weight fallback policy.
  • Fixed fraction-1 global fits ignoring configured robustness iterations; they now reweight observations and report iterations used.
  • Fixed Batch cross-validation candidate fits ignoring custom_weights; K-fold CV now rejects more folds than observations instead of returning zero scores.
  • Fixed Streaming and Online accepting invalid auto_converge tolerances; Online also rejects invalid explicit delta values while retaining NaN as its default sentinel.
  • Fixed GPU Batch fits misaligning fitted values with unsorted inputs; results now preserve input and requested sorted order.
  • Fixed GPU Batch silently ignoring custom_weights; GPU fit and CV candidate kernels now apply them directly.
  • Reduced GPU adapter buffer requirements from 30 storage/32 total buffer bindings to 7 storage/8 total per shader stage by using per-compute-pipeline resource layouts.
  • Fixed Online incremental mode accepting positive delta and auto_converge settings it cannot use; unsupported combinations now error.
  • Fixed grouped intervals discarding distinct confidence and prediction levels; each requested coverage is now applied independently.
  • Fixed the default boundary_policy ("extend") letting synthetic boundary points bias the shared robustness scale estimate used to reweight every point, compounding across robustness iterations.
  • Improved agreement with R/Cleveland on sparse, asymmetric, and high-iteration fits by aligning robustness stopping, local-linear degeneracy handling, neighborhood traversal, delta interpolation, and weighted accumulation.
  • Fixed zero-radius neighborhoods dropping tied observations or accumulating normalized weights in a different order under robust fits.
  • Fixed Gaussian fits clipping the unbounded kernel to the neighbor window and flooring far-tail weights; all observations now contribute under the standard Gaussian formula.
  • Fixed Julia model handles being finalized during native calls, concurrent Streaming/Online access, result appends with mismatched optional fields, unsafe GPU installer path handling and unsupported targets, and constructor errors hiding native validation details.

4.1.0

Added

  • Added musl release binaries for Julia.
  • Added retain_model and predict(model, new_x; kwargs...) for out-of-sample prediction.
  • Added return_derivative to Lowess, StreamingLowess, and OnlineLowess.
  • Added standard errors and confidence/prediction intervals to StreamingLowess and OnlineLowess; online intervals require update_mode = "full".

Fixed

  • Changed the OnlineLowess default iterations from 3 to 0, matching non-robust incremental updates; robustness iterations require update_mode = "full".

4.0.0

Added

  • Added a return_sorted option to Lowess.
  • Added a missing option to Lowess, StreamingLowess, and OnlineLowess to control non-finite input handling.
  • Added GPU installer support for a local prebuilt artifact and a prebuilt Linux AArch64 GPU library.

Changed

  • Breaking change: removed return_diagnostics, return_residuals, and parallel from OnlineLowess; these options were accepted but had no effect in online mode.
  • Breaking change: StreamingLowess's overlap default now resolves dynamically to chunk_size / 10 instead of a fixed 500.
  • Improved the Julia API documentation, including the dynamic overlap default.

Fixed

  • Clarified that installing the GPU library requires restarting Julia before the GPU backend can be used.

3.2.1

Changed

  • Clarified that results are returned in input order, even though the algorithm sorts internally.

Fixed

  • Corrected OnlineLowess defaults to min_points = 2 and update_mode = "incremental".
  • Fixed Windows local library discovery to resolve the native library at runtime instead of reusing a stale precompiled path.

3.2.0

Changed

  • Expanded and reorganized the Julia documentation with setup guidance, parameter explanations, and standardized examples.

Fixed

  • Corrected the Handling Outliers example to use a fraction that downweights the injected outlier.
  • Shortened confidence-interval and standard-error examples to display the first five values instead of all 100.

3.1.0

Changed

  • Moved Julia documentation to GitHub Pages with executable Documenter.jl examples and updated the README with package-specific guidance.

3.0.0

Added

  • Added an opt-in GPU backend through the gpu Cargo feature and a backend option on Lowess; install_gpu() downloads a prebuilt library.
  • Added cross-reference links in the Julia API documentation to the corresponding user guides.

Changed

  • Breaking change: renamed OnlineOutput fields smoothed and std_error to y and standard_error.
  • Organized Streaming and Online API documentation and tutorials into dedicated user-guide pages.

Fixed

  • Fixed LowessResult.iterations_used returning the raw FFI sentinel -1 instead of nothing when robustness iterations are not applicable.

2.0.0

Added

  • Added the custom_weights keyword to batch fit for non-negative per-observation weights.

Changed

  • Breaking change: replaced add_points(online, x, y) with add_point(online, x, y), which processes one point and returns its smoothed value or nothing until enough points are available.

1.3.0

Fixed

  • Fixed Windows local Julia runs to load the native library at runtime, avoiding stale library paths from precompiled packages.

1.2.0

Changed

  • Updated the GPU backend's wgpu dependency from v26.0 to v27.0.

1.1.2

  • Under-the-hood maintenance; no changes to the public API or runtime behavior.

1.1.1

Fixed

  • Fixed GPU configuration and initialization problems, and improved recovery after missing hardware/drivers or earlier GPU execution errors.

1.1.0

Added

  • Expanded GPU fitting support to selectable kernels, robustness and scaling methods, boundary policies, automatic convergence, prediction, and cross-validation.

Fixed

  • Fixed the Extend boundary policy not being applied and improved numerical precision through coordinate centering.
  • Fixed GPU integer overflow, initialization failures, and resource exhaustion.

1.0.0

Added

  • Added the mean scaling method (Mean Absolute Deviation).

Fixed

  • Caught Rust panics at the FFI boundary and reported them as Julia errors rather than unwinding across the language boundary.

0.99.9

Added

  • Registered the package in JuliaRegistries.

Changed

  • Introduced class-based builders for streaming and online processing.

0.99.8

  • Initial implementation of FastLOWESS.jl.