numpy.corrcoef: The Function Behind “numpy corr”

Short answer: NumPy has no function called numpy.corr. The function people mean is numpy.corrcoef, which returns the matrix of Pearson product-moment correlation coefficients. Calling np.corr(...) raises an AttributeError; call np.corrcoef(...) instead. This guide is for researchers and analysts who have a table of measurements in Python and need a correlation matrix, and it covers the one mistake that most often gives a wrong answer without any error: the rowvar orientation.

What numpy.corrcoef is used for in research

Pearson correlation measures the strength of the linear association between two quantitative variables, from -1 (perfect negative) to +1 (perfect positive). numpy.corrcoef computes it for every pair of variables you give it at once, so one call yields the whole matrix you would put in a descriptive-statistics table. It does not give a p-value or confidence interval; for those see the alternatives table below. For the concepts behind the number, read types of correlation and scatterplot interpretation, and remember that a correlation alone does not make a correlational study causal.

Install or access

NumPy’s install page (numpy.org/install, read 2026-11-08) lists these official routes: pip install numpy, conda install numpy (defaults or conda-forge channel), uv pip install numpy and pixi add numpy. It recommends working inside a virtual environment, for example one created with python -m venv my-env. Confirm the install:

import numpy as np
print(np.__version__)

Step-by-step: a correlation matrix with np.corrcoef

The signature in the current NumPy reference is numpy.corrcoef(x, y=None, rowvar=True, *, dtype=None).

1. Import NumPy. import numpy as np
2. Arrange the data so each variable is one row. By default (rowvar=True) each row of x is a variable and each column is an observation.

   x = np.array([[1, 2, 3, 4, 5],
                 [2, 4, 5, 4, 5],
                 [5, 4, 3, 2, 1]], dtype=float)

3. Call the function. R = np.corrcoef(x)
4. Read the result. This example returns a 3 x 3 matrix (rounded to four decimals when run locally with NumPy 2.5.1):

   [[ 1.      0.7746 -1.    ]
    [ 0.7746  1.     -0.7746]
    [-1.     -0.7746  1.    ]]

5. Optionally pass a second set with y. y has the same shape as x; the documentation states that it is treated as additional variables, and the NumPy example with two 3 x 3 arrays returns a 6 x 6 matrix. Stack your variables into one array instead when you can, so every pair appears in a single matrix.

The rowvar pitfall: samples as rows

Most research datasets are stored the other way round: one row per participant or sample, one column per variable. With the default rowvar=True, NumPy treats every row as a variable, so no error is raised and you receive a participant-by-participant matrix instead of a variable-by-variable one. The NumPy reference states that rowvar=False makes each column a variable and each row an observation.

d = np.array([[1, 5], [2, 4], [3, 3], [4, 2], [5, 1], [6, 0]], dtype=float)  # 6 samples, 2 variables
np.corrcoef(d).shape                 # (6, 6)  -> correlations between samples (wrong question)
np.corrcoef(d, rowvar=False)         # 2 x 2   -> correlations between the two variables

Quick check: the matrix side length must equal the number of variables. If you have 6 samples and 2 variables and get a 6 x 6 matrix, flip rowvar or transpose with d.T.

Output and how to read it

• The result R is a square ndarray. Per the documentation, R[i, j] = C[i, j] / sqrt(C[i, i] * C[j, j]), where C is the covariance matrix (see numpy.cov).
• Values lie between -1 and 1 inclusive. The matrix is symmetric and its diagonal is 1 (a variable correlates perfectly with itself).
• The documentation warns that because of floating-point rounding the result may not be exactly Hermitian and the diagonal may not be exactly 1; real and imaginary parts are clipped to the interval [-1, 1]. Round for reporting rather than testing equality with 1.
• The dtype argument (new in NumPy 1.20) sets the result type, with at least float64 precision.
• Report the coefficient with the sample size. A coefficient is a descriptive statistic; corrcoef supplies no significance test.

Troubleshooting

AttributeError: module ‘numpy’ has no attribute ‘corr’

The function name is corrcoef. The name corr is a method on pandas objects (DataFrame.corr, Series.corr), which is likely where the habit comes from.

A matrix of the wrong size

See the rowvar section above: the size should equal the number of variables.

nan in the output

In a local run with NumPy 2.5.1, a constant variable (for example np.corrcoef([1, 2, 3], [2, 2, 2])) produced nan entries and a RuntimeWarning, because a variable with zero variance has no defined correlation. A missing value (np.nan) in an input likewise produced nan in that run. The NumPy reference page does not describe missing-value handling, so remove or impute missing observations deliberately before calling it, and check your own output.

Need a p-value or confidence interval

corrcoef returns neither. Use scipy.stats.pearsonr (below).

Free alternatives and related functions

Function	Returns	Notes from the official documentation
numpy.corrcoef(x, y=None, rowvar=True, *, dtype=None)	Correlation matrix (ndarray)	Pearson product-moment coefficients; rows are variables by default.
scipy.stats.pearsonr(x, y, *, alternative='two-sided', method=None, axis=0)	Result object with statistic and pvalue; confidence_interval(confidence_level, method) method	Use when you need a p-value or confidence interval for one pair. Constant input gives a ConstantInputWarning and nan. The SciPy page cautions that zero correlation does not imply independence unless the pair is jointly normal.
pandas.DataFrame.corr	Correlation matrix (DataFrame)	A pandas method, not part of NumPy; see the pandas documentation for its arguments.

For a single pair, pearsonr(x[0], x[1]) on the example above returns the same coefficient as the matrix (0.7746), plus a two-sided p-value of about 0.124 with n = 5 in a local run under SciPy 1.18.0. Related NumPy how-tos on this site: numpy polyfit for fitting a line to the same variables and numpy percentile for distribution summaries.

FAQ

Is there a numpy.corr function?

No. Use numpy.corrcoef; np.corr raises an AttributeError.

Does numpy.corrcoef compute Pearson or Spearman correlation?

Pearson. The NumPy reference describes it as returning Pearson product-moment correlation coefficients. For rank correlation use a different function, such as one from scipy.stats.

How do I get the correlation between just two variables?

Call np.corrcoef(a, b) and read the off-diagonal element [0, 1]; the result is a 2 x 2 matrix.

Why do I get a huge matrix?

Your observations are probably in rows while rowvar is True. Set rowvar=False.

Sources

• NumPy documentation, numpy.corrcoef (stable), read 2026-11-08.
• NumPy, Installing NumPy, read 2026-11-08.
• SciPy documentation, scipy.stats.pearsonr, read 2026-11-08.