diff --git a/qlib/contrib/eva/alpha.py b/qlib/contrib/eva/alpha.py
index 2c2f9ea7e18..c7cdaf780b0 100644
--- a/qlib/contrib/eva/alpha.py
+++ b/qlib/contrib/eva/alpha.py
@@ -130,8 +130,8 @@ def pred_autocorr(pred: pd.Series, lag=1, inst_col="instrument", date_col="datet
     :param lag:
     """
     if isinstance(pred, pd.DataFrame):
-        pred = pred.iloc[:, 0]
         get_module_logger("pred_autocorr").warning(f"Only the first column in {pred.columns} of `pred` is kept")
+        pred = pred.iloc[:, 0]
     pred_ustk = pred.sort_index().unstack(inst_col)
     corr_s = {}
     for (idx, cur), (_, prev) in zip(pred_ustk.iterrows(), pred_ustk.shift(lag).iterrows()):
diff --git a/tests/misc/test_pred_autocorr_dataframe.py b/tests/misc/test_pred_autocorr_dataframe.py
new file mode 100644
index 00000000000..67f2c48c4e2
--- /dev/null
+++ b/tests/misc/test_pred_autocorr_dataframe.py
@@ -0,0 +1,24 @@
+import logging
+
+import pandas as pd
+
+from qlib.contrib.eva.alpha import pred_autocorr
+
+
+def test_pred_autocorr_dataframe_uses_first_column(caplog, monkeypatch):
+    index = pd.MultiIndex.from_product(
+        [list("ABC"), pd.date_range("2024-01-02", periods=3, freq="B")],
+        names=["instrument", "datetime"],
+    )
+    first = pd.Series([1, 2, 3, 2, 3, 1, 3, 1, 2], index=index, name="score")
+    # The second column must have different autocorrelations to detect selecting it.
+    second = pd.Series([1, 2, 4, 1, 5, 2, 3, 1, 9], index=index)
+    predictions = pd.DataFrame({"score": first, "unused": second})
+
+    monkeypatch.setattr(logging.getLogger("qlib"), "propagate", True)
+    pd.testing.assert_series_equal(pred_autocorr(predictions), pred_autocorr(first))
+    assert (
+        "qlib.pred_autocorr",
+        logging.WARNING,
+        f"Only the first column in {predictions.columns} of `pred` is kept",
+    ) in caplog.record_tuples