Learn & Understand

Why a High R-Squared Can Fool You

The companion calculator reports R-squared, the fraction of variance a model explains, a headline figure for regression quality. It is genuinely useful, but it is also one of the most misinterpreted numbers in statistics, because a high R-squared does not mean a model is good, correct, or trustworthy. R-squared can be inflated by adding useless features, can be high on a badly wrong model, and says nothing about whether the predictions will hold up. Understanding its traps keeps a reassuring number from creating false confidence.

It Always Goes Up With More Features

The first trap is that R-squared can only increase, never decrease, as you add more predictors to a regression, even if those predictors are pure noise. Add enough random variables and R-squared climbs toward a perfect fit, not because the model captures anything real, but because extra parameters let it contort to the training data. This means a high R-squared achieved by piling on features is often a sign of overfitting, a model memorizing noise rather than learning signal, and it will generalize poorly to new data. Comparing models by raw R-squared unfairly favors the one with more variables.

Adjusted R-Squared

The fix for this specific trap is adjusted R-squared, which penalizes the addition of predictors that do not improve the fit enough to justify their inclusion.

Adjusted R-squared can actually go down when a feature adds more complexity than value, so it gives a fairer comparison between models with different numbers of predictors. When deciding whether a new variable earns its place, adjusted R-squared is the honest referee that plain R-squared cannot be.

A Good Fit Isn't a Correct Model

The deeper trap is that R-squared measures how much variance is explained, not whether the model is appropriate. A famous illustration, Anscombe's quartet, shows four completely different datasets, one linear, one curved, one dominated by a single outlier, that produce nearly identical regression statistics, including R-squared. The number cannot tell them apart, yet a straight-line model is right for only one of them. This is why R-squared must never replace actually looking at the data and the residuals: a high R-squared on a fundamentally mis-specified model, fitting a line to a curve, is a high score for the wrong answer.

High R-Squared, Poor Predictions

Finally, a strong R-squared on the data you fit does not guarantee good predictions on new data. R-squared is typically measured on the training data, where overfitting flatters it, so it can look excellent while the model fails out-of-sample. It also gives no license to extrapolate: a model that fits well within the observed range can be wildly wrong outside it. And R-squared says nothing about causation, a high value only means the predictors track the outcome, not that they cause it. For real predictive quality, performance on held-out data matters far more than R-squared on the training set.

Reading R-Squared With Caution

Use the calculator's R-squared as one indicator of fit, not a verdict on model quality. Prefer adjusted R-squared when comparing models of different complexity, always inspect the data and residuals rather than trusting the number alone, and judge real predictive value on held-out data. The calculation reports variance explained; understanding its traps is what keeps a high R-squared from fooling you.