fix: correct WMA calculation to use nansum instead of nanmean

[ayushozha]

Summary

Closes #1993

The weighted_mean function in WMA._load_internal (qlib/data/ops.py) normalizes weights to sum to 1 via w = w / w.sum(), then incorrectly uses np.nanmean(w * x) which divides by the element count again. The correct function is np.nansum(w * x) since the weights already sum to 1.

This is consistent with how EMA.exp_weighted_mean is implemented (which already uses np.nansum).

Before

def weighted_mean(x):
    w = np.arange(len(x)) + 1
    w = w / w.sum()
    return np.nanmean(w * x)  # BUG: divides by count again

After

def weighted_mean(x):
    w = np.arange(len(x)) + 1
    w = w / w.sum()
    return np.nansum(w * x)  # correct: weights already sum to 1

Test Plan

• Change is consistent with EMA implementation pattern
• Single-line fix, minimal risk

[JKDasondee]

I ran into the same WMA bug independently and wrote a regression test for it. Rather than filing a duplicate PR I'd like to offer the test here for you to fold into this branch (or for a maintainer to pick up if this one has stalled).

Verified locally: with this branch's one-line fix applied, the test passes against the mock TW 0050 close series in tests/ops/test_elem_operator.py. Without the fix, WMA($close, 5) returns e.g. 29.57 instead of the correct 147.87 for prices around 146 — far below min(close), which any weighted average with non-negative weights summing to one must respect.

def test_WMA(self):
    # Regression test for issue #1993: WMA.weighted_mean divided by the
    # window length twice (normalized weights, then np.nanmean) producing
    # values ~N times smaller than a correct weighted moving average.
    # The sister operator EMA.exp_weighted_mean correctly uses np.nansum.
    N = 5
    field = f"WMA($close, {N})"
    result = ExpressionD.expression(self.instrument, field, self.start_time, self.end_time, self.freq)
    result = result.to_numpy()

    close = self.mock_df["close"].reset_index(drop=True).to_numpy()

    def reference_weighted_mean(x):
        w = np.arange(len(x)) + 1
        w = w / w.sum()
        return np.nansum(w * x)

    golden = (
        pd.Series(close)
        .rolling(N, min_periods=1)
        .apply(reference_weighted_mean, raw=True)
        .to_numpy()
    )
    np.testing.assert_allclose(result, golden, rtol=1e-5)

    # A weighted moving average with non-negative weights summing to one
    # must stay inside the input series' range. The pre-fix implementation
    # produced values ~close/N, far below close.min().
    close_min = float(np.nanmin(close))
    close_max = float(np.nanmax(close))
    self.assertGreaterEqual(float(np.nanmin(result)), close_min - 1e-9)
    self.assertLessEqual(float(np.nanmax(result)), close_max + 1e-9)

Requires import pandas as pd at the top of the file. Place it inside TestElementOperator in tests/ops/test_elem_operator.py.

Validated with:

python -m pytest tests/ops/test_elem_operator.py::TestElementOperator::test_WMA -v