Session 39: Native setup() phase 5a: time-series soft drain

The TS experiment finally adopts the native dispatcher. _can_use_native_setup accepts TaskType.TIME_SERIES; _native_setup_timeseries populates _fit_state with TS-shape slots so user-facing accessors work for time-series exactly like they do for the other tasks. This is a soft drain: the native path still calls legacy.setup() underneath because the TS verbs (create_model, predict_model, ...) read from legacy state and haven't been drained yet.

ADDED — engine#

• ADDED — Experiment._native_setup_timeseries(data, setup_kwargs) in packages/engine/pycaret/core/experiment.py. Calls legacy.setup (transitionally — phase 5b removes this) to build the sktime model registry, fold generator, and y_train / y_test, then snapshots them into _fit_state along with TS-specific slots (fh, seasonal_period).
• ADDED — Experiment._predict_model_legacy(estimator, *, data, verbose). Time-series fallback for predict_model — the drained sklearn path doesn't apply because sktime's ForecastingPipeline has no .transform() and forecasters use .predict(fh=...). Phase 5b will drain this.

CHANGED — engine#

• CHANGED — _can_use_native_setup predicate accepts TaskType.TIME_SERIES. The only "still legacy" branch left is caller-supplied setup_kwargs.
• CHANGED — TimeSeriesExperiment.fit() in packages/engine/pycaret/tasks/time_series.py mirrors the base dispatcher: native path → _native_setup_timeseries; legacy path → legacy.setup() + _snapshot_fit_state(). Both branches end with _fit_state populated; previously the legacy-only path didn't populate _fit_state at all and exp.y_train raised KeyError for TS.
• CHANGED — Experiment.predict_model() detects task == TIME_SERIES early and dispatches to _predict_model_legacy. Without this guard the drained path would call ForecastingPipeline.transform(X) and crash with AttributeError.
• CHANGED — Experiment.models() defers to legacy.models() for TS. The TS legacy registry has filter rules (model_type ∈ TSModelTypes) that exclude ensemble_forecaster (which requires runtime-built forecasters). Re-implementing those rules in the snapshot path would duplicate logic that lives in time_series/forecasting/oop.py; phase 5b moves both at once.

ADDED — tests#

• ADDED — packages/engine/tests/test_session39_native_setup_phase5.py — 9 new tests:
  • Predicate accepts TIME_SERIES; setup_kwargs still forces legacy.
  • Native path populates _fit_state with TS-shape slots (y / y_train / y_test / fh / fold_generator / preprocess_pipeline / model_registry).
  • User-facing accessors (exp.y_train, exp.y_test, exp.preprocess_pipeline) work for TS.
  • fold_generator is sktime's ExpandingWindowSplitter.
  • setup_kwargs path also populates _fit_state via the snapshot helper.
  • create_model('naive') returns a fitted forecaster.
  • predict_model(forecaster) returns a 12-step forecast matching fh.
  • models(internal=True) filters out ensemble_forecaster via legacy delegation.

CHANGED — existing tests#

• CHANGED — test_session35_native_setup::test_can_use_native_setup_predicate — TS assertion flipped from is False to is True. Docstring updated to reflect "Phase 5a — TS adopts the dispatcher (soft drain)".
• CHANGED — test_session38_native_setup_phase4::test_time_series_still_falls_back_to_legacy renamed to test_time_series_now_routes_through_native_dispatcher_phase5a. Inverted assertion. Same rename pattern as test_session35_native_setup::test_unsupervised_uses_legacy_setup → test_unsupervised_now_runs_natively_phase4 from s38.

INTERNAL#

• INTERNAL — Why a soft drain. The supervised + unsupervised drains worked one-shot because their verbs had already been drained (s24 / s28) to read from _fit_state. For TS, none of the verbs have been drained — they all delegate to self._legacy.<verb>. If _native_setup_timeseries skipped legacy.setup(), every verb would break. So phase 5a brings TS into the dispatcher and gives us accessor parity, but the actual skip of legacy.setup() waits for phase 5b/c when verbs migrate.
• INTERNAL — TS-specific slots in _fit_state. fh (sktime ForecastingHorizon) and seasonal_period (int or None) live alongside the standard slots. Other tasks set them to None. This keeps the dict shape stable across all five task types and lets future verb-drain code branch on self._fit_state["fh"] is not None rather than checking self.task.
• INTERNAL — Why models() defers, not _fit_state["model_registry"]. The snapshot already has the full registry (26 entries including ensemble_forecaster). The legacy models() filters by model_type in TSModelTypes and ensemble_forecaster has model_type="ensemble" (not a TSModelTypes member). Reimplementing that filter on the snapshot is straightforward but would couple core/experiment.py to TSModelTypes. Better to keep the deferral until phase 5b where TS verbs and TS-specific filtering both move together.

Session 39 delta summary#