Can Pearson r be larger than one?

A valid exact calculation cannot; values outside the range indicate a numerical or implementation issue.

Does r = -1 mean a weak relationship?

No; it means a perfect negative linear relationship for nonconstant paired data.

Can a coefficient be calculated without normality?

Yes. Normality concerns certain inferential procedures, not the descriptive arithmetic itself.

AI Glossary / Statistics and probability / Intermediate

Correlation coefficient

Definition

A correlation coefficient is a numerical summary of the direction and strength of a specified association between variables. In its usual Pearson form, it is the covariance divided by the product of the variables' standard deviations, measuring linear association on a unit-free scale from minus one to plus one. The coefficient's interpretation depends on which correlation measure is used and how the data were collected.

What it means

The sample Pearson coefficient is commonly written r, while the corresponding population correlation is often written rho. A value near plus one describes a strong positive linear pattern, and a value near minus one describes a strong negative linear pattern. Values near zero indicate little linear association, not necessarily little association of every kind.
The coefficient is symmetric: correlating X with Y gives the same value as correlating Y with X. That symmetry is one reason it cannot by itself specify a causal direction or a regression slope. Naming the coefficient is important because rank-based and other association coefficients answer different questions.

How it works

Start with paired numerical observations and inspect a scatter plot. For Pearson r, calculate each variable's deviations from its sample mean. Add the products of paired deviations to obtain S_xy, and add the squared deviations separately to obtain S_xx and S_yy. Then r = S_xy/sqrt(S_xx S_yy).
This form is equivalent to dividing sample covariance by sample standard deviations when consistent denominators are used. Both variables must vary. If either has zero variance, the denominator is zero and the coefficient is undefined; reporting zero would confuse lack of variation with lack of association.
The coefficient is unaffected by adding constants or multiplying a variable by a positive constant. A negative multiplier reverses the sign. Nonlinear transformations generally change Pearson r. Different units therefore do not prevent comparison of linear association, but different populations, ranges, or measurement quality still affect interpretation.
Separate estimation from inference. The usual exact null test for Pearson correlation relies on particular assumptions, commonly independent observations from a bivariate normal model. Other valid procedures may use appropriate resampling or modeling assumptions. Time-dependent or clustered pairs need special treatment.
Report the number of pairs, missing-data handling, uncertainty method, and relevant diagnostic findings. Avoid universal labels such as strong without context. The practical meaning of a given coefficient depends on the process, noise, measurement scale, and decision.

Three practical examples

A forecasting team summarizes the linear relationship between two measured business indicators before investigating trends and lags.
A researcher reports a Pearson coefficient and interval for paired continuous measurements while inspecting the scatter plot.
An AI evaluation compares predicted and observed values but supplements their correlation with error and calibration measures.

Think of it this way

Think of the coefficient as a standardized summary of alignment between two centered lists of values. It tells how consistently the lists point together or in opposite directions. Standardizing removes units, but it does not remove the need to understand what the lists represent.

Worked example

Consider X values 1, 2, 3, 4 and paired Y values 8, 6, 7, 3. Their means are 2.5 and six. The centered cross-products sum to -7, while the sums of squared deviations are five for X and fourteen for Y.
Pearson r = -7/sqrt(5 x 14), approximately -0.83666. The negative sign describes the downward linear tendency. Squaring gives r squared = 0.70. In a simple ordinary least-squares regression of Y on X with an intercept, fitted on these same observations, this equals the model's R-squared.
The regression slope is S_xy/S_xx = -7/5 = -1.4 Y-units per X-unit, which is different from the unit-free correlation. Multiplying every X value by ten leaves r unchanged but changes that slope to -0.14 per new X-unit.
The R-squared relationship does not turn 0.70 into a general forecast-accuracy rate, a causal percentage, or the R-squared of an arbitrary multivariable model. These four observations are a calculation example, not strong evidence about an unknown population correlation. A precision assessment would require an appropriate inferential model and considerably more context.

Illustrative case study

Fictional case: Elm Forecasting evaluates a model whose predictions correlate highly with observed demand. A reviewer notices that the model consistently predicts twice the actual values. The correlation remains high because the two series move together, but the forecasts are poorly calibrated and have substantial errors.
The team adds direct error measures and checks performance on later periods. It retains correlation as one descriptive statistic rather than using it as the sole acceptance criterion. The revised report explains both the model's ability to track variation and its shortcomings in predicting the required numerical level.

Common misconceptions and mistakes

Do not read r as the percentage of variance explained; even the interpretation of r squared requires a suitable model context. Correlation is not agreement, prediction error, or causation. A coefficient can be distorted by one influential point, restricted range, aggregation, or a shared time trend. A tiny p-value is not a large effect, and a nonsignificant test is not proof that the population coefficient equals zero.

Business use

A correlation coefficient provides a compact comparison useful in exploratory reporting, feature review, and monitoring. Keep its limitations visible: report the measure used, the population represented, and a plot or diagnostic summary. When evaluating an AI prediction system, pair correlation with measures aligned to the operational loss, such as absolute error or calibrated probabilities, and assess performance under realistic deployment conditions.

Check your understanding

What is r in the worked example?

Answer: Approximately -0.83666.

Why is the slope different?

Answer: It retains the variables' measurement units and scales.

If a predictor is constant, should its Pearson coefficient be reported as zero?

Answer: No, it is undefined.

Related terms

Covariance
Standard deviation
Coefficient of determination
Linear regression
Calibration
Pearson correlation
Spearman coefficient
Fisher transformation

Learn next

Calculate Pearson r from centered sums and verify its invariance to positive changes of units. Then study uncertainty intervals and compare correlation with agreement and prediction-error measures on the same dataset.

Sources

Makridakis, Wheelwright and Hyndman, Glossary of Forecasting Terms, correlation coefficient and covariance entries: https://robjhyndman.com/mwh3/FG4.pdf SciPy documentation, pearsonr, definition and inference assumptions: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.pearsonr.html

Previous: Correlation clustering Next: Correlation does not imply causation