Session 41: Phase 5c: drain TimeSeriesExperiment.predict_model

The second TS verb is fully native. exp.predict_model(forecaster) no longer touches legacy.predict_model — predictions go through the standalone get_predictions_with_intervals helper and metrics are computed against _fit_state["y_test"] via the standalone calculate_metrics utility.

ADDED — engine#

• ADDED — TimeSeriesExperiment.predict_model in packages/engine/pycaret/tasks/time_series.py. Reconciles pipeline (deepcopy if ForecastingPipeline; wire bare forecaster via _add_model_to_pipeline), forecast horizon (arg → _fit_state["fh"] → estimator.fh), and exogenous X (arg / data shim → _fit_state["X_test"]). Calls get_predictions_with_intervals for predictions; computes test metrics via calculate_metrics + update_additional_scorer_kwargs when y_test is available. Returns a typed PredictResult. Supports return_pred_int, custom fh, custom alpha / coverage for prediction intervals.

ADDED — tests#

• ADDED — packages/engine/tests/test_session41_ts_predict_model_drain.py — 8 new tests:
  • Drain-lock for predict_model (poison legacy.predict_model, verify native).
  • Metrics DataFrame: 9 columns (Model + 8 standard TS metrics).
  • pull() returns the predict_model metrics.
  • return_pred_int=True includes lower / upper columns.
  • Custom fh=[1,2,3] produces 3 predictions.
  • Bare forecaster (no pipeline) gets wired into the experiment's preprocess.
  • Object without .predict → TypeError.
  • End-to-end create + predict chain with both legacy verbs poisoned.

INTERNAL#

• INTERNAL — Why reuse get_predictions_with_intervals + calculate_metrics. Same reasoning as cross_validate reuse in s40 — these are free functions (no legacy state), already handle the alpha/coverage interval plumbing and metric registry iteration with additional_kwargs. Re-implementing them on top of bare sktime would either lose interval support or duplicate care code.
• INTERNAL — Bare forecaster path. When the user passes a bare fitted sktime forecaster (not a ForecastingPipeline), we wire it into the experiment's preprocess pipeline via _add_model_to_pipeline. Lets joblib.dump-ed forecasters from older code work without forcing the user to reconstruct a pipeline.
• INTERNAL — The data parameter shim. Supervised predict_model takes data as the second positional arg. TS legacy used X as the exogenous-data parameter. The native override accepts both — data defers to X if both are passed; data becomes X otherwise. Keeps the verb signature compatible with how callers might invoke from the supervised side.

Session 41 delta summary#