Changes in version 0.4.0 (2026-04-21)                  

  - Added scaffolded download/read support for IDEAL through
    elcf4r_download_ideal() and elcf4r_read_ideal(), focused on
    extracted hourly aggregate-electricity summaries from
    auxiliarydata.zip.
  - Added scaffolded download/read support for GX through
    elcf4r_download_gx() and elcf4r_read_gx(), with support for either
    the official SQLite database or flat exports normalized into the
    common panel schema.
  - Added offline tests for IDEAL and GX download-resolution helpers and
    normalization readers, including GX SQLite-table detection.
  - Updated package docs and the dataset vignette to describe IDEAL as
    an unshipped household-level scaffold under the current CC BY 4.0
    source record, and GX as an unshipped secondary
    transformer/community-level scaffold with explicit licence
    re-verification guidance before redistribution.
  - Removed implicit RETICULATE_PYTHON mutation from the LSTM backend
    probe and added explicit, user-driven Python selection through
    elcf4r_use_tensorflow_env().
  - Tightened CRAN-facing package metadata and examples, including a
    shorter elcf4r_benchmark() help example that runs on a toy
    single-entity panel.

                        Changes in version 0.3.0                        

  - Replaced the previous KWF baseline with a wavelet-based
    implementation using wavelets, deterministic calendar groups, kernel
    weighting and approximation/detail mean correction.
  - Replaced the unused src/kwf_core.cpp placeholder with compiled KWF
    helper routines for distances, kernel weights, group restriction and
    mean-corrected forecasts, and wired the R KWF path to those
    accelerators.
  - Added a first-class clustered KWF workflow with thermosensitivity
    classification, wavelet-feature clustering helpers, cluster
    assignment, and a dedicated elcf4r_fit_kwf_clustered() model path.
  - Generalized dataset ingestion around a common normalized panel
    schema and added dataset adapters for iFlex, StoreNet, Low Carbon
    London and REFIT.
  - Implemented elcf4r_download_storenet() with figshare API resolution
    for known household article IDs and an archive fallback for broader
    StoreNet retrieval.
  - Added a generic rolling-origin benchmark API through
    elcf4r_build_benchmark_index() and elcf4r_benchmark(), with saved
    predictions, backend metadata and support for gam, mars, kwf,
    kwf_clustered and lstm.
  - Completed benchmark metric coverage so shipped benchmark artifacts
    now carry populated NMAE, NRMSE, sMAPE and MASE values for all
    shipped result rows.
  - Added shipped example panels and saved benchmark-result datasets for
    StoreNet, Low Carbon London and REFIT, complementing the existing
    iFlex example and benchmark artifacts.
  - Expanded the shipped benchmark cohorts to stronger rolling windows:
    iFlex now uses 15 households with 28 train days and 7 test days; the
    shipped LCL and REFIT benchmark cohorts are now filtered to
    thermosensitive seasonal windows so kwf_clustered rows are
    benchmarked rather than skipped.
  - Reworked dataset-facing documentation to describe the supported
    reader surface, shipped artifacts and reproducible data-raw/ rebuild
    scripts.
  - Clarified the dataset roadmap around IDEAL and GX: IDEAL is a future
    candidate dataset with a currently verified CC BY 4.0 source record,
    while GX is treated as a secondary transformer-level benchmark
    candidate that requires explicit licence re-verification before any
    shipped subset is added.

                        Changes in version 0.2.0                        

  - Added an iFlex preprocessing pipeline with normalized panel readers,
    daily-segment builders, compact shipped example data, and saved
    benchmark result datasets.
  - Added package documentation and vignettes for the shipped iFlex
    workflows and benchmark outputs, and documented the bundled
    elcf4r_elmas_toy dataset.
  - Replaced the placeholder KWF/LSTM paths with working model wrappers,
    unified predict.elcf4r_model(), and migrated the LSTM implementation
    to keras3 with automatic detection of the r-tensorflow virtualenv.
  - Cleaned up package metadata, namespace declarations, tests, and
    examples so package checks now pass apart from environment-specific
    CRAN notes.

                        Changes in version 0.1.0                        

  - Package creation and initial release containing estimators, autoplot
    helpers, and reliability utilities.