Roadmap#

sktime-cli is an early alpha. This page records what the current version covers, where it stops, and what comes next.

Coverage#

sktime’s registry holds 697 object registrations across 25 scitypes. Version 0.0.1 dispatched run on four of them, about 40% of the registry. Version 0.0.2 dispatches on eight, about 66%.

scitype

objects

sktime-cli surface

forecaster

156

run (forecasting)

transformer

153

run transform

classifier

78

run (panel)

metric

46

--metric, metrics score

metric_forecasting

32

--metric, metrics score

regressor

31

run (panel)

detector

28

run detect

dataset

21

datasets load

transformer-pairwise-panel

21

used inside other estimators

param_est

20

not yet exposed

splitter

15

--cv, data split --cv

interval_scorer

14

used inside detectors

dataset_forecasting

11

datasets load

network

11

used inside classifiers and regressors

clusterer

11

run (panel)

dataset_classification

9

datasets load

catalogue

9

catalogues list

metric_forecasting_proba

8

--metric, metrics score

aligner

7

not yet exposed

metric_detection

6

--metric, metrics score

reconciler

5

run transform

object

2

abstract category

transformer-pairwise

1

used inside other estimators

early_classifier

1

run (panel)

dataset_regression

1

datasets load

Regenerate this table with python scripts/scitype_coverage.py. Every scitype is classified in sktime_cli/_scitypes.py, and tests/test_scitypes.py fails if sktime adds one that is not, so the gap cannot widen unnoticed.

What version 0.0.2 covers#

  • Registry search, describe, tags, and types, backed by a disk cache.

  • Dataset listing, description, and loading from sktime’s dataset objects plus the UCR/UEA, Monash, and fpp3 sources.

  • Benchmark catalogues.

  • Data inspection, format conversion, temporal splitting, and cross-validation fold generation.

  • Fit, predict, transform, detect, and evaluate across forecasters, classifiers, regressors, clusterers, transformers, reconcilers, and detectors.

  • Probabilistic forecasting: intervals, quantiles, variance, and residuals.

  • Long-format and hierarchical file input, including global forecasting.

  • Metric listing and scoring.

  • API-contract checking with sktime-cli check.

  • Saved model artifacts with a spec round trip through model inspect.

  • Environment reporting through version, env, doctor, and cache.

There are no sessions, handles, async jobs, or daemons, and adding them isn’t planned. For the reasoning, see Design.

Known limitations#

  • run evaluate covers forecasters, classifiers, and regressors. sktime has no cross-validation utility for clusterers or detectors, so neither does this.

  • Aligners, parameter estimators, and pairwise transformers are not exposed by run. A generic call command is the likely route rather than one command each.

  • --fh accepts relative horizons only, so absolute and dated horizons are not supported.

  • Spec strings naming non-sktime objects fall back to sktime.registry.craft, which needs pytest installed in lean environments. This is an upstream issue; see the note in sktime_cli/_specs.py.

Planned for later versions#

Model workflows

run update for streaming updates, run tune over ForecastingGridSearchCV, and a generic call_method for any estimator method.

Data

Absolute or dated --fh values.

Reproducibility

A full export_code command that emits a runnable Python script for a run.

Ecosystem integration

The benchmarking module, plotting, an mlflow flavor, and pushing models to the Hugging Face Hub.