feat(automl): time-series backtesting curve support

.github/scripts/check_imports/import_exceptions.yaml

@@ -13,6 +13,7 @@ files:
     - pandas
   "components/**/shared/*":
     - pandas
+    - numpy
     - component
   "pipelines/**/tests/*":
     - tests

components/training/automl/autogluon_timeseries_models_training/README.md

@@ -125,3 +125,12 @@ def example_pipeline(
 Writes ``component_status.json`` under ``component_status`` with ``component_id`` ``autogluon_timeseries_models_training`` and training stages (``load_data``, ``model_selection``, ``refit_full``, ``evaluate_models``). Artifact metadata display name: **Timeseries Models Training Status**.
 
 Inference notebooks are loaded from ``shared/notebook_templates/timeseries_notebook.ipynb`` at runtime (same shared package data as tabular training).
+
+### Model insight artifacts (per refitted model)
+
+Under each ``{model_name}_FULL/metrics/`` directory:
+
+- **`metrics.json`**: Holdout test metrics from ``TimeSeriesPredictor.evaluate`` (finite values only).
+- **`back_testing.json`**: Multi-window backtest summary with ``per_window_metrics`` and ``series_analysis`` (best/worst series forecast timelines). Generated best-effort after refit; omitted when backtest APIs or history length are insufficient.
+
+The timeseries notebook template loads ``back_testing.json`` when present for model insights.

components/training/automl/autogluon_timeseries_models_training/component.py

@@ -346,6 +346,30 @@ def replace_placeholder_in_notebook(notebook, replacements):
                 with (metrics_path / "metrics.json").open("w", encoding="utf-8") as f:
                     json.dump(metrics_dict, f, indent=2)
 
+                from kfp_components.components.training.automl.shared.back_testing import build_back_testing_json
+
+                try:
+                    back_testing_payload = build_back_testing_json(
+                        predictor_refit,
+                        model_name=model_name,
+                        model_name_full=model_name_full,
+                        train_data=full_train_ts_df,
+                        eval_metric=eval_metric,
+                        target=target,
+                        id_column=id_column,
+                        timestamp_column=timestamp_column,
+                        prediction_length=prediction_length,
+                        metrics=list(AVAILABLE_METRICS.keys()),
+                    )
+                    with (metrics_path / "back_testing.json").open("w", encoding="utf-8") as f:
+                        json.dump(back_testing_payload, f, indent=2)
+                except Exception as backtest_exc:
+                    logger.warning(
+                        "Could not generate back_testing.json for model %r: %s. Skipping backtest artifact.",
+                        model_name_full,
+                        backtest_exc,
+                    )
+
                 notebook_file = "timeseries_notebook.ipynb"
                 with (shared_automl_dir() / "notebook_templates" / notebook_file).open("r", encoding="utf-8") as f:
                     notebook = json.load(f)
@@ -372,6 +396,7 @@ def replace_placeholder_in_notebook(notebook, replacements):
                         "predictor": str(Path(model_name_full) / "predictor"),
                         "notebook": str(Path(model_name_full) / "notebooks" / "automl_predictor_notebook.ipynb"),
                         "metrics": str(Path(model_name_full) / "metrics"),
+                        "back_testing": str(Path(model_name_full) / "metrics" / "back_testing.json"),
                     },
                     "metrics": {
                         "test_data": metrics_dict,