fix(automl): treat a DatetimeIndex as the ts_forecast time column

flaml/automl/task/time_series_task.py

@@ -126,7 +126,9 @@ def validate_data(
             else:
                 target_names = label
 
+            time_col_inferred = False
             if self.time_col is None:
+                time_col_inferred = True
                 if isinstance(X_train_all, pd.DataFrame):
                     assert dataframe is None, "One of dataframe and X arguments must be None"
                     self.time_col = X_train_all.columns[0]
@@ -141,6 +143,18 @@ def validate_data(
             if X_train_all is not None:
                 assert y_train_all is not None, "If X_train_all is not None, y_train_all must also be"
                 assert dataframe is None, "If X_train_all is provided, dataframe must be None"
+                if isinstance(X_train_all, pd.DataFrame):
+                    X_train_all, promoted_time_col = self._promote_datetime_index_if_needed(
+                        X_train_all,
+                        time_col=self.time_col,
+                        is_inferred=time_col_inferred,
+                    )
+                    if promoted_time_col is not None:
+                        self.time_col = promoted_time_col
+                if isinstance(y_train_all, (pd.DataFrame, pd.Series)) and isinstance(X_train_all, pd.DataFrame):
+                    if not y_train_all.index.equals(X_train_all.index):
+                        y_train_all = y_train_all.copy()
+                        y_train_all.index = X_train_all.index
                 dataframe = TimeSeriesDataset.to_dataframe(X_train_all, y_train_all, target_names, self.time_col)
 
             elif dataframe is not None:
@@ -150,6 +164,14 @@ def validate_data(
             else:
                 raise ValueError("Must supply either X_train_all and y_train_all, or dataframe and label")
 
+            dataframe, promoted_time_col = self._promote_datetime_index_if_needed(
+                dataframe,
+                time_col=self.time_col,
+                is_inferred=time_col_inferred,
+            )
+            if promoted_time_col is not None:
+                self.time_col = promoted_time_col
+
             try:
                 dataframe.loc[:, self.time_col] = pd.to_datetime(dataframe[self.time_col])
             except Exception:
@@ -161,6 +183,16 @@ def validate_data(
 
             if X_val is not None:
                 assert y_val is not None, "If X_val is not None, y_val must also be"
+                if isinstance(X_val, pd.DataFrame):
+                    X_val, _ = self._promote_datetime_index_if_needed(
+                        X_val,
+                        time_col=self.time_col,
+                        is_inferred=False,
+                    )
+                if isinstance(y_val, (pd.DataFrame, pd.Series)) and isinstance(X_val, pd.DataFrame):
+                    if not y_val.index.equals(X_val.index):
+                        y_val = y_val.copy()
+                        y_val.index = X_val.index
                 val_df = TimeSeriesDataset.to_dataframe(X_val, y_val, target_names, self.time_col)
                 val_len = len(val_df)
             else:
@@ -385,7 +417,60 @@ def _preprocess(self, X, transformer=None):
             X = transformer.transform(X)
         return X
 
+    @staticmethod
+    def _promote_datetime_index_if_needed(df, time_col=None, is_inferred=False):
+        """Promote a DatetimeIndex to a guaranteed-unique column if needed.
+
+        Args:
+            df: A pandas DataFrame.
+            time_col: The current timestamp column name (if known).
+            is_inferred: Whether time_col was inferred rather than explicitly specified.
+
+        Returns:
+            Tuple of (df_processed, promoted_col_name or None).
+        """
+        if not isinstance(df, pd.DataFrame) or not isinstance(df.index, pd.DatetimeIndex):
+            return df, None
+
+        # Precondition check:
+        # 1. If time_col was inferred: promote if df lacks a datetime-typed time_col.
+        # 2. If time_col was explicit: promote if time_col is missing from df.columns.
+        needs_promotion = False
+        if is_inferred:
+            if time_col is None or time_col not in df.columns or not pd.api.types.is_datetime64_any_dtype(df[time_col]):
+                needs_promotion = True
+        else:
+            if time_col is not None and time_col not in df.columns:
+                needs_promotion = True
+
+        if not needs_promotion:
+            return df, None
+
+        base_name = df.index.name or "ds"
+        promoted_col = base_name if not is_inferred and time_col else base_name
+        if not is_inferred and time_col:
+            promoted_col = time_col
+
+        # Guarantee unique column name avoiding collisions with existing columns
+        target_col = promoted_col
+        i = 1
+        while target_col in df.columns:
+            target_col = f"{promoted_col}_{i}"
+            i += 1
+
+        df_out = df.copy()
+        df_out.insert(0, target_col, df.index)
+        df_out = df_out.reset_index(drop=True)
+        logger.info(f"Timestamp column not found, promoted DataFrame index as '{target_col}'.")
+        return df_out, target_col
+
     def preprocess(self, X, transformer=None):
+        if isinstance(X, pd.DataFrame):
+            X, _ = self._promote_datetime_index_if_needed(
+                X,
+                time_col=self.time_col,
+                is_inferred=False,
+            )
         if isinstance(X, (pd.DataFrame, np.ndarray, pd.Series)):
             X = normalize_ts_data(X.copy(), self.target_names, self.time_col)
             return self._preprocess(X, transformer)

flaml/automl/time_series/ts_data.py

@@ -549,6 +549,9 @@ def normalize_ts_data(X_train_all, target_names, time_col, y_train_all=None):
         elif isinstance(y_train_all, pd.Series):
             y_train_all = pd.DataFrame(y_train_all)
             y_train_all.index = X_train_all.index
+        elif isinstance(y_train_all, pd.DataFrame):
+            y_train_all = y_train_all.copy()
+            y_train_all.index = X_train_all.index
 
         dataframe = pd.concat([X_train_all, y_train_all], axis=1)
 

test/automl/test_ts_forecast_datetime_index.py

@@ -0,0 +1,163 @@
+"""Regression tests for ts_forecast input where the timestamp is the DataFrame index.
+
+Flaml expects the timestamp to live in a *column* (`time_col`). When a user passes the
+idiomatic pandas shape instead -- timestamps as a `DatetimeIndex` -- `validate_data`
+defaults `time_col` to `dataframe.columns[0]`, which is a value column (often the label
+itself), so the real timestamps are never looked at.
+"""
+
+import numpy as np
+import pandas as pd
+import pytest
+
+from flaml import AutoML
+
+
+def _make_train_df(periods=120):
+    return pd.DataFrame(
+        {"y": np.sin(np.arange(periods) / 6) * 10 + 50},
+        index=pd.date_range("2018-01-01", periods=periods, freq="MS"),
+    )
+
+
+def _fit(automl, **kwargs):
+    settings = {
+        "task": "ts_forecast",
+        "label": "y",
+        "period": 12,
+        "time_budget": 5,
+        "estimator_list": ["lgbm"],
+        "metric": "mape",
+        "log_file_name": False,
+        "verbose": 0,
+    }
+    settings.update(kwargs)
+    automl.fit(**settings)
+
+
+def test_unnamed_datetime_index_becomes_the_time_column():
+    automl = AutoML()
+    _fit(automl, dataframe=_make_train_df())
+    assert automl._state.task.time_col == "ds"
+
+
+def test_named_datetime_index_keeps_its_name():
+    df = _make_train_df()
+    df.index.name = "month"
+    automl = AutoML()
+    _fit(automl, dataframe=df)
+    assert automl._state.task.time_col == "month"
+
+
+def test_promoted_index_keeps_every_training_row():
+    """Pre-promotion the label column was read as epoch nanoseconds, collapsing the 120 real
+    timestamps onto a handful of 1970-01-01 values and dropping the duplicates."""
+    df = _make_train_df()
+    automl = AutoML()
+    _fit(automl, dataframe=df)
+    assert automl._state.data_size[0] == len(df)
+
+
+def test_explicit_time_col_still_wins_over_the_index():
+    df = _make_train_df()
+    df["ts"] = df.index
+    automl = AutoML()
+    _fit(automl, dataframe=df, time_col="ts")
+    assert automl._state.task.time_col == "ts"
+
+
+def test_genuine_missing_timestamp_still_raises():
+    df = pd.DataFrame({"note": ["not-a-date"] * 120, "y": np.arange(120, dtype=float)})
+    automl = AutoML()
+    with pytest.raises(ValueError) as exc:
+        _fit(automl, dataframe=df)
+    assert "must contain timestamp values" in str(exc.value)
+
+
+def test_timestamp_column_input_is_unchanged():
+    df = _make_train_df().reset_index().rename(columns={"index": "ds"})
+    automl = AutoML()
+    _fit(automl, dataframe=df, time_col="ds")
+    assert automl._state.task.time_col == "ds"
+    assert len(automl.predict(df[["ds"]].tail(12))) == 12
+
+
+def test_index_name_collision_resolves_to_unique_column():
+    """When dataframe already has a column named 'ds' (or 'index') that is not datetime,
+    promoting an unnamed DatetimeIndex must not overwrite or conflict with the existing column."""
+    df = _make_train_df()
+    df["ds"] = np.arange(len(df), dtype=float)  # Collision candidate
+    df["ds_1"] = np.arange(len(df), dtype=float)  # Double collision candidate
+    automl = AutoML()
+    _fit(automl, dataframe=df)
+    # Target column should be uniquely chosen as ds_2 avoiding collision
+    assert automl._state.task.time_col == "ds_2"
+    assert "ds" in automl._feature_names_in_
+    assert "ds_1" in automl._feature_names_in_
+    assert "ds_2" in automl._feature_names_in_
+    assert automl.data_size_full == len(df)
+
+
+def test_validation_data_with_datetime_index():
+    """Verify that validation data passed as DataFrame with DatetimeIndex is properly promoted."""
+    df_train = _make_train_df(periods=100)
+    val_index = pd.date_range("2026-05-01", periods=20, freq="MS")
+    df_val = pd.DataFrame({"y": np.sin(np.arange(100, 120) / 6) * 10 + 50}, index=val_index)
+    X_val = pd.DataFrame(index=val_index)
+    y_val = df_val["y"]
+    automl = AutoML()
+    _fit(
+        automl,
+        dataframe=df_train,
+        X_val=X_val,
+        y_val=y_val,
+    )
+    assert automl._state.task.time_col == "ds"
+    assert automl._state.eval_method == "holdout"
+    assert len(automl.predict(X_val)) == len(val_index)
+
+
+def test_validation_data_with_dataframe_target_and_datetime_index():
+    """Verify that validation data with DataFrame y_val and DatetimeIndex preserves aligned indexes."""
+    df_train = _make_train_df(periods=100)
+    val_index = pd.date_range("2026-05-01", periods=20, freq="MS")
+    df_val = pd.DataFrame({"y": np.sin(np.arange(100, 120) / 6) * 10 + 50}, index=val_index)
+    X_val = pd.DataFrame(index=val_index)
+    y_val = df_val[["y"]]  # DataFrame target with DatetimeIndex
+    automl = AutoML()
+    _fit(
+        automl,
+        dataframe=df_train,
+        X_val=X_val,
+        y_val=y_val,
+    )
+    assert automl._state.task.time_col == "ds"
+    assert automl._state.eval_method == "holdout"
+    assert len(automl.predict(X_val)) == len(val_index)
+
+
+def test_xtrain_ytrain_with_datetime_index():
+    """Verify that (X_train, y_train) inputs with DatetimeIndex are promoted and work seamlessly."""
+    train_idx = pd.date_range("2018-01-01", periods=100, freq="MS")
+    val_idx = pd.date_range("2026-05-01", periods=20, freq="MS")
+    X_train = pd.DataFrame({"feat": np.arange(100, dtype=float)}, index=train_idx)
+    y_train = pd.Series(np.sin(np.arange(100) / 6) * 10 + 50, index=train_idx, name="y")
+    X_val = pd.DataFrame({"feat": np.arange(100, 120, dtype=float)}, index=val_idx)
+    y_val = pd.DataFrame({"y": np.sin(np.arange(100, 120) / 6) * 10 + 50}, index=val_idx)
+    automl = AutoML()
+    _fit(automl, X_train=X_train, y_train=y_train, X_val=X_val, y_val=y_val)
+    assert automl._state.task.time_col == "ds"
+    assert len(automl.predict(X_val)) == 20
+
+
+def test_predict_with_datetime_index():
+    """Verify that future prediction data passed with DatetimeIndex works without explicit time_col."""
+    df = _make_train_df(periods=120)
+    automl = AutoML()
+    _fit(automl, dataframe=df)
+
+    # Future prediction input using DatetimeIndex
+    future_index = pd.date_range("2028-01-01", periods=12, freq="MS")
+    future_df = pd.DataFrame(index=future_index)
+    preds = automl.predict(future_df)
+    assert len(preds) == 12