Clean imported numeric columns
A column contains numbers stored as strings alongside malformed entries.
Pandas · Function
pd.to_numeric
Converts scalar or sequence-like input to numeric data. Invalid values raise by default or can be converted to missing values with coercion; downcast and dtype_backend provide output-type control.
How it works
Pass a scalar, sequence, or Series to convert its values into numeric data. Parsing errors raise by default; choosing coercion turns invalid entries into missing values, while downcast can request a smaller compatible numeric dtype and dtype_backend can control nullable output representation.
Syntax
pandas.to_numeric(arg, errors='raise', downcast=None, dtype_backend=<no_default>)Parameters
| Name | Type | Description |
|---|---|---|
arg | scalar, list, tuple, 1-d array, or Series | Value or collection to convert to numeric data; supported inputs include scalars, lists, tuples, one-dimensional arrays, and Series. |
errors | {'raise', 'coerce'}, default 'raise' | How invalid parsing is handled: raise an exception or convert invalid values to NaN. |
downcast | str, default None | Requests a smaller compatible numeric dtype; accepted categories are integer, signed, unsigned, and float. |
dtype_backend | {‘numpy_nullable’, ‘pyarrow’}, default ‘numpy_nullable’ | Controls the nullable dtype backend used for the result. |
Returns
Numeric data after parsing. The return type depends on the input: a Series for Series input, otherwise an ndarray; with dtype_backend settings, nullable representations may be used.
Key points
- Use the function to parse input values into numeric data.
- Without downcasting, non-numeric input generally produces float64 or int64 according to the parsed data.
- Set downcast when a smaller integer, unsigned integer, signed integer, or floating-point dtype is desirable.
- The argument may be a scalar, sequence, one-dimensional array, or Series.
- Coercion converts unparseable entries into NaN instead of stopping the operation.
- The output type follows the input category rather than always being a Series.
Common mistakes
- The default error policy raises when any value cannot be parsed; use errors="coerce" when invalid values should become missing.
- Very large numbers may lose precision during conversion.
Example: Convert mixed values safely
import pandas as pd
values = pd.to_numeric(["10", "2.5", "unknown"], errors="coerce")
print(values)
# [10. 2.5 nan]Converts parseable values to numbers and represents the unparseable value as NaN.
When to use it
Reduce numeric storage
A parsed column is larger than necessary for its value range.
Continue after invalid parsing
A pipeline needs missing values instead of conversion failures.
Interview questions
QWhat does pandas.to_numeric do, and what does it return?
The function converts a scalar, list, tuple, one-dimensional array, or Series to numeric data. Its default output is float64 or int64 for non-numeric input, depending on the values.
Follow-up: What happens to a Series? The result remains a Series; otherwise the result is an ndarray.
QHow do you handle values that cannot be parsed?
errors="raise" throws for invalid parsing, whereas errors="coerce" replaces invalid values with NaN. The default is "raise".
Follow-up: How can invalid values be handled without silently ignoring them? Catch the conversion exception explicitly rather than using the deprecated ignore option.
QWhat is the purpose of downcast and dtype_backend?
downcast accepts integer, signed, unsigned, or float and requests a smaller compatible dtype. It does not force an unsafe conversion when the values cannot fit.
Follow-up: What additional output choice is available? dtype_backend can select the nullable data representation, including nullable output for coerced values.
Related
pandas.to_datetimepandas.to_timedeltapandas.DataFrame.astypepandas.api.types.is_numeric_dtype