What Does R Squared Mean?
R-squared means the share of the variation in your outcome that a regression model can explain. In plain English, it is the variance explained by the predictors in the model, expressed as a proportion of variance in the dependent variable. If the R-squared value is 0.80, the model accounts for 80% of the total variance, and 20% remains as unexplained variance.
So, what does r squared mean in practice? It shows how much of the outcome your model can predict, and how closely the line of best fit follows your data points on a scatter plot. Its formal name is the coefficient of determination, written with the symbol R². This simple explanation covers the formula, worked numbers, and common mistakes, in simple terms.
Table of Contents
What Does R Squared Mean in Simple Terms?
Take an everyday example. Suppose study hours are the predictor and test scores are the outcome. If students who study more score higher, R-squared is high. If test scores vary for other reasons, such as sleep, study hours explain only a small part, and R-squared is low.
The same idea works for house prices. Price is the dependent variable, and size, location, and age are the independent variables that might explain it. The explained variance is the part the model can explain, also called the explained variation, and the rest is the unexplained variation. Together they make up the total variation in prices, and R-squared is the explained part divided by the total part.
An R-squared of 0 means the model explains none of the variation, so it is no better than guessing the average every time. An R-squared of 1 means every point sits exactly on the regression line, which almost never happens with real data. An R-squared of 0.50 means half the variation is explained, and an R-squared of 0.80 means four-fifths is explained. R-squared is a measure of goodness of fit, describing model fit in one number.
How Is R-Squared Calculated?
The R-squared formula is R² = 1 – SSres / SStot, where SSres is the sum of squared residuals and SStot is the sum of squares total. To calculate R-squared, find how far each prediction misses its real value, square those misses so positive and negative errors do not cancel, and add them up. Then compare that total with the total spread of the data around its mean. To interpret R-squared, compare the error with predicting the mean; interpreting R-squared this way keeps it honest.
Two sums drive the calculation. The residual sum of squares, or SSres, adds up each residual: the difference between an observed value and its fitted value. Observed values are the actual values measured, and fitted values are the predicted values from the regression line. The total sum of squares, or SStot, measures each observed value against the mean of the observed values, the mean value you would predict with no model, as in the mean, median and mode idea. The coefficient of determination reference gives the full definition, and the same formula is used in linear regression, simple linear regression, multiple regression, and multiple linear regression.
The line is usually fitted by ordinary least squares, often shortened to OLS, the least squares method that makes the sum of squared residuals as small as possible. It has an intercept and a slope, and the slope is a regression coefficient, also called the slope coefficient or the beta coefficient. The unexplained part is the error term, whose spread is measured by the variance and the standard deviation of the residuals.
Worked Example 1: A Strong Model
This worked example runs step-by-step. Suppose the sum of squared residuals is 20 and the sum of squares total is 100, so SSres = 20 and SStot = 100. Divide 20 by 100 to get 0.20, then subtract from 1 to get 1 – 0.20 = 0.80. The result is 0.80, or 80%, so the model explains 80% of the variance and leaves 20% unexplained.
Worked Example 2: Study Hours and Test Scores
Take study hours and test scores again. Suppose the total sum of squares for the test scores is 180, and the residual sum of squares after fitting the line is 45, so SSres = 45 and SStot = 180. Divide 45 by 180 to get 0.25, then 1 – 0.25 = 0.75. The result is 0.75, or 75%, so study hours explain 75% of the variation in test scores, and 25% comes from other factors. The Pearson correlation between study hours and test scores, the correlation coefficient r, is 0.866 here, and the squared correlation is 0.866 x 0.866 = 0.75.
Worked Example 3: A Weak Model and the Adjusted Value
For a weak model, suppose SSres = 90 and SStot = 100. Divide 90 by 100 to get 0.90, then 1 – 0.90 = 0.10. The result is 0.10, or 10%, so the model explains only 10% of the variation. Now suppose a model with three predictors has R-squared of 0.75 on a sample size of 30. The adjusted R-squared, which corrects for the number of predictors and the degrees of freedom, is 1 – [(1 – 0.75) x (30 – 1) / (30 – 3 – 1)] = 1 – [0.25 x 29 / 26] = 1 – 0.279 = 0.721, which rounds to 0.72.
How Do You Read an R-Squared Value?
To read an R-squared value, multiply it by 100 to turn it into a percentage of variance. An R-squared value of 0.80 becomes 80%, an R-squared value of 0.50 becomes 50%, and an R-squared value of 0.25 becomes 25%. That percentage is the proportion of variance the model explains, and the rest is the share it does not explain. It is a share of variance, not a measure of prediction accuracy.
On a scatter plot, a high R-squared shows up as data points hugging the line of best fit, with only small gaps between the points and the regression line. A low R-squared shows up as a wide cloud of points around the line, with large residuals in both directions. A five-number summary can show that spread, and the 5 Number Summary Calculator gives it in seconds.
| R-squared value | What it means | Everyday reading |
|---|---|---|
| 0.00 | The model explains none of the variance | No better than guessing the average |
| 0.10 | The model explains 10% of the variance | A weak pattern, with most variation unexplained |
| 0.50 | The model explains 50% of the variance | Half the variation is explained, half is not |
| 0.80 | The model explains 80% of the variance | A strong pattern, with most variation explained |
| 1.00 | The model explains all of the variance | A perfect fit, which is rare in real data |
The table shows why a single number needs context. A value of 0.95 in a controlled laboratory study might be ordinary, while the same value in a survey of human behaviour would be surprising.
What Is a Good R-Squared Value?
There is no single good R-squared value, because what counts as good depends on the field of study and on what the model is for. In physics, where experiments are tightly controlled, values above 0.90 are common, and 0.70 might look weak. In social science and psychology, values between 0.10 and 0.30 are common, and 0.30 can be a genuinely useful finding. The Duke University guide to R-squared makes the same point: the right benchmark depends on the subject, the data, and the question.
In economics and finance, a model of investment returns on the stock market will usually have a modest R-squared, because prices move for countless reasons no single model can capture, while a fund tracking its benchmark closely has a high R-squared against that benchmark. A model of sales against advertising spend may score 0.60 in one market and 0.20 in another. When outcomes vary a lot, relative spread matters too, and the Coefficient of Variation Calculator can compare that spread across different scales.
What Is the Difference Between R-Squared and Adjusted R-Squared?
Adjusted R-squared is a corrected version of R-squared that takes the number of predictors and the sample size into account. Ordinary R-squared never falls when you add a predictor, even if that predictor is useless, because adding a variable can only make the line fit the training data a little better. Adjusted R-squared, written as adjusted R², penalises each extra predictor, so it only rises when the new predictor improves the fit by more than you would expect by chance. That makes it the fairer number for model comparison, and for model evaluation, when two models use different numbers of predictors.
The difference matters most in multiple regression, where several explanatory variables are used to predict one response variable. In a multiple linear regression with many variables, the gap can be large. Model selection is therefore safer with the adjusted value, and variable selection should follow it too: adding variables that do not help should be avoided, and removing variables that do not help can raise the adjusted value even while the ordinary value falls slightly. The adjusted formula uses the degrees of freedom, which shrink as the number of predictors grows relative to the sample size.
What Are the Most Common R-Squared Mistakes?
Mistake 1: Treating a High R-Squared as Proof of Causation
A high R-squared does not prove causation, because correlation does not imply causation. Two variables can move together closely without one causing the other, and the link can be driven by a third factor. The classic example is temperature and ice cream sales: both rise in summer, so they are closely related, yet buying ice cream does not cause hot weather. A confounding variable of that kind can make a model look strong while the causal relationship it suggests is false, and a spurious correlation can appear whenever two things share a season, a trend, or a common cause. Height and weight, and income and education, show the same trap.
Mistake 2: Adding Variables Until R-Squared Looks Good
Adding variables will always raise ordinary R-squared, even when the new variables are random, so a model can be made to look good by piling in predictors. That is overfitting: the model learns the quirks of the training data instead of the real pattern, and it then performs poorly on test data it has never seen. The opposite error is underfitting, where the model is too simple to capture the real pattern at all. The cure is validation, such as cross-validation on data held back from training, and careful variable selection.
Mistake 3: Judging a Model by R-Squared Alone
Judging a statistical model by R-squared alone misses model performance and model quality, so look at the residuals as well. A residual plot, which plots each residual against its fitted value, should show a shapeless cloud, and residual analysis of that plot can reveal problems that R-squared hides. If the spread grows as the fitted values grow, the errors show heteroscedasticity, while a steady spread is homoscedasticity, and a curved pattern can mean the straight line is the wrong shape. The normality of the residuals matters for some tests, and a few outliers or influential points with high leverage can pull the line toward themselves and inflate R-squared.
The other checks are just as important. The standard error of the predictions, the root mean square error, shortened to RMSE, the mean squared error, shortened to MSE, and the mean absolute error, shortened to MAE, all measure error in the original units, and they tell you how far predictions are likely to be wrong, which R-squared cannot do. A p-value, a significance level, and a hypothesis test tell you about statistical significance, whether a predictor is likely to matter at all, and the F-statistic tests the model as a whole while the t-statistic tests a single coefficient. A confidence interval shows the range of plausible values for a coefficient, a prediction interval shows the range for a single new prediction, and an ANOVA table brings the F-statistic and R-squared together. The Minitab guide to interpreting R-squared puts it well: a good predictive model needs a sensible R-squared, small errors, and residuals that behave.
How Can You Check R-Squared With a Calculator?
If you are still asking what does r squared mean for your own numbers, try them in the calculator and watch the percentage change as the residuals fall. The Coefficient of Determination Calculator is the free R-squared calculator on this site: enter your sums of squares, and it returns the R-squared value at once, so you can check a worked example or your own data without doing the division by hand. This coefficient of determination calculator tests the examples above.
Squaring a value is at the heart of this topic, and it appears in other everyday calculations too. The Right Triangle Calculator squares the sides of a triangle, because the square of the hypotenuse equals the sum of the squares of the other two sides. The AP Lang Score Calculator works with exam scores, the kind of outcome a regression model might try to predict from study time, while a running model might predict a finish time, as the Negative Split Calculator does for runners. In finance, the Credit Card Interest Calculator and the SIP Calculator UK work with interest and growth a model might predict, and the how much you need to retire guide shows how a long-term prediction is built. You can find more guides like this one in the Daily Life Blog category on Daily Life Calculators.
Frequently Asked Questions About R-Squared
Can R-squared be negative?
Yes, in some models R-squared can be negative, which means the model fits worse than simply predicting the mean of the observed values every time. That usually happens when a model is forced through a fixed point, when it is tested on new data, or when it has no intercept. In ordinary least squares with an intercept, R-squared sits between 0 and 1, so a negative value is a warning.
Is an R-squared of 0.5 good?
An R-squared of 0.50 can be good, ordinary, or weak, depending on the field. In physics it would usually be low, in social science it can be a solid result.
What is the difference between r and R-squared?
The letter r is the correlation coefficient, which runs from -1 to 1 and shows both the strength and the direction of a straight-line relationship. R-squared is the square of r in simple linear regression, so it runs from 0 to 1 and shows only the strength, not the direction. A correlation of -0.80 and 0.80 both give an R-squared of 0.64, because squaring removes the sign.
Can R-squared be 1?
An R-squared of 1 means a perfect fit, with every point exactly on the line and no residuals at all. In real data that is almost always a sign of a problem, such as a predictor that is really the outcome in disguise, a model with as many predictors as data points, or data copied by mistake.
Does a low R-squared mean my model is useless?
No. A low R-squared does not always mean a bad model, especially when the outcome is hard to predict or when the goal is to test whether a predictor matters rather than to predict exact values. A model with a low R-squared can still show a real, statistically significant pattern, and it can still be useful for ranking, for planning, or for further research.
Disclaimer: This article is for general information only and does not give financial, medical, or professional advice. Use the calculator and the examples to learn the idea, and check important decisions with a qualified professional.







