Changes in version 1.2.0                        

  - Fixed mixed-type detection bug (class_index now per-column; strict
    all-numeric OR all-categorical check).

  - Corrected best_deriv() p-values length/off-by-one and added
    short-series guard.

  - Patched doxa_filter() binary clamp (now clamps to [0,1] instead of
    < 1 typo).

  - Hardened metrics: added safe_scale() to avoid zero/NA scale; fixed
    benchmark to use mean(actuals) for "average".

  - Validated inputs (ci, windows, samples, dates length) and added
    early error for too-short series.

  - Stabilized weight rescaling/min-max with zero-range handling;
    vectorized ensemble normalization.

  - Improved date handling for plots/outputs (consistent step; correct
    future index).

  - Reduced brittleness/perf costs (preallocation where useful, safer
    splitting, fewer coercions) and kept seed propagation for
    reproducibility.

  - Added compute_distance() with memoized caching and new methods:
    "correlation" (1−corr) and "dtw" (guarded on dtw pkg).

  - Replaced Gaussian weight sampling with kernel weights vs. the most
    recent window + Dirichlet draws for uncertainty (faster, stabler).

  - Implemented conformal prediction: CI bounds now calibrated from
    rolling residuals; plot uses these bounds at the requested ci.

  - Introduced optional parallelism (use_parallel, parallel_workers)
    using furrr/future with seeded execution; falls back to sequential
    if unavailable.

  - Enabled distance-matrix caching (via digest when present) for reuse
    across samples and to cut repeated computation.

  - Kept API backward-compatible; new options default off and legacy
    distance methods unchanged.

                 Changes in version 1.1.0 (2022-08-15)                  

  - Rationalization and streamlining of processes.

                 Changes in version 1.0.0 (2022-06-28)                  

  - Added a NEWS.md file to track changes to the package.