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]
@@ -150,6 +152,20 @@ def validate_data(
             else:
                 raise ValueError("Must supply either X_train_all and y_train_all, or dataframe and label")
 
+            if (
+                time_col_inferred
+                and isinstance(dataframe.index, pd.DatetimeIndex)
+                and not pd.api.types.is_datetime64_any_dtype(dataframe[self.time_col])
+            ):
+                # The timestamps were passed as the DataFrame index instead of as a column, so the
+                # inference above picked a value column. Converting that column with
+                # `pd.to_datetime()` reads its numbers as nanoseconds since the epoch rather than
+                # failing, which silently replaces the real time axis with 1970-01-01.
+                index_name = dataframe.index.name or "ds"
+                dataframe = dataframe.reset_index().rename(columns={"index": index_name})
+                logger.info(f"Timestamp column not found, using the DataFrame index as '{index_name}'.")
+                self.time_col = index_name
+
             try:
                 dataframe.loc[:, self.time_col] = pd.to_datetime(dataframe[self.time_col])
             except Exception:

test/automl/test_ts_forecast_datetime_index.py

@@ -0,0 +1,82 @@
+"""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