np.round and Python round disagree on 43% of x.xx5 values
NumPy 2.5.4 / pandas 3.1.0 / CPython 3.12. I took the 100,000 decimal literals i.jk5 (0.005 to 999.995) and rounded each to 2 places. Python round(x, 2) and np.round(x, 2) differed on 43,414 of them. Python's round is correctly rounded from the exact binary value (100,000 of 100,000 match Decimal(x).quantize, half-even). np.round computes rint(x*100)/100, and the multiply rounds first, so only 56,586 match. Examples: round(2.675, 2) = 2.67 (the stored value is 2.67499999...) but np.round gives 2.68; round(1234.565, 2) = 1234.57 but np.round gives 1234.56. pandas Series/DataFrame.round follow NumPy (2.68), and f-string '.2f' and np.format_float_positional follow Python (2.67). Neither does commercial half-up: round(1.005, 2) is 1.0 in both. For money or reproducible reports, use Decimal(str(x)).quantize(Decimal('0.01'), ROUND_HALF_UP) or integer cents. Never assert that a NumPy-rounded column equals a Python-rounded one.
Reproduced exactly on CPython 3.14.10 / NumPy 2.5.4 / pandas 3.1.0. For the 100,000 literals i.jk5, Python round and np.round differ on 43414. Python round matches Decimal(x).quantize(0.01, HALF_EVEN) in 100000 of 100000 cases, np.round in 56586. Spot checks also match: round(2.675,2)=2.67 vs np 2.68, 1234.57 vs np 1234.56, pd.Series.round 2.68, f'{2.675:.2f}' and format_float_positional 2.67, round(1.005,2)=1.0 in both. Your Decimal(str(x)).quantize(Decimal('0.01'), ROUND_HALF_UP) recipe gives 2.68, 1.01, 1234.57 and 0.13 for 2.675, 1.005, 1234.565 and 0.125, i.e. commercial half-up on the shortest repr.