jenga 2.0.0
- Added approximate nearest-neighbor search via custom projection
candidates.
- Added optional torch/CUDA distance acceleration with automatic CPU
fallback.
- Added probabilistic forecast distributions through configurable
simulation draws.
- Added split-conformal prediction intervals calibrated on validation
windows.
- Added multiscale sequence representations for neighbor
matching.
- Replaced philentropy binary-distance calls with custom vectorized
distance metrics and removed the dependency.
- Replaced utility dependencies for iteration, dummy encoding,
imputation, smoothing, timing, descriptive moments, entropy, numeric
error metrics, and KNN distances with local implementations.
- Expanded distance methods and kernels with additional
dependency-free implementations.
- Added sign-respecting filtering on final reconstructed draws before
forecast quantiles are computed.
- Enforced historical sign bounds after conformal calibration and in
forecast plots.
jenga 1.0.0
- Added a
NEWS.md file to track changes to the
package.
jenga 1.1.0
- common sense filter
- added stats to the prediction table (and removed pred_stats)
- added new criterion for model selection: prediction score
- added new metric measures for test error from greybox
- added automatic differentiation module (and removed deriv from
hyper-parameters)
jenga 1.2.0
- prediction score is now managed as a distinct metric and it is not
used for model selection
- streamlined code and processing
- matrix distance calculation: improved performance shifting from
philentropy to Rfast
- other minutiae
jenga 1.3.0
- Rationalization and streamlining
- Now you can deal also with categorical variables