• R-squared, or the coefficient of determination, measures how much of the variation in a dependent variable (Y) is explained by one or more independent variables (X) in a statistical model.
– It is commonly expressed between 0 and 1 (or 0%–100%). A value closer to 1 means the model explains a larger share of the variance.
Key formula
– R2 = 1 − (Unexplained variation / Total variation)
– In regression terms:
• TSS (Total Sum of Squares) = Σ (yi − ȳ)2
• RSS or SSE (Residual Sum of Squares / Unexplained) = Σ (yi − ŷi)2
• R2 = 1 − RSS / TSS
What R-squared tells you (and what it doesn’t)
– Tells you: the proportion of the dependent-variable variance explained by the independent variable(s).
• Example: R2 = 0.50 implies about 50% of observed variation in Y is explained by the model inputs.
– Does NOT tell you: causation, whether the model is correctly specified, whether predictions are unbiased, or whether the model will generalize to new data.
Practical steps — how to calculate R-squared (manual / conceptual)
1. Collect paired observations (yi, xi1, xi2, …).
2. Fit the regression model (e.g., ordinary least squares) to get predicted values ŷi.
3. Compute the mean of actual outcomes ȳ = (1/n) Σ yi.
4. Compute TSS = Σ (yi − ȳ)2.
5. Compute RSS = Σ (yi − ŷi)2.
6. Compute R2 = 1 − RSS / TSS.
7. Interpret R2 in context (see interpretation guidance below).
Quick calculation examples
– Simple numeric example:
• Observations y = [3, 4, 5], predicted ŷ = [2.8, 4.1, 5.0].
• ȳ = 4.0, TSS = (3−4)2 + (4−4)2 + (5−4)2 = 1 + 0 + 1 = 2.
• RSS = (3−2.8)2 + (4−4.1)2 + (5−5)2 = 0.04 + 0.01 + 0 = 0.05.
• R2 = 1 − 0.05 / 2 = 0.975 (97.5% of variance explained).
How to calculate in common tools
– Excel: fit regression via Data Analysis → Regression (or use LINEST), then read R Square in output.
– Python (statsmodels):
• Fit model: model = sm.OLS(y, X).fit()
• Get R2: model.rsquared
– R:
• fit 0.9).
• Social sciences: lower R2 (e.g., 0.2–0.5) may still be meaningful.
• Investing: R2 of 85%–100% means asset/fund moves closely with the benchmark; ≤70% implies weak tracking.
Can R-squared be negative?
– Yes. While R2 is usually between 0 and 1 for models with an intercept and using the standard definition, it can be negative when:
• The model fits worse than the horizontal mean line (RSS > TSS).
• Models are fit without an intercept.
• Alternative R2 definitions or poor-specified models are used.
– Adjusted R2 can also be negative if the model fits poorly relative to the mean.
R-squared vs. adjusted R-squared
– Problem with plain R2: it never decreases when you add variables, even if they add nothing useful—this can encourage overfitting.
– Adjusted R2 penalizes added predictors and only increases if a new variable improves the model more than would be expected by chance.
– Formula (common form):
• Adjusted R2 = 1 − (1 − R2) × (n − 1) / (n − k − 1)
• n = number of observations, k = number of predictors (not counting intercept)
– Use adjusted R2 when comparing models with different numbers of predictors.
R-squared vs. beta (in investing)
– R2 measures how closely an asset’s returns move with a benchmark (goodness of fit).
– Beta measures the sensitivity (magnitude of returns) of an asset relative to the benchmark (systematic risk).
– Combined use: R2 tells you whether beta is meaningful. High R2 + a given beta → that beta is a useful descriptor of relative movement.
Limitations and pitfalls
– No causality: a high R2 does not mean X causes Y.
– Overfitting: R2 can be artificially high by adding many predictors; it may not translate to predictive power.
– Misspecified functional form: linear R2 may be low when relation is nonlinear.
– Outliers: can distort R2 upward or downward.
– Multicollinearity: can inflate variances of coefficient estimates even if R2 is high.
– Not a measure of prediction accuracy on new data—use cross-validation, RMSE, MAE, or out-of-sample R2 for predictive performance.
Why R-squared may be low
– High inherent variability / noise in Y.
– Important explanatory variables are omitted.
– Measurement error in X or Y.
– Wrong functional form (use transformations or non-linear models).
– Small sample size.
What is a “good” R-squared?
– Context-dependent:
• In finance: R2 > 0.85 often indicates strong tracking to a benchmark; <0.70 typically indicates weak correlation.
• In many empirical fields, modest R2 (e.g., 0.2–0.5) can still be useful if effects are meaningful and statistically significant.
– Never judge model quality by R2 alone—consider statistical significance, residual diagnostics, predictive checks.
Is a higher R-squared always better?
– Not necessarily. A higher in-sample R2 can reflect overfitting or irrelevant variables. Use adjusted R2, cross-validation, and other model-selection criteria (AIC, BIC) to assess true usefulness.
Practical steps to improve R-squared (without overfitting)
1. Re-examine model specification:
• Try transformations (log, power) or polynomial terms if the relationship is nonlinear.
2. Feature selection / engineering:
• Add relevant predictors that theory or data suggest matter.
• Create interaction terms if variables jointly affect Y.
3. Address data quality:
• Check for measurement error, coding mistakes, and correct them.
• Detect and treat outliers if they’re errors or understand their influence.
4. Reduce irrelevant predictors:
• Use stepwise selection, LASSO, or domain knowledge to remove noise predictors.
5. Handle multicollinearity:
• Check variance inflation factors (VIF).
• Combine correlated predictors (principal components) or drop redundant ones.
6. Use regularization and cross-validation:
• Methods like LASSO or ridge penalize complexity and improve out-of-sample performance.
• Use k-fold cross-validation to evaluate predictive R2 on held-out data.
7. Consider non-linear or more flexible models:
• Decision trees, random forests, gradient boosting, or splines can capture complex relationships.
8. Evaluate with multiple metrics:
• Use adjusted R2, RMSE, MAE, AIC, BIC, and out-of-sample R2 to judge model improvements holistically.
Using R-squared in investing
– R2 of a fund vs. benchmark indicates % of returns explained by the benchmark.
– High R2 (85%–100%): fund closely follows the benchmark — beta is more meaningful.
– Low R2 (≤70%): fund’s returns are less tied to the benchmark; using the benchmark to explain fund returns is less useful.
Bottom line
– R-squared is a useful, intuitive measure of how much variation in a dependent variable a model explains, but it has important limitations. Always interpret it in context, combine it with adjusted R2 and other diagnostic/predictive metrics, guard against overfitting, and use domain knowledge when building or assessing models.
Source
– Investopedia — “R-Squared”
Editor’s note: The following topics are reserved for upcoming updates and will be expanded with detailed examples and datasets.