Roadmap¶
This file should describe future work, not already-shipped functionality.
For the full prioritized technical roadmap with implementation details, see ./technical_roadmap.md.
Current Focus¶
- tighten and stabilize the public API surface ahead of wider release
- improve docs so the maintained guides, architecture notes, and API spec stay aligned
- expand verification around newer optional components such as benchmarking, transformers, and torch-backed models
Near-Term Priorities (Phase 1)¶
- Quantile-based CRPS — replace Gaussian approximation with exact quantile CRPS
- Remove dead scipy import in
distribution.py:sample() - Unified metric API — single
score()entry point for all metrics .summary()onDistributionPrediction— one-call overview- Multivariate plot facets — fix silent first-target-only plotting
See ./technical_roadmap.md#phase-1--fix-whats-broken-v020 for details.
Mid-Term Priorities (Phase 2)¶
- PIT histograms and continuous calibration curves
- Isotonic recalibration wrapper (close the calibration loop)
- Sequential conformal updates (ACI) for non-stationary data
- Rolling-origin and sliding-window time series evaluation
- Non-crossing quantile training
See ./technical_roadmap.md#phase-2--core-statistical-rigor-v030 for details.
Longer-Term Ideas (Phases 3–4)¶
| Idea | Phase | Notes |
|---|---|---|
| Parametric distribution fitting | 3 | Normal, Student-t, LogNormal from quantiles |
| Log-score metric | 3 | Strictly proper scoring rule alongside CRPS |
| Energy score for multivariate | 3 | Proper multivariate scoring rule |
| Extended copula support (>2 targets) | 3 | Vine copulas or pairwise chaining |
| SHAP and leverage analysis in model API | 3 | Wire standalone modules into BaseUncertaintyModel |
| EnbPI for time series conformal | 4 | Bootstrap + sequential conformal |
| Conformal classification / prediction sets | 4 | APS prediction sets |
| Batch and streaming prediction API | 4 | Production serving support |
| Model comparison suite | 4 | Skill scores, Diebold-Mariano tests |
| Risk function integration | 4 | Wire risk_functions.py into ConformalRiskControl |
| Test coverage to 80%+ | 4 | From current 65% floor |
Completed Work¶
Features that have already landed should stay in the changelog or archive, not in this roadmap. See ./changelog.md and ../archive/README.md for historical context.