pandas skew/kurt vs scipy.stats: different estimators, kurtosis can flip sign

pandas 3.1.0 / SciPy 1.18.1. Series.skew() and Series.kurt() are the bias-adjusted estimators (G1 and adjusted excess kurtosis). scipy.stats.skew and kurtosis default to bias=True, the plain moment ratios g1 and g2. For [1, 2, 3, 4, 10]: skew 1.697 (pandas) vs 1.138 (scipy), and kurtosis 3.152 vs -0.212, so the sign of excess kurtosis flips. scipy with bias=False reproduces pandas exactly. The skew ratio is exactly sqrt(n(n-1))/(n-2): 1.49 at n=5, 1.054 at n=30, 1.015 at n=100 (checked on lognormal samples). Kurtosis gaps stay visible further out (lognormal, n=30: 8.92 vs 7.30). scipy.stats.describe mixes conventions: variance uses ddof=1 (12.5) while skewness and kurtosis are biased. With NaN, pandas skips it (skipna), scipy returns nan unless nan_policy='omit', and even then it is still the biased estimator (0.848 vs pandas 1.469 on [1, 2, nan, 4, 10]); and Series.corr / DataFrame.corr use pairwise-complete rows, while np.corrcoef returns nan, so df.corr() and df.dropna().corr() differ when NaNs are scattered (-0.8857 vs -0.8846 in a 5-row example). State the estimator (bias, fisher, ddof, NaN policy) whenever you compare these numbers across tools.