The Statistical Definition of R-Squared (Coefficient of Determination)

In linear regression analysis, econometric modeling, and machine learning, the Coefficient of Determination (R²) quantifies the proportion of total variance in the dependent response variable ($) that is predictable from the independent explanatory variables ($).

R² ranges from 0.0 (0%) — indicating that the regression model explains none of the variation around the mean — to 1.0 (100%) — indicating that the model's fitted predictions perfectly track observed data with zero residual error.

The Sum of Squares Decomposition

The mathematical foundation of R² is derived from partitioning total variation into explained and residual components:

1. Total Sum of Squares Decomposition:
SS_total = SS_regression + SS_residual
• SS_total (Total Variation): ∑ (y_i − ȳ)²
• SS_regression (Explained Variation): ∑ (ŷ_i − ȳ)²
• SS_residual (Unexplained Error): ∑ (y_i − ŷ_i)²

2. R-Squared Formulas:
R² = SS_regression / SS_total = 1 − [ SS_residual / SS_total ]

3. Simple Linear Regression Equivalence:
In simple single-variable OLS regression, R² equals the square of Pearson's correlation coefficient (r):
R² = (r_xy)²

R-Squared Benchmarks Across Scientific and Financial Disciplines

Discipline / Domain	Typical Expected R² Range	Analytical Context & Modeling Standard
Physics & Laboratory Chemistry	0.95 – 0.999	Controlled experimental systems governed by deterministic physical laws with minimal noise.
Automotive & Mechanical Dyno Testing	0.90 – 0.98	Crankshaft torque vs RPM sweeps, mass airflow calibration curves.
Finance: Index Mutual Funds vs S&P 500	0.95 – 0.99	Confirms pure passive index tracking; low active management risk.
Finance: Active Equity Hedge Funds	0.30 – 0.70	Indicates idiosyncratic stock-picking alpha with low market factor tracking.
Social Sciences, Psychology, & Economics	0.10 – 0.40	Human behavioral data with high individual variance; low R² models can still have highly statistically significant predictor variables ( < 0.001$).

R-Squared Traps: Anscombe's Quartet and Non-Linear Data

A high R² value does not guarantee that a linear model is appropriate or valid:

Anscombe's Quartet: Four distinct datasets constructed by statistician Francis Anscombe share identical linear regression lines and identical R² values (0.67), despite one dataset being parabolic, another having an extreme outlier, and a third having a vertical line. Always visually inspect residual plots.
Non-Linearity: Fitting a straight line to a U-shaped quadratic curve produces a low R², even though the relationship between $ and $ is 100% deterministic.

Step-by-Step Practical Calculation: Regression Model Fit

A model predicts house prices (Y) from square footage (X). Total sum of squares SS_tot = 500,000, and sum of squared residual errors SS_res = 75,000:

Step 1: Calculate Explained Sum of Squares:
SS_reg = SS_tot − SS_res = 500,000 − 75,000 = 425,000.
Step 2: Calculate R-Squared:
R² = 425,000 / 500,000 = 0.8500 (85.0%).
Interpretation: 85.0% of the variation in house selling prices is explained by home square footage; the remaining 15.0% is driven by unmeasured factors (neighborhood, lot size, renovation age).

Frequently Asked Questions About R-Squared

Can R-Squared ever be negative?

In standard OLS linear regression with an intercept term, R² is mathematically constrained between 0.0 and 1.0. However, if a regression model is forced through the origin (no intercept) or tested out-of-sample on new validation data where the model performs worse than a simple horizontal average line, R² can become negative.

What is the difference between R-Squared and Adjusted R-Squared?

Standard R² increases automatically whenever any new predictor variable is added to the model, even if the variable is pure random noise. Adjusted R² penalizes model complexity based on degrees of freedom, rising only when a new variable improves the model more than expected by chance.

Does a low R-Squared mean a regression model is useless?

No. In medical clinical trials and econometric studies, human health outcomes have high natural variance (R² ≈ 0.15), yet finding that a medication or policy intervention has a statistically significant t-statistic (p < 0.01) provides vital, life-saving conclusions.

How does R-Squared relate to Active Share in mutual funds?

In portfolio management, an equity mutual fund with an R² > 0.95 relative to the S&P 500 is classified as a "closet indexer" (charging active management fees while merely replicating the index). Genuine active managers exhibit lower R² values (0.60 to 0.85).

Matrix Formulation of R-Squared in Multiple Linear Regression

In generalized multiple linear regression (Y = Xβ + ε), the explained sum of squares is computed using the orthogonal Projection Hat Matrix (H):

Matrix Algebra Formulation:
Hat Matrix (H) = X(X^TX)⁻¹X^T
Fitted Values Ŷ = HY | Residual Vector e = (I − H)Y
SS_tot = Y^T(I − (1/n)J)Y
SS_res = e^Te = Y^T(I − H)Y

R² = [ Y^T(H − (1/n)J)Y ] / [ Y^T(I − (1/n)J)Y ]
where J is an n × n matrix of all ones, and I is the identity matrix.

Spurious Regression in Non-Stationary Time Series

A critical hazard in macroeconomic and financial econometrics is Spurious Regression, identified by Nobel laureates Clive Granger and Paul Newbold:

Econometric Condition	Underlying Data Characteristics	Observed R² & t-Statistics	Diagnostic & Remediation Protocol
Spurious Time-Series Regression	Two completely unrelated non-stationary random walks ((1)$ series with unit roots, e.g., US GDP vs solar sunspot count).	Artificially massive R² (> 0.85) and extreme t-statistics ( > 10$), despite zero causal connection.	Check Durbin-Watson statistic (DW < R²). Take first differences (ΔY, ΔX) or test for cointegration (Engle-Granger test).
Cointegrated Regression	Two non-stationary series that share a genuine long-run equilibrium relationship (e.g., spot price vs futures price).	High valid R²; residual errors are stationary ((0)$).	Fit Vector Error Correction Models (VECM) to model short-term adjustments and long-term equilibrium.

Multicollinearity and Variance Inflation Factors (VIF)

When independent explanatory variables are highly correlated with each other, standard R² remains high, but individual regression coefficients become unstable with inflated standard errors:

Variance Inflation Factor (VIF) Formula:
VIF_j = 1 / (1 − R_j²)
where R_j² is the coefficient of determination obtained by regressing predictor $ on all other remaining predictor variables.
• VIF = 1.0: Complete orthogonality (zero correlation).
• VIF > 5.0 to 10.0: Severe multicollinearity; standard errors are inflated by √VIF, destabilizing hypothesis testing.

Non-Linear R-Squared and Pseudo R-Squared in Logistic Regression

In non-linear regression models (such as binary logistic regression or probit models), standard sum-of-squares R² is mathematically invalid. Statisticians utilize Pseudo R-Squared metrics based on log-likelihoods:

Pseudo R² Formulation	Mathematical Definition	Interpretation & Scaling
McFadden's Pseudo R²	R²_McFadden = 1 − [ ln(L_full) / ln(L_null) ]	Values between 0.20 and 0.40 represent exceptional model fit (equivalent to 0.70–0.90 in OLS).
Cox & Snell R²	R²_CS = 1 − [ L_null / L_full ]^{(2 / n)}	Cannot reach 1.0 (maximum value is bounded below 1.0).
Nagelkerke / Cragg-Uhler R²	R²_Nagelkerke = R²_CS / [ 1 − (L_null)^{(2 / n)} ]	Scales Cox & Snell to span the complete 0.0 to 1.0 range.

R-Squared Variance Decomposition: Shapley Value and LMG Metric

When predictor variables are correlated, determining the exact percentage of R² contributed by each individual predictor is accomplished via Shapley Value Regression (LMG Decomposition):

LMG Decomposition Principle:
The Lindeman, Merenda, and Gold (LMG) metric computes the sequential increase in R² when predictor $ is added across all possible $2^{k-1}$ variable subset orders, averaging across all permutations:
∑ LMG_j = Total Model R²
This cleanly decomposes total explained variance among correlated predictors without ordering bias.

The 10-Point Statistical Regression and Goodness-of-Fit Protocol

Inspect Scatter Plots Before Regressing: Visually verify linearity; never rely on R² alone to confirm model validity.
Examine Residual Error Plots: Confirm that residuals exhibit zero mean, constant variance (homoscedasticity), and absence of non-linear patterns.
Check for Outliers and High-Leverage Points: Calculate Cook's Distance ( > 4/n$) to identify single data points artificially inflating R².
Audit Variance Inflation Factors: Ensure all predictor VIFs are < 5.0 to eliminate multicollinearity distortion.
Verify Stationarity in Time-Series Models: Run Augmented Dickey-Fuller (ADF) tests to eliminate spurious time-series regressions.
Compare Against Adjusted R-Squared: Ensure raw R² gains are matched by Adjusted R² improvements.
Perform Cross-Validation: Calculate Predicted R² (Q²) via Leave-One-Out cross-validation to detect in-sample overfitting.
Differentiate Predictive vs Explanatory Power: A low R² model can provide vital causal insights, while a high R² model can be purely correlational without causal truth.
Report Confidence Intervals for Predictors: Accompany R² with regression coefficients, standard errors, t-statistics, and exact p-values.
Validate on Holdout Test Data: Evaluate final model R² on an independent out-of-sample test split (e.g., 80% train / 20% test).

Detailed R-Squared FAQs

What is the relationship between R-Squared and Pearson Correlation (r)?

In simple bivariate linear regression (Y = β₀ + β₁X), R-Squared is mathematically identical to the square of Pearson's correlation coefficient: R² = r². If the correlation between two variables is r = 0.80, the regression model explains R² = (0.80)² = 0.64 (64%) of the variance.

Why can R-Squared be high when all t-statistics are insignificant?

This is the classic symptom of extreme multicollinearity. The predictors collectively explain the response variable (R² > 0.80, significant overall F-test), but because the predictors are heavily correlated with each other, the regression algorithm cannot isolate which individual variable deserves credit, causing individual t-statistics to appear statistically insignificant.

Does R-Squared measure the slope of the regression line?

No. R-Squared measures the tightness of the scatter around the fitted line (goodness of fit), not the steepness of the slope. A nearly horizontal line with tightly clustered points has a high R², while a steep slope with wide point scatter has a low R².

What is the difference between R-Squared in OLS vs Machine Learning Random Forests?

In OLS, R² is computed in-sample on training data. In machine learning (Random Forests, Gradient Boosting), R² is evaluated out-of-sample on unseen validation test sets to measure generalization accuracy.

Can a transformation of variables (like Log-Transform) increase R-Squared?

Yes. If the underlying physical or financial relationship is multiplicative or exponential (e.g., compound growth, population expansion), log-transforming the variables (ln(Y) = β₀ + β₁ × ln(X)) linearizes the data, dramatically increasing R² and stabilizing residual variances.

How does sample size affect R-Squared?

In very small sample sizes (e.g., = 3$), standard R² is mathematically biased upward. As sample size increases to large numbers ( > 1,000$), sample R² converges asymptotically to the true population coefficient of determination.

ANOVA Table Structure and the Global F-Test

In regression reporting, the relationship between R-Squared, Sum of Squares, and the overall ANOVA Global F-Test is formalized through the analysis of variance table:

Global F-Statistic Formula:
F = [ SS_reg / k ] / [ SS_res / (n − k − 1) ] = [ R² / k ] / [ (1 − R²) / (n − k − 1) ]

Interpretation: The F-test evaluates the joint null hypothesis that all k slope coefficients are simultaneously zero (H₀: β₁ = β₂ = ... = β_k = 0). Even when individual t-statistics are borderline, a high R² combined with a statistically significant F-statistic (p < 0.001) confirms that the regression model explains significant collective variance.

Partial R-Squared and Incremental Explanatory Power

When evaluating the added value of a specific predictor variable ($) controlling for existing predictors ($), econometricians calculate the Partial Coefficient of Determination (r²_{Y,X1 | X2}):

Partial R-Squared Formula:
Partial R² = [ SS_res(reduced) − SS_res(full) ] / SS_res(reduced) = [ R²_full − R²_reduced ] / [ 1 − R²_reduced ]

Example: A base model with 3 variables has ^2 = 0.70$. Adding a 4th variable increases ^2$ to $0.76$:
Partial R² = [ 0.76 − 0.70 ] / [ 1 − 0.70 ] = 0.06 / 0.30 = 0.200 (20.0% of remaining unexplained variance explained).

Additional R-Squared FAQs

What is the difference between R-Squared and Goodness-of-Fit?

R-Squared is one specific quantitative metric of goodness-of-fit measuring variance explained. Comprehensive goodness-of-fit evaluation also requires analyzing residual error distributions, normality tests (Jarque-Bera), heteroscedasticity tests (White test), and out-of-sample cross-validation.

Can two datasets have identical R-Squared but completely different predictive accuracy?

Yes. If Dataset A has a total variance of SS_tot = 100 and R² = 0.80, its residual error is SS_res = 20. If Dataset B has SS_tot = 10,000 and R² = 0.80, its residual error is SS_res = 2,000. The absolute prediction errors in Dataset B are 100 times larger despite identical R² values.

Why is R-Squared in time-series models often artificially higher than in cross-sectional models?

Time-series economic data often shares strong underlying macroeconomic trends (GDP growth, population expansion, inflation). Cross-sectional individual micro-data contains massive person-to-person idiosyncratic variance, naturally yielding lower R² values.

What is the coefficient of non-determination?

The coefficient of non-determination is ^2 = 1 - R^2$, representing the exact proportion of total variance left unexplained by the regression model.

Advanced Regression Specification Diagnostics: Ramsey RESET and White's Test

A high R-Squared does not guarantee that a regression model is properly specified. Econometricians apply rigorous post-estimation diagnostic hypothesis tests:

1. Ramsey RESET Test (Functional Form Misspecification):
Regresses residuals on powers of fitted values: Y = Xβ + γ₁Ŷ² + γ₂Ŷ³ + ε.
If the F-test for γ₁ = γ₂ = 0 is rejected (p < 0.05), the relationship is non-linear, and high linear R² is illusory.

2. White's General Heteroscedasticity Test:
Regresses squared residuals (e_i²) on original predictors, cross-products, and squared predictors.
LM = n × R²_auxiliary ∼ χ²(df). Rejection indicates non-constant error variance, requiring White-Huber robust standard errors.

R-Squared Troubleshooting and Modeling Diagnostics Matrix

Observed R² Diagnostic Anomaly	Underlying Model Defect	Statistical Consequence	Remediation Protocol
Extremely High R² (> 0.98) with Spurious T-Statistics	Non-stationary unit roots in time-series data; trend persistence.	Spurious regression; false causal attribution; severe forecast failure.	First-difference variables (ΔY, ΔX) or fit Vector Error Correction Models (VECM).
High R² (> 0.85) but None of the Individual Predictors are Significant	Severe multicollinearity among explanatory predictors.	Unstable coefficient signs and massive standard error inflation.	Calculate VIF; drop redundant collinear variables or apply LASSO/Ridge regularization.
R² Increases Substantially but Adjusted R² Drops	Added predictors are pure noise and fail the \|t\| > 1.0 threshold.	Model degrees of freedom penalty exceeds minuscule explained variance.	Prune non-significant variables to restore parsimonious model specification.
Low R² (< 0.10) with Highly Significant P-Values (p < 0.001)	High natural population variance in cross-sectional or epidemiological data.	Model captures a genuine causal effect amidst high baseline noise.	Valid for scientific inference; report effect sizes and confidence intervals rather than raw R².

Glossary of Econometric Goodness-of-Fit Terminology

Coefficient of Determination (R²):: Proportion of total variation in the dependent variable explained by the regression model (SS_reg / SS_tot).
Sum of Squares Total (SS_tot):: Total squared dispersion of observed dependent values around their sample mean (∑(y_i − ȳ)²).
Sum of Squares Residual (SS_res):: Unexplained error variation remaining after fitting the regression plane (∑e_i²).
Projection Hat Matrix (H):: Orthogonal linear algebra operator X(X^TX)⁻¹X^T mapping observed vectors onto fitted hyperplane coordinates.
Anscombe's Quartet:: Famous set of four distinct datasets with identical summary statistics (mean, variance, correlation, R² = 0.67) but completely different physical geometries.
Cook's Distance:: Diagnostic metric identifying influential outlier observations that exert excessive leverage on regression coefficients and R².
Homoscedasticity:: Core Gauss-Markov assumption that error disturbances maintain constant finite variance across all values of predictor variables.
Pseudo R-Squared (McFadden):: Log-likelihood ratio metric evaluating goodness of fit in non-linear binary logistic and multinomial logit models.

Geometric Interpretation of R-Squared in Hilbert Vector Spaces

In linear algebra and geometric statistics, multiple linear regression is viewed as an orthogonal projection in n-dimensional Euclidean space (ℝⁿ):

The Response Vector (Y): An n-dimensional vector representing observed data points in sample space.
The Column Space of X: A k-dimensional linear subspace spanned by the predictor vectors (X₁, X₂, ..., X_k).
The Fitted Vector (Ŷ): The orthogonal projection of Y onto the subspace of X, minimizing Euclidean distance ||Y − Ŷ||.
The Geometric R-Squared: Equal to the squared cosine of the geometric angle (θ) between the mean-centered response vector (Y − Ȳ) and the fitted vector (Ŷ − Ȳ):
R² = cos²(θ).
When the angle θ = 0°, cos(0) = 1.0 and R² = 1.0 (perfect fit). When θ = 90° (orthogonal vectors), cos(90°) = 0 and R² = 0.0 (zero explanatory power).

Historical Origins: Sewall Wright's Path Analysis and Ronald Fisher

The coefficient of determination was pioneered in the 1920s by evolutionary geneticist Sewall Wright (1921) in the development of Path Analysis to partition genetic inheritance correlations. Statistician Sir Ronald A. Fisher (1925) integrated R² directly into the Analysis of Variance (ANOVA) framework, establishing the sum-of-squares decomposition that forms the backbone of modern econometric regression.

Executive Summary: Guidelines for Rigorous R-Squared Interpretation

To avoid common statistical traps, follow these foundational guidelines when analyzing regression R²:

Never Equate High R² with Causality: Correlation and variance explanation do not establish causal mechanisms.
Check Residual Error Plots First: Verify homoscedasticity, normality, and absence of non-linear curvature before interpreting R².
Use Adjusted R² for Multi-Variable Models: Prevent artificial inflation caused by adding redundant predictor features.
Verify Out-of-Sample Performance: Confirm in-sample R² translates to high Predicted R² (Q²) on independent holdout test data.

Empirical Applications: R-Squared Benchmarks by Scientific Discipline

Expected and acceptable R-Squared thresholds vary fundamentally across scientific and industrial disciplines depending on baseline physical noise:

Astrophysics and Mechanical Engineering: R² > 0.98 expected due to precise deterministic physical laws (Newtonian mechanics, thermodynamics).
Chemistry and Pharmacology: R² > 0.90 expected in controlled laboratory assays and spectroscopic calibration curves.
Macroeconomics and Financial Econometrics: R² between 0.30 and 0.70 typical for quarterly forecasting models; higher values frequently indicate non-stationary spurious regressions.
Microeconomics, Psychology, and Medicine: R² between 0.05 and 0.25 common due to massive individual behavioral and genetic diversity, where modest R² values still provide decisive, life-saving medical insights.

R-Squared Calculator

Related Statistics Calculators