Session 44: Phase 5c (cont.): drain TimeSeriesExperiment.finalize_model

The fifth (and last user-facing) TS verb is fully native. After this session, every callable TS verb runs without touching legacy — create_model / predict_model / compare_models / tune_model / finalize_model all bypass self._legacy.<verb>. (assign_model doesn't exist for TS — it's an UnsupervisedExperiment-only verb.) The only legacy.* callsite remaining for TS is legacy.setup() inside _native_setup_timeseries itself.

ADDED — engine#

• ADDED — TimeSeriesExperiment.finalize_model in packages/engine/pycaret/tasks/time_series.py. Refits on the full y (y_train + y_test from _fit_state["y"]) using the experiment's preprocess pipeline. Optional exogenous X_full from _fit_state["X"]. Defaults fit_kwargs["fh"] to the experiment's fh. Returns a FinalizeResult whose pipeline is a deployment-ready ForecastingPipeline. Doesn't mutate the input — input pipeline keeps its train-only fit. Accepts both bare sktime forecasters and ForecastingPipeline instances; raises TypeError for anything else.

ADDED — tests#

• ADDED — packages/engine/tests/test_session44_ts_finalize_model_drain.py — 7 new tests:
  • Drain-lock for finalize_model.
  • Refits on the full y (forecaster's _y length matches _fit_state["y"]).
  • Bare forecaster path (auto-wires into preprocess).
  • predict_model(final.pipeline, fh=[1,2,3,4,5,6]) after finalize produces 6-step forecasts.
  • Object without .fit → TypeError.
  • End-to-end create + tune + finalize + predict chain with all 5 legacy verbs poisoned.
  • Input pipeline isn't mutated by finalize (the original res.pipeline keeps its train-only fit).

INTERNAL#

• INTERNAL — No assign_model for TS. PyCaret's assign_model is defined on UnsupervisedExperiment (clusters/anomaly labels back onto rows). TimeSeriesExperiment inherits from SupervisedExperiment, not UnsupervisedExperiment, so the verb is genuinely absent — not a missed drain. Confirmed with grep -n "def assign_model" across the TS module tree: zero hits.
• INTERNAL — Why phase 5d isn't in this session. Stripping legacy.setup() from _native_setup_timeseries requires porting two pieces: (1) seasonality auto-detection via Fourier analysis (legacy's setup() derives seasonality_present, primary_sp_to_use, strictly_positive, seasonality_type and TS containers read all of those when constructed), and (2) sktime-native temporal_train_test_split + ExpandingWindowSplitter build. Substantive enough to be its own session.
• INTERNAL — finalize_model doesn't go through create_model. Unlike tune_model (which refits via create_model to get standard CV metrics), finalize_model is fundamentally a "refit on more data" operation — there's no holdout left to CV against. So we just call pipeline.fit(y=y_full, X=x_full) directly. Matches legacy semantics.
• INTERNAL — fh defaulting in fit_kwargs. Some sktime forecasters (e.g. ARIMA with auto order) need to know the forecast horizon at fit-time. Defaulting to the experiment's fh lets predict_model work post-finalize without the user having to re-specify it. The try/except TypeError retry handles forecasters that don't accept fh in fit (e.g. NaiveForecaster).

Session 44 delta summary#