Is explained variation always a percentage?

No. It can refer to a sum of squares or variance component; a ratio gives a proportion.

Can training R² rise with an irrelevant predictor?

Yes. More flexible fitting can capture noise.

AI Glossary / Statistics and probability / Intermediate

Explained variation

Definition

Explained variation is the part of variation in an outcome that is accounted for by a specified statistical model or set of predictors. In ordinary least-squares regression with an intercept, it commonly refers to the regression sum of squares, the variation of fitted values around the observed mean. Dividing this by total variation gives R². The word “explained” describes a statistical relationship and does not, by itself, establish causation.

What it means

Outcomes differ across observations. A model can represent some of those differences systematically, leaving residual differences between observed and fitted values. Explained variation summarizes the part represented by the fitted model under a stated decomposition.
The quantity depends on the observations, predictors, fitting procedure, and definition of variation. It is not an intrinsic percentage of an outcome that science has permanently understood. A different sample or model can give a different answer.

How it works

For observed outcomes y_i, calculate their mean ȳ. The total sum of squares is SST = Σ(y_i − ȳ)². It measures variation around a model that predicts the same sample mean for everyone.
Fit an ordinary least-squares linear model that includes an intercept. Its residual sum of squares is SSE = Σ(y_i − ŷ_i)². Its regression, or explained, sum of squares is SSR = Σ(ŷ_i − ȳ)².
For this fitted model on its fitting data, the residuals are orthogonal to the fitted values and sum to zero. The cross term vanishes, giving SST = SSR + SSE. If SST is positive, R² = SSR/SST = 1 − SSE/SST.
Those equalities require their conditions. For arbitrary predictions, a model without an intercept, or evaluation on new data, the same orthogonal decomposition need not hold. A predictive score computed as 1 − SSE/SST can be negative, indicating worse squared-error performance than predicting that evaluation sample's mean.
At a population level, another useful decomposition is Var(Y) = Var(E[Y|X]) + E[Var(Y|X)], assuming finite variance. The first component describes variation in the conditional mean. An estimated model may approximate that conditional mean imperfectly, so its sample fit is not automatically this population quantity.
Adding predictors to an ordinary least-squares model cannot worsen its training residual sum of squares when the expanded model contains the original. Training R² can therefore increase even for unhelpful variables. Assess complexity, residual patterns, and performance on suitable new data rather than selecting models only by explained variation.
Do not use the named explained variance score interchangeably with every explained-variation measure. That prediction metric subtracts a residual-variance ratio and ignores a constant prediction offset; ordinary R² generally does not.

Three practical examples

A service analyst reports how much observed variation in processing time is represented by a fitted workload model.
A manufacturing study separates variation associated with fitted operating factors from residual variation.
An AI evaluation compares model fit with held-out prediction performance to check whether apparent explanatory strength survives beyond training data.

Think of it this way

Think of a map that captures some features of a landscape. The map's coverage of observed patterns is useful, but it does not establish why the landscape was formed or guarantee accuracy in another region.

Worked example

Consider invented data x = [0, 1, 2, 3] and y = [1, 3, 2, 4]. The least-squares line with an intercept is ŷ = 1.3 + 0.8x, giving fitted values [1.3, 2.1, 2.9, 3.7]. The observed mean is 2.5.
Total variation is SST = 2.25 + 0.25 + 0.25 + 2.25 = 5. Explained variation is SSR = 1.44 + 0.16 + 0.16 + 1.44 = 3.2. Residuals are [−0.3, 0.9, −0.9, 0.3], so SSE = 1.8.
The decomposition checks: 3.2 + 1.8 = 5. R² = 3.2/5 = 0.64. Thus the fitted linear relationship accounts for 64% of the sample's squared variation around its mean.
This does not mean 64% of individual predictions are correct or that x causes 64% of y. The sums of squares have squared outcome units; their ratio is dimensionless.

Illustrative case study

Fictional case: Pine Logistics reports a high training R² for a model of delivery duration. Reviewers find that some predictors were recorded only after deliveries finished. The team removes these unavailable inputs and evaluates the revised model on later deliveries. This invented case illustrates how apparently explained variation can reflect information leakage rather than useful advance knowledge.

Common misconceptions and mistakes

Do not equate explained variation with a causal share, prediction accuracy percentage, or probability that the model is correct. A high score can coexist with biased data, poor extrapolation, and important residual patterns.
Do not assume all sum-of-squares decompositions apply to arbitrary model predictions. State the intercept and fitting conditions.
A low R² need not make an estimated relationship scientifically unimportant, especially when outcomes contain substantial individual variability.

Business use

Explained variation can communicate model fit and help compare nested analyses on the same data. Business users also need errors in practical units, performance across relevant groups, and the availability of predictors when decisions are made. A model that explains historical variation may still be unsuitable for intervention or forecasting.

Check your understanding

Which identity holds in the worked example?

Answer: SST = SSR + SSE.

Why is the intercept condition important?

Answer: It supports the centered orthogonal decomposition for ordinary least squares.

What does 0.64 describe?

Answer: A proportion of squared sample variation represented by that fitted line.

Related terms

Learn next

Learn the geometry of least squares and the law of total variance. Then compare training R², adjusted R², and held-out predictive scores.

Least squares

Sources

Penn State, “The Coefficient of Determination, r-squared”: https://online.stat.psu.edu/stat462/node/95/ NIST/SEMATECH, “Least Squares”: https://www.itl.nist.gov/div898/handbook/pmd/section4/pmd431.htm Independent review consulted NIST above and Hyndman and Athanasopoulos, regression goodness-of-fit and residual properties: https://otexts.com/fpp3/regression.html . The Penn State page returned an access error during this review.

Previous: Explained variance score Next: Explainer