IdxMax and IdxMin return the position of NaN instead of the extreme value

Bug Description

IdxMax and IdxMin return the position of the extreme value inside the rolling window. Both use ndarray.argmax / ndarray.argmin, which do not skip NaN: as soon as the window holds a NaN, they return its position instead.

Max and Min go through pandas.rolling.max/min and do skip NaN, so the two contradict each other on the same series. Mad and Rank in the same file filter NaN explicitly (x[~np.isnan(x)]); IdxMax and IdxMin do not.

This reaches Alpha158. qlib/contrib/data/loader.py builds three field families on these two operators:

$high and $low are NaN while an instrument is suspended.

To Reproduce

1. Run the script below on a clean checkout. It only replaces the data access of a leaf feature; the operators under test run their real code.

import numpy as np
import pandas as pd
from qlib.data.base import Feature
from qlib.data.ops import IdxMax, Max


class SeriesFeature(Feature):
    def __init__(self, name, values):
        super().__init__(name)
        self._values = values

    def _load_internal(self, instrument, start_index, end_index, freq):
        return pd.Series(self._values, dtype="float64")


values = [1.0, 5.0, np.nan, 2.0, 3.0]
window = 3
idx = IdxMax(SeriesFeature("a", values), window).load("x", 0, len(values) - 1, "day")
mx = Max(SeriesFeature("b", values), window).load("x", 0, len(values) - 1, "day")

for i in range(len(values)):
    w = values[max(0, i - window + 1) : i + 1]
    print(f"{w}  Max={mx.iloc[i]}  IdxMax={idx.iloc[i]}  points at {w[int(idx.iloc[i]) - 1]}")

1. Output on be725493:

[1.0]  Max=1.0  IdxMax=1.0  points at 1.0
[1.0, 5.0]  Max=5.0  IdxMax=2.0  points at 5.0
[1.0, 5.0, nan]  Max=5.0  IdxMax=3.0  points at nan
[5.0, nan, 2.0]  Max=5.0  IdxMax=2.0  points at nan
[nan, 2.0, 3.0]  Max=3.0  IdxMax=1.0  points at nan

Every window holding a NaN is wrong.

Expected Behavior

IdxMax points at the value Max reports for the same window, and IdxMin at the value Min reports.

Environment

• Qlib version: 0.9.8.dev34 (editable install)
• Python version: 3.12.13
• OS: MacOS 26.6.2, arm64
• Commit number: be725493eb1a6bbb42bf11b37aa7669f59610ff1
• numpy 2.5.3, pandas 2.3.3

Additional Notes

Scale on real data. csi300, 30 instruments, 2019, qlib_data_simple:

One NaN spoils the window for N days, so the effect is wider than the share of NaN itself.